The greatest theory of all time?
The more I learn about the Bayesian brain, the more it seems to me that the theory of predictive processing is about as important for neuroscience as the theory of evolution is for biology, and that Bayes’ law is about as important for cognitive science as the Schrödinger equation is for physics.
That is quite an ambitious statement: if our brains really are Bayesian, which is to say that predictive processing is the fundamental principle of cognition, it would mean that all our sensing, feeling, thinking, and doing is a matter of making predictions.
In this blog post, I won’t discuss the ample evidence we have for this theory,1 but limit myself to presenting the idea itself. Though it is probably not the greatest theory of all time, it does have the potential to become the greatest theory in the history of cognitive science.
What is predictive processing?
During every moment of your life, your brain gathers statistics to adapt its model of the world, and this model’s only job is to generate predictions. Your brain is a prediction machine. Just as the heart’s main function is to pump blood through the body, so the brain’s main function is to make predictions about the body. For example, your brain predicts incoming sensory data: what you’re about to perceive from within (interoception) as from without (exteroception).
We used to think that perception is a simple feed-forward process. Say, you open your eyes and sensory data travels from your retina into your brain where it is fed forward through multiple levels of cortical processing, ultimately leading to a motor reaction. If your brain is Bayesian, however, it doesn’t process sensory data like that. Instead, it uses predictive processing (also known as predictive coding)2 to predict what your eyes will see before you get the actual data from the retina.
Your brain runs an internal model of the causal order of world that continually creates predictions about what you expect to perceive. These predictions are then matched with what you actually perceive, and the divergence between predicted sensory data and actual sensory data yields a prediction error. The better a prediction, the better the fit, and the less prediction error propagates up the hierarchy. Wait, what’s a hierarchy?
Your internal model (also called generative model because it generates predictions) is structured as a bidirectional hierarchical cascade:
- The model is a cascade because it involves multiple levels of processing, multiple cortical areas in the brain.
- The model is hierarchical because it comprises higher and lower processing layers: lower levels process simple data (e.g., visual stimuli, affective signals, or proprioceptive data), higher level process categorizations (e.g., object recognition, emotion classification, action selection), and the highest levels process mental states (e.g., mental imagery, emotion experience, conscious goals, planning, reasoning).3
- The model is bidirectional because signals continually propagate in both directions: predictions move downward to pyramidal neurons of lower levels while prediction errors move upward to pyramidal neurons of higher levels.
Each processing layer predicts what’s happening at the layer below and receives an error signal from that lower layer—a dynamic process that may or may not loop all the way down to incoming sensory data. A model is successful when the predictions it generates at every layer of the hierarchical cascade are accurate, producing only minimal prediction errors.
If your predictions don’t fit the actual data, you get a high prediction error that updates your internal model—to reduce further discrepancies between expectation and evidence, between model and reality. Your brain hates unfulfilled expectations, so it structures its model of the world and motivates action in such a way that more of its predictions come truer. Here’s a schematic diagram of how this all works:
The black circles in the diagram (at the prediction error sources) depict something I haven’t explained yet, namely expected precisions. Precisions determine the weight of prediction errors. If your brain expects that a certain prediction error won’t be particularly reliable or important, it decreases its weight and thus the extent to which the error can update your internal model. For example, if it’s dark and foggy outside, you can’t expect the visual information you receive to be particularly reliable, so you add that expectation—in the form of expected precisions that modulate error signals—to your predictions in order to prevent your model from being unduly distorted by unreliable data. In other words, precisions are a measure of uncertainty that expresses your brain’s trust in the expected sensory data and the ensuing the error signals. The brain wants prediction errors to have a high signal-to-noise ratio.
Expected precisions are formally equivalent to inverse variance and functionally equivalent to attention. When you pay a lot of attention to an object, you become confident that the information you get from it is relatively accurate. Say, you’ve been mindfully looking at a Ferrari under good lighting conditions, so you’re pretty confident that it is red. The more attentive you are, the more weight you put on potential error signals. By contrast, if you’re not really focused, your brain turns down the “error volume,” i.e., the impact a prediction error can have on your model.
What makes the brain Bayesian?
The basic idea of Bayesian probability is that you update your beliefs in the light of new evidence. For instance, if you see a dog-like creature running towards you, your belief that it will attack you may be 34%; if you recognize that it’s an unmuzzled pit bull, your belief that you’ll be attacked may rise to 78%; if it starts barking, 92%; if another barking dog passes you from behind, the probability of your attack-belief being true drops again. That’s a vague and incomplete example, but it illustrates how your certainty of a belief may change as you gain new evidence. Initial belief + new evidence = updated belief.
Bayes’ law is the mathematical formalization of that idea: P(B|E)=P(E|B)*P(B)/P(E), which calculates the conditional probability of your belief B being true, given evidence E. This is what the brain approximates4 during predictive processing. In particular, your brain updates its statistical model of the world by integrating prediction errors in accordance with Bayes’ theorem; hence the name Bayesian brain.5
With your model’s prediction or prior probability P(B) and the lower-level data E within a broader hypothesis space P(E), your brain learns about the likelihood P(E|B) of E, given hypothesis B. Applying Bayes’ rule, this yields a posterior probability P(B|E), which determines the prediction error.6 The error signal, modulated by attention (an expected precision), then propagates back upward to the next higher level where it updates the model, correcting your brain’s probabilistic map of reality, and on goes the hypothesis testing. Imagine, for instance, you want to cross a street:7
- Based on a vast set of hypotheses B about the current situation, your brain computes a hierarchical cascade of predictions P(B), including
- priors about how shapes, colors, and noises will change (for low-level perceptual inference),
- priors about you moving your eyes, head, and legs (for low-level active inference),
- priors about vehicles in motion and changing traffic lights (for high-level perceptual inference),8 and
- a prior about you standing on the other side of the street (for high-level active inference).
- Your brain receives sensory data E about what’s actually happening on the street and in your body, including
- exteroceptive data representing shapes, colors, and noises,
- proprioceptive data representing eye, head, and leg movements,
- conceptual data representing vehicles and traffic lights, and
- agentic data representing your relation to the goal state.
- Your model contextualizes that data to get P(E) and estimates the likelihoods P(E|B) of the data, given your hypotheses.
- Your brain implicitly applies Bayes’ theorem to estimate the posterior probabilities P(B|E) and determines prediction errors at the relevant processing layers by
- matching posteriors and priors regarding shapes, colors, and noises,
- matching posteriors and priors regarding your eye, head, and leg movements,
- matching posteriors and priors regarding vehicles and traffic lights, and
- matching the posterior and the prior regarding your goal state.
- The prediction errors, weighed by expected precisions, propagate up the hierarchical cascade, level by level, and alter the corresponding set of hypotheses, thus updating the model to reduce future prediction errors; or they are actively reduced through motor signals to the muscles that move your eyes, head, and legs (more about that second option in a second).
After enough cycles of hypothesis testing and model updating during constant sensorimotor interaction with the world, you’ll find yourself at the other side of the street (if all goes well). This complex predictive process can be described as your brain continually implementing Bayes’ theorem.
What does this explain?
The scope of the Bayesian brain hypothesis is extremely ambitious. It’s meant to be a unifying framework for all neural, cognitive, and psychological phenomena. If true, predictive processing explains, at a computational level, everything about the brain and mind—for reasons we shall see soon. But let’s first take a look at how the framework of predictive processing applies to various cognitive phenomena:
- Perception is the prediction of sensory inputs and the inferential9 process of minimizing prediction error by changing the internal predictive model. This is perceptual inference: reducing prediction error by updating your model so that sensory inputs match with prior expectations. Through perception, you make your model more similar to the world.
- Imagination may be a side-function of the brain’s predictive machinery based on high-level predictions that don’t travel all the way down the hierarchical cascade and aren’t matched with current sensory data, for example, by radically lowering the precision weighting on low-level error signals.
- Delusions and hallucinations may stem from alterations in the ability to integrate incoming data with perceptual predictions (see Griffin & Fletcher 2017).
- Action is proprioceptive10 prediction and the inferential process of minimizing prediction error by changing sensory inputs. This is active inference (also called predictive control): reducing prediction error by moving the body so that sensory inputs match with prior expectations. Through motor action, you make the world conform to your model. As the predicted proprioceptive states are not yet actual, actions change the world to make them so (which means that all your actions are essentially self-fulfilling prophecies).
- Reflex arcs operate at the lowest layer of the active inferential process and work to fulfil proprioceptive predictions until the expected sensory input is obtained. Unlike with perceptual inference, the model parameters are not updated, but kept stable.
- Intentional behavior is a high-level prediction about an abstract future goal state. Once a concrete opportunity to act arises, it entrains multilevel cascades of lower-level predictions to change the world in a way that makes the high-level prediction come true (i.e., achieve the goal state), thereby reducing long-term prediction error. The achievement of conscious goals is a good example of predictive processing operating at large spatiotemporal scales, for it may take minutes, months, or years to achieve a goal, and it may require one, say, to travel overseas).
- Emotion is interoceptive11 prediction and the active inferential process of minimizing prediction error by triggering physiological changes (see Seth 2013).
- Drive is the active-inferential process of minimizing interoceptive prediction error through high-level action. For example, when the brain detects low blood sugar levels through interoceptive inference and the resulting prediction errors aren’t resolved through autonomic control (e.g., metabolize body fat), you will feel driven to minimize error signals through voluntary action (e.g., eat sugary food).
- Mental disorder may be the inability to reduce interoceptive prediction errors.
- Attention is expected precision optimization; it modulates the weight of prediction errors. Shifts in attention result from the hyperprior that the same sensorimotor hypotheses (lower-level priors) should not be retained for too long because the world is an ever-changing place that will kill you if you don’t adapt (which you can’t if you never shift your attention, if you always give the same weight to the errors of related predictions).
- Learning is the updating of your internal model based on prediction errors so that your predictions gradually improve. The better your predictions about the causal, probabilistic structure of the world, the more effectively you can engage with it. This is why teenagers move more fluently through space than toddlers, why you’re less clumsy at playing a sport or a musical instrument after having practiced the relevant motor skills, and why it’s important to pursue truth—enhanced effectiveness due to better predictions.
- Memory consists of the learned parameters of your internal model, whereas its non-acquired parameters would be the innate knowledge evolution has genetically built into your nervous system. Both parts determine your brain’s predictions.
- Self-awareness is the inferential process of minimizing prediction error by changing your internal self-model, i.e., the model that generates predictions about what’s most likely to be “you” (see Braun et al. 2018).
- Agency arises from a good enough fit between your self-model’s predictions and exteroceptive input.12 If the prediction error is too high, you feel like you’re not in control of your actions.
- Ownership arises from a good enough fit between your self-model’s predictions and proprioceptive input. If the prediction error is too high, you might experience that one of your limbs doesn’t belong to you, or you might have an out-of-body experience.
- Belief is a hyperprior; a systemic prior with a high degree of abstraction; a high-level prediction that entails general knowledge about the world. Some examples:
- Physical beliefs. You expect things to change over time, you expect heavy objects to fall fast, and you expect that you can’t move left and right at the same time.
- Physiological beliefs. You expect to see something when you open your eyes and you expect fire to hurt and burn your skin.
- Psychological beliefs. You expect a great performance to make you feel proud and you expect to feel regret when you shy away from a challenge.
- Social beliefs. You expect happy people to smile, offensive words to trigger a reaction, and power to corrupt.
- Cultural beliefs. You expect cars to slow down as they approach stop signs, special offers to be highlighted in stores, and handshakes to not last an hour.
Notice how the Bayesian brain also manifests in some of our deepest values:
- We value truth because accurate beliefs about the world allow us to make good predictions.13
- We value honesty and authenticity because we know what we can expect from honest, authentic people, which improves our prediction making in social contexts that involves them.
- We value simplicity because simple beliefs enable us to generate high-level predictions quickly.
- We value wisdom because reflected life experience equips us with relatively reliable hyperpriors that minimize long-term prediction error.
At the same time, the predictive nature of human cognition explains why we tend to get enslaved by habits, feel drawn to comfort zones, and shy away from uncertainty. Beyond habits, comforts, and certainties, our priors and hyperpriors are less reliable, which leads us to make worse predictions and get higher prediction errors. But high prediction errors are precisely what the brain works so hard to prevent. The effort of radically updating our models to accommodate novel, unfamiliar situations can be so high that the brain simply decides to trigger feelings of fear, anxiety, or discomfort to motivate us to stick to what we can more reliably predict, i.e., our habitual behavior patterns in comfort zones of soothing regularity.
If we want to leave or expand our comfort zones, we must convince our brains that the immediately resulting high prediction errors are worth it at a broader timescale. Accordingly, we can conceptualize willpower as a hyperprior—namely, as the abstract systemic prediction that, in certain contexts, higher prediction errors in the short term will lead to lower prediction errors in the long term. Self-control may thus minimize overall prediction error. But how can the average prediction error over time even matter to a brain that constantly has to deal with immediate errors in any given moment? The answer is to be found in…
The free energy principle
Due to the second law of thermodynamics, the majority of your body’s possible states are death and dysfunction. You’re lucky that you’re alive and able to read this. Sooner or later, entropy will catch up with each of us, and we will enter the physiologically unexpected yet physically probable and inevitable state of death. Yet this is precisely what we evolved to avoid! Evolution “designed” us to struggle for survival, to fight chaos, to resist entropy. The human body evolved to maintain itself within expected states that are easily predictable.
According to the free energy principle, we achieve such maintenance by suppressing our free energy, the information-theoretic equivalent of overall (long-term average) prediction error. Everything we do (and everything any living creature does) is, on average and over time, done to minimize free energy, which corresponds to the brain’s job of minimizing prediction error.
And all this to stay within expected, relatively stable states… Why? Because we can survive only within a certain range of physiological states. Too hot and we die, too cold and we die, too much oxygen and we die, too little oxygen and we die, et cetera. Too much or too little of anything will kill us, hence the fundamental physiological principle of homeostasis.
To make a long story short, the following principles are hypothesized to be equivalent, though operating at different levels of description:
- cognitive level: minimize prediction error, maximize model fit
- information-theoretic level: suppress free energy,14 reduce surprisal15
- physiological level: maintain homeostasis
- physical level: resist entropy
- biological level: survive
The upshot here is that predictive processing captures what the brain contributes to the body’s evolutionary fitness.16 According to Friston (2009), the free energy principle also explains all structural and functional aspects of the brain, including its anatomy, connectivity, synaptic physiology, electrophysiology, and psychophysiology. I must concede, however, that there’s still much about free energy I don’t understand, and so much math I haven’t gone through yet, that it would be preposterous for me to keep writing as if I actually knew what I’m talking about here.
Five objections
Sensory deprivation
Problem. If all our brains want is low prediction errors, shouldn’t the winning strategy be to motivate us to lock ourselves up in a dark, silent room? What better way to make good predictions about expected sensory input than depriving the body of sensory stimulation? In total darkness, we can’t get high prediction errors for expecting to see black, black, and black.
Answer. Even though this strategy may be effective (and relaxing) in the short term, the demands of the world and our bodies cannot be avoided for long. The brain, being a bodily organ, never ceases to generate interoceptive prediction cascades, no matter the degree of exteroceptive deprivation. Soon we’ll feel bored, hungry, and driven to leave the room again, which will put our bodies in increasingly unexpected states. Locked in a dark, silent room for a long period, we can’t maintain homeostasis. In Bayesian terms, we have a hyperprior that occupying the same state for too long will increase long-term prediction error. To always stay within a range of expected states, it’s usually more efficient to engage with a stimulus-rich environment (except, of course, when we need to sleep).
Human rationality
Problem. Humans are notoriously bad at probabilistic reasoning, at thinking rationally about probabilities.17 So if mental processes can barely handle Bayesian inference, how can we assume that the brain, which produces the mind, is Bayesian?
Answer. The Bayesian brain is not inhabited by some homunculus doing complicated math. Its probabilistic inferences are not explicit, but implicit. Brain cells fire in a way that “naturally” approximates Bayes’ law, just like ants move in a way that “naturally” approximates Gaussian and Pareto distributions without there being the need for a chief ant strategist who calculates the optimal path for his colony. What does “naturally” mean here? In the case of ants: pheromone secretion. In the case of neurons: Bayesian sampling.
The Bayesian brain is a sampler, not a calculator: it samples information from a local landscape of probabilities (its environment). Over time, action and perception, associated with a practically infinite number of predictions, yield an effectively infinite number of samples from the unimaginably complex probability distributions that constitute reality. This enormous amount of data makes the brain (asymptotically) conform to the laws of probability, thus mechanistically realizing Bayes’ theorem. Of course, the brain doesn’t store all those samples; it just uses them to update its internal model of the world before it draws new samples from its current local environment.
The human mind, by contrast, having only a finite number of samples, must rely on cognitive shortcuts (heuristics) to generate predictions. Although this has proven evolutionarily adaptive, it now leads us to be cognitively biased and statistically irrational. In Bayesian Brain without Probabilities, Sanborn & Chater (2016) explain how sampling produces various reasoning biases. Remember, however, that laboratory studies on cognitive biases are conducted in extremely scarce environments. When people make everyday cognitive judgments in rich, familiar, realistic contexts, their reasoning can be much closer to Bayes optimality than with random judgments in a lab context where sampling is radically limited (see Griffiths & Tenenbaum 2006 and Maguire et al. 2018).
Moreover, consider how the confirmation bias (our tendency to focus on information that confirms our preconceptions) follows directly from the Bayesian brain hypothesis. After all, preconceptions are vital for the internal model our brain uses to make predictions. What we notice and focus on is highly determined by our preconceptions, by what we already know. And since predictive processing makes us search for whatever is most likely to verify our predictions, it will generally favor confirmatory evidence, rather than surprising evidence, as a strategy to minimize prediction error.
Offline cognition
Problem. The brain does so much more than perceive the world and command motor actions. Not everything we do is directly related to perception and action. We plan holidays, make financial decisions, reflect on philosophical arguments, and ruminate about events that happened years ago. Such cognitive activities happen “offline,” i.e., without immediate real-world interaction. How can predictive processing account for that?
Answer. Offline cognition is only a form of high-level prediction making that doesn’t necessarily propagate predictions all the way down to the lowest layers of the hierarchical cascade (or suppresses low-level model updates by radically lowering respective precision weights)—although it might at some point, in which case the timescales are simply much larger than those of “online” cognition, i.e., cognition concerned with immediate sensorimotor interactions in the present environment.
Phenomenology
Problem. The world doesn’t look like a complex set of intertwined probability density distributions. It looks rich and colorful and usually unambiguous, often beautiful. So how can we say that cognition is nothing but predictive processing? Is this not overly reductionistic?
Answer. Predictive processing is a theory about how the brain encodes information about the world, not how people experience it. Phenomenology is simply on a higher level of description than cognitive science, similar to how physics could (on an even lower level and to little avail) describe cognitive processes in terms of molecules, atoms, and electrons. A more interesting question would be how much of psychology can be reduced to predictive processing.
Falsifiability
Problem. If predictive processing can explain everything about the brain and mind, does this not expose its triviality? Even more, researchers might retroactively tweak their prior assumptions to make any experimental result fit a Bayesian interpretation post hoc. This would mean that the theory effectively explains nothing. What keeps the Bayesian brain hypothesis from being trivial and unfalsifiable?
Answer. Evolution, too, is an extremely ambitious theory that aims to explain everything about life and biological systems with only a few basic tools. Naturally, this invites a lot of post-hoc theorizing, especially in the field of evolutionary psychology where just-so stories are often isolated from alternative interpretations. But this doesn’t automatically invalidate the theory. All it means is that we must think harder to come up with viable alternative hypotheses and stipulate priors based on independent evidence. “For example,” writes Hohwy (2015, p. 14), “there is independent evidence that we expect light to come more or less from above […], that objects move fairly slowly […], and that we expect others to look at us.” Such kinds of evidence should be integrated with neuroscientific studies.
Predictive processing and the free energy principle can be falsified. If descending prediction signals were found not to carry expected precisions, this would falsify the theory. Or it might turn out that some brain areas are not best described in terms of prediction error signaling. Lastly, the discovery of an organism that doesn’t at all act to stay within expected states (to maintain homeostasis) would be a wholesale falsification, though highly unlikely.
Conclusion
Predictive processing provides a framework for understanding all areas of neuroscience and cognitive science at a computational level. Although the Bayesian brain theory is still in its fledgling stage, confirmatory evidence is flowing in on a weekly basis from a vast range of different fields. And although it offers a highly integrative and ambitious account of the brain and human cognition, it does leave much unspecified. For specification, we must also engage with the paradigms of evolution (as cognition has an evolutionary history), embodied embeddedness (as cognition is embedded in physical environments), and sociocultural situatedness (as cognition is situated in social and cultural contexts). Only if we cover and combine all levels of analysis can we truly understand how humans work.
Read more
- Philosophy and Predictive Processing
- How the Brain Makes Emotions
- The Bayesian Brain: Placebo Effects Explained
References
Braun N, Debener S, Spychala N, Bongartz E, Sörös P, Müller HHO, Philipsen A (2018). The Senses of Agency and Ownership: A Review, Frontiers in Psychology, Vol. 9(535).
Buckley CL, Kim CS, McGregor S, Seth AK (2017). The free energy principle for action and perception: A mathematical review, Journal of Mathematical Psychology, Vol. 81, pp. 55-79.
Clark A (2013). Whatever next? Predictive brains, situated agents, and the future of cognitive science, Behavioral and Brain Sciences, Vol. 36(3), pp. 181-204.
Friston KJ, Stephan Ke (2007). Free-energy and the brain, Synthese, Vol. 159(3), pp. 417-458.
Friston KJ (2009). The free-energy principle: a rough guide to the brain?, Trends in Cognitive Sciences, Vol. 13(7), pp. 293-301.
Griffin JD, Fletcher PC (2017). Predictive Processing, Source Monitoring, and Psychosis, Annual Review of Clinical Psychology, Vol. 13, pp. 265-289.
Griffiths TL, Tenenbaum JB (2006). Optimal predictions in everyday cognition, Psychological Science, Vol. 17(9), pp. 767-773.
Harkness DL & Keshava A (2017). Moving from the What to the How and Where – Bayesian Models and Predictive Processing, Philosophy and Predictive Processing, 16.
Hohwy J (2015). The Neural Organ Explains the Mind, Open MIND by Metzinger T & Windt JM (Eds.), 19(T).
Maguire P, Moser P, Maguire R, Keane MT (2018). Why the Conjunction Effect Is Rarely a Fallacy: How Learning Influences Uncertainty and the Conjunction Rule, Frontiers in Psychology, Vol. 9(1011).
Metzinger T, Wiese W (2017). Vanilla PP for Philosophers: A Primer on Predictive Processing, Philosophy and Predictive Processing, 1.
Sanborn AN, Chater N (2016). Bayesian Brains without Probabilities, Trends in Cognitive Science, Vol. 20(12), pp. 883-893.
Seth AK (2013). Interoceptive inference, emotion, and the embodied self, Trends in Cognitive Science, Vol. 17(11), pp. 565-573.
Thornton C (2016). Predictive processing simplified: The infotropic machine, Brain and Cognition, Vol. 112, pp. 13-24.
- Some noteworthy evidence includes predictive coding by the retina, binocular rivalry, perceptual completion, repetition suppression, multisensory integration, motion perception, synesthesia, and non-classical receptive-field effects.
- Predictive coding is a data compression strategy (used, e.g., for the lossless compression of digital images, audio, and video), whereas predictive processing is the use of that strategy to explain how hierarchical generative models flexibly combine up- and downward flows of information in the nervous system to minimize precision-weighted prediction errors. In short, predictive processing is the application of predictive coding in a cognitive context; even shorter, predictive processing is hierarchical predictive coding.
- In general, lower levels (also called unconscious or subpersonal levels) are more automatic, concrete, detailed, spatially and temporally precise, and associated with specific sensory or motor areas, whereas higher levels (with the agentic or personal level on top) are more voluntary, abstract, complex, integrative, and distributed over multiple cortical areas, especially associative ones. A very clear example of that hierarchy comes from language processing where we have first, on the lowest level, an auditory stimulus, then phonetics, then phonology, then morphology, then syntax, then semantics, then a situation model, and finally the communicated message.
- Your internal model only ever approximates Bayes optimality because there is no homunculus sitting in your brain doing math and because your neural resources are limited: your brain has to deal with limited memory, limited attentional capacity, limited space, limited time, and incomplete information.
- I should mention here that, according to Thornton (2017), predictive processing could also be specified by Shannon information theory as an alternative to Bayesian inference.
- Formally, the prediction error is calculated as the KL divergence—DKL(posterior||prior)—between prior probabilities P(B) and posterior probabilities P(B|E); see Jakob Hohwy (2014), The Predictive Mind, p. 58.
- I’m again very casual in my description here; for a scientific example, see Harkness & Keshava 2017.
- In addition, your high-level expectation includes hyperpriors such as the belief “that whenever auditory and visual estimations of the spatial location of a signal source are reasonably close, the best overall hypothesis is that there is a single source—in this case, a rapidly moving car” (Clark, 2016, Surfing Uncertainty, p. 64).
- Perception is inferential because it infers the latent causes of observations. Predictions are inferences about the causal structure of the world.
- Proprioception is the ability to sense the relative position of your body parts through
signals frompredictions about the dynamic states of muscle spindles, tendons, and joints. - Interoception is the ability to sense the internal condition of your body through
signals frompredictions about physiological states. - Compare this to Andy Clark’s conceptualization of agency: “When proprioceptive prediction error is highly weighted yet suitably resolved by a stack of top-down predictions (some of which reflect our goals and intentions), we feel that we are the agents of our own actions” (2016, Surfing Uncertainty, p. 158).
- I see predictive processing as a profound reason for my belief that rationality is important, although I am aware that it is somewhat philosophically scandalous to provide a ‘reason for reason’.
- I’m putting free energy at the information-theoretic level because the variational free energy discussed here is not the same as the concept of free energy in thermodynamics (see Buckley et al. 2017 and Friston & Stephan 2007).
- Surprisal is a technical term for “surprise” that denotes how unlikely an event is, given a model. For example, your brain gets “surprised” by an unmet expectation (surprising sensory input) and a high prediction error. Since your brain hates nothing more than “surprise,” it does everything it can to avoid it. Mathematically, surprisal is the negative logarithm of an event’s probability.
- At least in terms of an animal’s survival; the survival of its genes through reproduction is yet another story.
- Examples include numerous logical fallacies and cognitive biases such as base rate neglect, probability neglect, gambler’s fallacy, prosecutor’s fallacy, subadditivity effect, conjunction fallacy (but see Maguire et al. 2018) and many more.
What a great piece – the best one I’ve read on PP (Clark’s book was almost painful for a layman like me). I am a language teacher in Japan and this year I started producing a magazine that connects brain science to teaching. See site link below. Our readership hit 500 first year, and we have people subscribed from about 18 countries. We expect 5000 in 30 soon, maybe by the end of 2019. Would you like to write a 2-4 page article for us? The first half of this piece (modify it, so that it is new) would be good.
Very interesting and excitant the Bayesian model theory since shed new light the way human brain knows the world.
My question is regarding we knows the world as extrrnal or the world conception is internas.
Nice work.
This is an excellent description but I have a very different perspective on the operation of the brain and mind. This is a semantic perspective. I don’t subscribe to the predictive brain hypothesis.
I have identified something linguists have failed to recognise, for reasons that are not hard to explain. I call it the alphabet or code of meaning, a compact system of semantic features that constitute the necessary structure of the meanings of words in all languages. It can also be regarded as the language or metalanguage of the mind because it underpins, in addition to word meaning, semantic memory and human interaction with the world. We share this alphabet of meaning with animal and organisms of all types. It can also be described as ‘the dimensions of the space of the interactions of organisms with their umwelts’. This system of meaning is arguably as necessary to life as the genetic code. The semantic code is of two types: one that relates to parameters of the physical world and the other that is of a normative character (feedback): this type is the basis of action/behaviour. The semantic code is substantiated in neural nodes and hubs.
I will not go into more detail here but papers on my website provide the necessary background. I am happy to discuss further.
Hey Trevor,
your approach sounds very intriguing. I will certainly take a look at your papers. What would be a key point of contradiction between your hypothesis and the predictive brain perspective?
Hi Dom,
I’m pleased you find my angle interesting. I’ll mention just two points.
1. Predictive brain theorists focus on perception, but perception is just one form of cognition. I have seen nothing in what I have read about how the mind handles the normal form of thought, concept after concept flowing through the conscious mind. Bayesian prediction seems a clunky way of explaining this process which is the structured recovery of concepts from semantic memory.
2. So far as I can see predictive theorists are silent on the semantics of cognition. How do they account for meaning? What precisely are concepts and word meanings and, more relevantly, what is the semantic form of the interpretation of percepts that are subject to predictive processing?
I have a model of the semantic form of word meaning (described in most of my papers.) I hold that the process of interpretation of percepts is best conceived semantically, as working out the precise parameters of the meaning of a percept that fits the context. The most important parameters are what I call semantic factors, the small system of necessary,universal structural parameters of meaning. This structure is elaborated on by sensory images and encyclopaedic meaning. I may have mentioned a key aspect of the factoral code is its duality – it consists of two types, physically descriptive and affective or normative. This is crucial to my conception of meaning.
Cheers,
Trevor
Hey Trevor,
I reckon that to an extent this depends on how essential one finds the paradigm of embodied cognition, particularly as it relates to off-line cognition. Strong embodiment would radically shift the idea of “the normal form of thought” and make Bayesian prediction appear more integrated for cases of seemingly pure mental association.
I agree that Bayesian brain theorists say little about meaning and the semantics of cognition, but I still fail to see how this contradicts their predictive processing hypothesis.
Hi Dom,
Embodiment is integral to the system of semantic factors, the dimensions of bodily interaction with the world. The system is nothing without the body. The descriptive factors require sensory cognition and the affective factors, gut reactions, basic normative intuitions, are dependent on specific neural regions that are expressed interoceptively .
The normal form of thought is intrinsically embodied, in my book. Even when I put together in my mind an abstract idea like ‘integral to the system’, body and mind are working in concert.
I can’t follow the logic of your sentence, “Strong embodiment….” How does embodiment make the predictive process more integral to thought?
You are probably aware of Jacob Hohwy’s paper “The self-evidencing brain” which argues strongly that Bayesian error minimization is the right approach to understanding cognition. I have only recently read it and I’m still absorbing it. The idea of self-evidencing is powerful and it expresses the nub of my approach. I plan to write to Howhy contesting his well-argued position.
The basic idea as I understand it is that the brain interprets sensory evidence as categorial and fit to act on, but only via a Bayesian probability analysis of a percept.
In my view the brain self-evidences in a different way. It interprets the evidence through meaning: the agreement of the abstract ontological structure of the percept with its own factorally structured model of the percept. A difficulty for PP including Friston’s version, is that it has no model of meaning. Mine is the factoral model with the biological mandate that I claim for it. This model is substantiated in the neural web of the brain. When the nature of factoral meaning is grasped I believe the redundancy of PP becomes clear.
PP theorists assume that there is no way that the brain can simultaneously access valid internal estimates of externalities and true states of affairs in the world, and act on this correlation. This is precisely what the factoral model allows.
Friston postulates Markov Blankets with hidden variables outside and inside the mind that the MB functions to reconcile by PP. My view is that the external hidden variables are the structural features of things and events that our brains are equipped by evolution to recognise, and that are represented by the descriptive semantic factors alone. The internal variables are the two kinds of factoral dimension that are substantiated in our brains by evolutionary processes. Their function is to enable us to interact adaptively with the external world.
To illustrate, when I see a spoon on the table I view a physical object as the same kind of entity that an animal could identify, a perceptual gestalt. But the structure of the human brain enables me to intuit something else, the spoon’s factoral structure that is roughly made up of the following semantic factors: materiality, particularity, smallness, extension, capacity to contain, tactility, artifactuality, utility. I also have extensive encyclopaedic knowledge of spoons derived from past experience. The sum of this material enables me to instantaneously identify the spoon and a meaningful entity and to interact with it in a useful way. PP has no place in this account.
I hope I am making all of this clear.
Cheers,
Trevor
“With your model’s prediction or prior probability P(B) and the lower-level data E within a broader hypothesis space P(E), your brain learns about the likelihood P(E|B) of E, given hypothesis B. Applying Bayes’ rule, this yields a posterior probability P(B|E), which determines the prediction error.”
I have a doubt about this, isn’t the evidence the same as the as the prediction error in predictive processing? I mean I thought it was the prediction error which enabled to infer the posterior hypothesis by comparing with the prior hypothesis. I mean by knowing the prediction error (wich I assume is the same as the sensory evidence) you know the deviation from the mean so you can estimate the likelihood of the evidence or the error given the prior and then applying bayes estimate the true posterior. So when you say that the posterior probability determines the prediction error couldn’t you say too that the prediction error determines the posterior probability?
From what I understood nothing but prediction errors was computed as sensory evidence, how can it be that you need the posterior probability to determine the prediction error via its KLdivergence with the prior. I don’t know I must be misunderstanding something. If error units, which seem to replace the classical feature detections units, signal mismatches between predicted and observed data isn’t that error or mismatch signal waht is used to infer the posterior hypothesis? maybe the description of the prediction error signal as the KL divergence between pirors and posteriors is a mathematical way of expressing it but I don’t think it is actually the way brain computes the prediction error, it does it by signaling the mismatch between the prior and the evidence and this prediction error signal serves to infer the posterior.
“So when you say that the posterior probability determines the prediction error couldn’t you say too that the prediction error determines the posterior probability?”
Yes, this is exactly right. Every prediction error is influenced by previous prediction errors and influences the prediction errors to come. This is the main idea of Bayesian updating. The key is not to view the integration of Bayes’ theorem as a standalone formula, but as a continuous (and approximative) process.
I have seen that prediction error is not only KLdivergence, but suprisal+perecptual divergence or KL divergence between priors and posteriors. But I still think this is not the way brain calculates prediction error, it does it via error neurons which signal the mismatch between predicted (priors) and observed data (evidence or sensory input) and this yields a prediction error that is the evidence that brain uses to infer the likelihood of the sensory input given the prior and then applying bayes infer the true posterior. What we call sensory signals are nothing but error signals (maybe).
What are Prediction Errors, and Where in the Brain are they Encoded?
Principally, a prediction error can be defined as the mismatch between a prior expectation and reality. Prior expectations are based on an agent’s model of the world, which is partly hard-wired in the structure of neural circuits and partly derived from statistical regularities in the sensory inputs that the agent experiences over a lifetime. As such, a PE signals a deviation of the current state with respect to what is predicted based on the current model of the world, and calls for an update. Exposure to experimentally manipulated environments indeed alters prior expectations, even when such expectations are the result of a lifetime of experience, such as the prior that light comes from above (Adams et al., 2004), or that slow motion speeds are more likely than fast ones (Weiss et al., 2002; Sotiropoulos et al., 2011). The notion of the coding of PEs as a ubiquitous strategy is supported by the observation that PE signals appear ubiquitously throughout the brain, in relation to the processing of sensory signals, value, motor actions, and cognitive control. One may object that, if PEs are so general, to the degree that all our brains do is code predictions and PEs – then how can we achieve such extensive neural specialization for actions, emotions, and generally functional specialization of areas? The answer lies in the notion that, while coding in terms of PEs is a common currency across brain regions, the exact content and nature of these error signals vastly differs between areas and functional specializations. Below, we will first discuss two main classes of PEs that have been reported in the last decades. Roughly, these can be divided into on the one hand perceptual and cognitive PEs, which report the degree of surprise with respect to a particular outcome, and on the other hand motivational PEs, which also report the valence (sign) of a PE, i.e., not only whether the outcome was surprising, but also whether it was better or worse than expected.
here https://www.frontiersin.org/articles/10.3389/fpsyg.2012.00548/full
I still don’t know how this relates to KLdivergence, but my mathemathical knowlegde is very low. I can see that the diference between prior and posterior is guided by the diference between prior and input (I think), but I don’t really know how this is exactly related. What I don’t know neither is if the upper levels update only in base to prediction errors sent by lower levels, are the predictions about the magnitude of this prediction errors or what else? and, is the prediction about the input that will be received or about the distal causes that produce it?
What is inferred I guess are the distal causes but it must be that the distal causes are just the statitical structure of the input.
I think I have understood now the schema, at each level there are superficial pyramidal cells or error neurons that receive afferents which convey sensory evidence and afferents which convey predictions from deep pyramidal neurons at te upper level, and the comparison between the priors or predictions and the sensory evidence yields a prediction error that is conveyed to the next level by the error neuron to update the generative model.
Congratulations for the post, it is very well explained.
Maybe I don’t use the term to infer very precisely, I think the likelihood is not infered but calculated what is infered are the parameters that will constitute perception according to the posterior probability or the likelihood of B given E.
About this other:
” The prediction errors, weighed by expected precisions, propagate up the hierarchical cascade, level by level, and alter the corresponding set of hypotheses, thus updating the model to reduce future prediction errors; or they are actively reduced through motor commands to the muscles that move your eyes, head, and legs (more about that second option in a second).”
I would say that motor commands are not needed in active inference, what gives rise to movement is the reduction of the prediction erros via reflex arcs. Propioceptive predictions about the expected consecuences of a movement are generated and translated into reflex arcs via hardwired connections between sensorimotor and motor areas, according to the active inference that seeks to reduce the prediction errors.
about this you can read that article
https://arxiv.org/abs/1806.02649
“Maybe I don’t use the term to infer very precisely, I think the likelihood is not infered but calculated what is infered are the parameters that will constitute perception according to the posterior probability or the likelihood of B given E.”
I would say that the likelihood is estimated for a given sample of data.
Yes I have some misundertandings about this of inferring, estimating and calculating. From what I’ve seen I think what is inferred is the true posterior according to its estimated likelihood given E. But at the same time priors (posteriors from the previous iteration) are inferences about the distal causes that we are gonna encounter and predictions about the next input. That cast me some doubts since I don’t trully understande how priors can be inferences about (the structure of?) hidden causes and predictions about sensory input at the same time. Maybe it is just how I said, that priors are probability ditributions or the parametes that define them and is its statistical structure what determines which and how are the hidden or distal causes that produce sensation.
I don’t know maybe i’m missing something since I’m not any expert on this field and I don’t get to understand the more formal aspects of the theory.
Well this goes beyond my knowledge but I have seen this about it, which could be interesting (I guess you would understand it better than me, I see it is related to what I said but it’s something I can’t fully understand, I only put what I still understand a bit, because it explains the relation between estimating and inferring)
” Making inferences about causes (e.g. the
most likely cause, or how certain we are that the cause falls
within some interval) depends on some representation of the
relative probabilities of values the causes can take. This
entails representing the probability distribution or density of
the causes. The key density is the conditional or posterior
density that summarises the likelihood of any cause given the
input”
and
“Representational learning is framed in terms of estimating
probability densities of the causes. This is referred to as
posterior density analysis in the estimation literature and
posterior mode analysis if the inference is restricted to
estimating the most likely cause. The mode of a distribution
is the location of its maximum. Although density learning is
formulated at a level of abstraction that eschews many
issues of neuronal implementation (e.g. the dynamics of
real-time learning), it provides a unifying framework that
connects the various schemes considered below. It is
important to appreciate the distinction between simply
estimating the most likely cause (i.e. mode) and the broader
problem of inference. Inference entails estimating the
conditional density not just its mode”
https://www.fil.ion.ucl.ac.uk/~karl/Learning%20and%20inference%20in%20the%20brain.pdf
Well as I supossed inferring and estimating is mostly the same, but not all estimations are inferences. In inferential statistics I think estimating the parameters of a probability density function is what is called inference, this is farthest I can go xD.
Well not of a probability density function, but of any distribution of probabily. The PDF is what let you know the probability of the parameters being like those of your sample given the null hypothesis. But as I can se in the context of predictive processing the inference has an specificic meaning which I think is not exactly the same as in statistical inference.
Yes, estimation is the part of inference that is directed at unknown population parameters. Insofar as we also want to include hypothesis testing, we can simply speak of inference altogether. What’s more important, however, is that during the inferential process, there is no real calculation, but merely approximation via sampling.
Also, there is likelihood inference on the one hand, determined by neuronal connections (what the brain has learned and how it applies that), and Bayesian posterior inference on the other, which appears to be comparatively more process-oriented (what the brain is learning), if such a comparison is meaningful.
“I don’t trully understande how priors can be inferences about (the structure of?) hidden causes and predictions about sensory input at the same time.” The answer to this, I think, lies in what you have written in the sentence before, that is, “priors (posteriors from the previous iteration)” – so it’s not precisely at the same time, but shifted in time from one iteration to the next.
Yes thank you for clarifying this, I think what you call likelihood inference (to help me clarify my doubts) is called perceptual learning in this context and Bayesian posterior inference is called perceptual inference.
“As information ascends
the cortical hierarchy, sensory information is therefore replaced by
prediction error signals that carry the information that has yet to be
explained (Clark, 2013). The prediction errors drive representations in
higher levels of the cortical hierarchy to provide better predictions –
and thereby suppress prediction error signals in lower levels. In addition
to this online evidence accumulation (i.e., perceptual inference),
prediction error signals also drive associative plasticity to update the
generative model; thereby, minimising prediction errors when a similar
situation is encountered in the future (i.e. perceptual learning).”
https://pubmed.ncbi.nlm.nih.gov/32446883/
Based on this and something I read on the article of learning and inference but I dindn’t trully understand.
Thanks a lot you have been very helpful
“I don’t trully understande how priors can be inferences about (the structure of?) hidden causes and predictions about sensory input at the same time.” The answer to this, I think, lies in what you have written in the sentence before, that is, “priors (posteriors from the previous iteration)” – so it’s not precisely at the same time, but shifted in time from one iteration to the next.”
Yes this helps me understand it, but I still think that priors sometimes are said to be predictions about the causes and sometimes to be predictions about the sensory input (if I’m correct if not all the argumentation that follows means nothing xD), doesn’t this means that when we talk about causes we are talking about the statistical structure of the data and what is predicted is this. The thing with perception as unconscious inference is that sensory data per se is not enough to make a good representation (sorry to non-representationalists xD) of the distal causes, so we need to apply perceptual learning and perceptual inference to reach a good interaction with the real world. I mean to some extent sensory input talk about the causes but it doesn’t carry enough information to make a good guess about what is producing it so we need to apply perceptual learning and inference in order to reach an almost direct perception of reality.
Then is the problem of the one-to-many and many-to-one mapping, but this doesn’t means to me that sensory input doesn’t talk about the causes but that perceptual learning and inference are needed in order to correctly interpret it. So there is no difference in saying that we predict the input or the causes, the only thing is that these predictions are needed to correctly infer the real causes behind the input. Or how else can posteriors from one iteration, which talk about the causes, be used in the next iteration as predictions of the input, must be that in this scenario predicting the causes and the input is the same, since input actually carry information about the causes.
And even posterior expectations which are inferences about the causes are conveyed as sensory input for the next level. It is just that in order to correctly interpret this data at each level it has to be completed and compared with the top down predictions to yield a prediciton error and generate a new posterior of higher level of representation which hold more general and abstract information than those from the levels below.
Hahaha I’m obsessed with this of the statistical structure, I think the key point is what you said, but maybe I would need to take in account priors over policies to understand this (and a graduate in mathematics…) and how probabilities over outcomes in one time step are related to probabilities over hidden states in the next step (or the opposite as I posited it). I hope my english is not as horrible as I imagine lol, thaks for your attention you have been very nice.
And maybe or in addition that I’m not having in account that one thing are priors over causes or the marginal likelihood of the belief or hypothesis, and other thing are priors over evidence or the marginal likelihood of the evidence (which in bayes formula serve to weight, these two priors acquired through learning, the likelihood of E given B and obtain the posterior probability; which means making the model inversion and resolving the one-to-many and many-to-one mapping problem).
About this Hohwy says that in Bayes formula dividing for the marginal probability of E serves to maintain the probability between 0 and 1, but I think this might not be totally correct since the joint probability of E and B can never be more than 1 (nor than the marginal likelihood of E). Maybe sometimes probabilities are represented as being more than one and dividing by the marginal probability of E actually serves to maintain it between 0 and 1. I’ve used the term likelihoods and probabilities as equivalent I don’t know if this is correct neither (my english is kinda poor).
The division by the marginal likelihood is the formality to normalize the posterior density. In practice, it is usually an intractable integral that sometimes need not even be computed if the proportional relationship between posterior and likelihood*prior is all that matters in a given problem context (e.g., if you are only interested in maximum a posteriori estimates).
I give up xD. No, I’ve benn re-reading the chapter 2 of Hohwy’s book and I find that what I said about the statistical structure being what talk about the causes was kinda correct. What you pointed out is true too, but my point wasn’t about being inferences or predictions, but about being about sensory input or about hidden causes. All I said after this is kinda misleading and mix too much concepts. It is only that beliefs or hypothesis refers to the statistical model that describe the data (which talk about causes) and that predictions are based on these models but refers to data (what I don’t know is if to say predictions is equivalent to say prior hypothesis). As Friston say,” cortical hierarchies are trying to generate sensory data from high level causes. This means that causal structure of the world is embodied in the backward connecions”.
Maybe I make too literal translations from spanish so I don’t know if saying that hypothesis or statistical structures talk about something sounds weird or even meaningless. I guess potential readers can bear with it if they are interested in what I say (what might be even weirder than saying that abstract entities talk). Hehe
but the normalization factor doesn’t serve to maintain the probability between 0 and 1, what wikipedia says is “The normalizing constant is used to reduce any probability function to a probability density function with total probability of one.” this doesn’t seem the same as Hohwhy says.
And as I qouted before, what is called inference in this context is estimating the PDF, not its mode, I think.
“but the normalization factor doesn’t serve to maintain the probability between 0 and 1, what wikipedia says is “The normalizing constant is used to reduce any probability function to a probability density function with total probability of one.” this doesn’t seem the same as Hohwhy says.”
It means effectively the same: each area under the PDF (i.e., each specified probability) lies between 0 and 1 such that, when adding up all areas, you get a total probability of 1.
Ok thanks so the sum of E given all causes remains 1 (the marginal probability of E). I understood that it served to maintain the probability of every likelihood between 0 and 1 and I found no sense on this.
haha I don’t understand what I said last, I’ll have to read more and talk less but thank you anyway…hehe
I’ve been thinking about this of the normalizing constant and how Howhy explains it and I still find some contraarguments that show his explanation is kinda ambiguous or imprecise. The normalization constant is set to yield a posterior probability PDF that is equal to 1 but in a broader sense to normalize the posterior density what does is that for differents values o E the area under the curve of each posterior remains the same (what doesn’t mean necessarily that it is equal to 1, we could choose a normalizing constant that yields an area for the curve of the posterior different from 1 but is still always the same for all posteriors given differents values of E). So in one hand, to normalize what does is that the area under the curve of all posteriors is equal, what in the case of the normalizing constant is choosen to yield a value of 1 since we are worknig with probabilities.
But with this same logic, to say, as he does, that the role of the denominator in Bayes’ rule is to normalize the resulting posterior to a value between 0 and 1 is, at the very least, vague or imprecise, because it is used to normalize the resulting PDF to 1 (not between 0 and 1), but each specific posterior for a given joint probability (a concrete value) is never gonna be more than 1 since we are working with probabilities and they can never be more than 1 (so when he adds that probabilities are never reported as being 7 or 42 he is, at same time, making it more confuse, and contradicting what he has said before). That is, I don’t understand this in depth and I know it but I don’t see the point to not accept that his explanation has little meaning, except for the sake of the argument from authority.
And saying that it normalizes posterior probabilities to a value between 0 and 1 doesn’t imply that the resulting PDF is gonna be exactly 1, so it is at the very least vague or imprecise and I would even say that it is meaningless because probabilities are always a value between 0 and 1.
I’ve been thinking about this of the normalizing constant and how Howhy explains it and I still find some contraarguments that show his explanation is kinda ambiguous or imprecise. The normalization constant is set to yield a posterior probability PDF that is equal to 1 but in a broader sense to normalize the posterior density what does is that for differents values o E the area under the curve of each posterior remains the same (what doesn’t mean necessarily that it is equal to 1, we could choose a normalizing constant that yields an area for the curve of the posterior different from 1 but is still always the same for all posteriors given differents values of E). So in one hand, to normalize what does is that the area under the curve of all posteriors is equal, what in the case of the normalizing constant is choosen to yield a value of 1 since we are worknig with probabilities.
https://stats.stackexchange.com/questions/481390/why-is-the-normalisation-constant-in-bayesian-not-a-marginal-distribution
Here you can see what I say about the meaning of normalizing the posterior distribution
But with this same logic, to say, as he does, that the role of the denominator in Bayes’ rule is to normalize the resulting posterior to a value between 0 and 1 is, at the very least, vague or imprecise, because it is used to normalize the resulting PDF to 1 (not between 0 and 1), but each specific posterior for a given joint probability (a concrete value) is never gonna be more than 1 since we are working with probabilities and they can never bee more than 1 (so when he adds that probabilities are never reported as being 7 or 42 he is, at same time, making it more confuse, and contradicting what he has said before). That is, I don’t understand this in depth and I know it but I don’t see the point to not accept that his explanation has little meaning, except for the sake of the argument from authority.
And saying that it normalizes posterior probabilities to a value between 0 and 1 doesn’t imply that the resulting PDF is gonna be exactly 1, so it is at the very least vague or imprecise and I would even say that it is meaningless because probabilities are always a value between 0 and 1.
sorry last comment didn’t appear that’s why I repeated it all.
I don’t know this seems kinda tautological to me… probabilities are never reported to be more than one because we do the normalization or we do the normalization because we want to report probabilities? I think what should prevail is that probabilities are always (by definition) between 0 and 1, so we can’t report probabilities being more than one because this wouldn’t be a probability. So saying that probabilities are never reported to be more than one is meaningles and contradictory.
I know that the normalisation constan is not choosen arbitrarily but that is the fact that it is the marginal probability of E what makes the posterior distribution to have a value of 1. Howhy points toward this but he doesn’t explain why and his explanation is plagued of shortcomings and contradictions. And what you say that each sècified probability is maintained between 0 and 1 doesn’ explain it anyway.
Probabilities are, as you said, by definition, always in the interval [0,1]; however, likelihoods (and probability densities) need not be: they can be >1; and we have a likelihood in the numerator of Bayes’ theorem.
I think it is very clear, the numerator of Bayes is already a probability so it lies between 0 and 1. What makes diving it by P(E) is to give us the proportion of each joint probability in relation to P(E), so the sum of all them gives us a value of 1 (that is, the sum of all posterior probabilities for a given E is one, and not less what would be the case is we just sum the joint probabilities without dividing them by P(E)). This is easy to see if you consider that the sum of all joint probailities (numerator of Bayes) is equal to P(E) (indeed this is the definition of the marginal probaility of E). So what means saying that dividing by P(E) maintains the probability of posteriors between 0 and 1?… nothing I guess
I was gonna say only this but since you say likelihoods can be more than 1 I understand that dividing by P(E) gives a probability (so it lies between 0 and 1). So it should be said, in any case, that it serves to obtain a probability, not that it serves to maintain probabilities between 0 and 1 (what doesn’t actually makes sense).
You could have started with this and not just neglecting what I say when it actually made a lot of sense (I had already said that if probabilities can be reported as being more than one then dividing by P(E) actually served to maintain them between 0 and 1, but as I see now this make no sense because for something to be a probability it must be between 0 and 1).
Thanks anyway I understand it better now
I won, Howhy is wrong, I’m left…I mean right XD
Put simply: If we don’t report green being red is because green is not red (unless we are daltonics so maybe we would need to apply some kind of normalisation constant to maintain green being green and red being red). We don’t say apples are lemons because apples are not lemons (maybe in virtue of our perceptive and cognitive system, but this is not the point). I would never say apples are apples because I maintain them being apples and not lemons (despite this makes more sense than saying that that we maintain probabilities between 0 and 1).
Indeed being able to maintain (not actually) probabilities between 0 and 1 is more like being daltonic, so you are able to maintain green being red.
That feels kinda solipsist but if, as Hohwy says, we are at one remove from the world, who knows if apples aren’t the same as lemons and 1 isn’t the same as 7 or 42 . Anyway I’m more on the side of constructivism but as a kind of simbolic mediation and performative action, not in the sense that our brains alone shows us representations that are at one remove from the world.
This of the likelihood on the numerator posit another question. As far as I’ve seen the sum of the probabilities of a likelihood function can be more than 1, but since for each specific value of the parameters there is an asociated value that tells you the probability of obtaining the data for this specific parameter, I think that the numerator of Bayes’ theorem can never be more than one, what can be more than 1 is the sum of all the numerators. So, again, normalisation serves to obtain a PDF for the posterior whit an area of 1 (or to make all the posteriors sum 1) , but not for making each specific value lie betwen 0 and 1 (each specific likelihood is indeed a probaility, but the likelihood function is not a probability distribution so doesn’t necessarily sum to 1).
Another error I find in Hohwhy’s explanation is that he says that P(e) is the marginal probabilty, and this is not correct, P(e) is the marginal likelihood which, I think, is not the same (can’t be the same because in the numerator we have a likelihood so its integral can be more than 1; indeed the sum or the integral of the numerator is P(e) or the marginal likelihood).
“The marginal likelihood, also known as the evidence, or model evidence, is the denominator of the Bayes equation. Its only role is to guarantee that the posterior is a valid probability by making its area sum to 1.
Therefore, its only effect in the posterior is that it scales it up or down, but the shape of the posterior does not change.”
https://www.quora.com/What-is-the-difference-between-marginal-likelihood-and-posterior
I don’t want to get more ino this since it is getting tired from my point of view. But I don’t think that statistical models can be seen accurately as being about the causes, what I think is that the statistical model (or the generative model) is a function that serves to map causes into data to generate predictions about the incoming sensory input. In this sense prior beliefs or hypothesis hold the prior probability of some causes being true and the statistical model to map them into data, in order to generate predictions which are not exactly the same as the priors but are generated by them. More or less this would be the picture (from my mathematical naive point of view).
About this Howhy says “we get that the data in that case is the sensory input and the statistical model is the hypotheis, maintained in the brian, about the causes in the world” I find this is somehow misleading beacuse it makes the statistical model and the causes equivalent, when, I think, the statistical model is what map causes into data, but no the causes themselves (despite he doesn’t exactly say that, maybe this is a problem of myself being a circumstancial thinker who gives to much importance to irrelevant details).
Hey pharmakoi,
in your quote, Hohwy equates the statistical model with a hypothesis about causes in the world, not with the causes themselves. Insofar as inference can be thought of as a mapping of sorts, I see no reason for confusion here, or am I missing something?
Yes, that’s why I say that he doesn’t exactly say that, but if he’s using hypothesis in the sense of priors (i’m not sure but it seemed to me when reading the chapter) then it is not correct because the priors are the parameters of the generative model (the causes), but not the model. Of course the model anyway is a hypothesis about causes in the world and how they are related to data.
I said I have a tendency towards overfitting or circumstancial thought but some pages after he explains that PE minimization serves to get a better fit between evidence and hypothesis so likelihood is maximized, I think this means changing the priors to produce data that better predicts the sensory input but maybe it is about optimizing the generative model too, so he might be using the term hypothesis for the generative model and the priors indistinctly. But as I said this is a bit imprecise because the priors are the paremeters of the model not the model itself (which is about how priors or causes map into data, so it is about them two I think).
Anyway, whatever… haha
About this of the marginal serving only to normalize te posterior, you must be true, but I don’t see how divding by the evidence can be so trivial if it it says what is the proportion that beliefs and evidence occur at the same time over the total times that E occur. If this proportion is high the the posterior must be high and if it is low posterior must be low. I don’t understand how can it be that only joint probabilty informs about posterior. As I said my mathematical knowledge is very low. But how can you know the proportional relationship between likelihood*prior and posterior before knowing posterior? (don’t take me too serious maybe it is impossible for me to understand this whit my little knowledge…)
The problem is that the exact posterior cannot be known because it is computationally intractable. The encompassing model formalizes neural processing via Bayes’ theorem. Actual research in computational neuroscience, however, requires computable models in order to create evidence for that larger model, for the general theory. Maybe you find this paper helpful, which also mentions MAP estimates as “the simplest approach to inference” on page 13.
Hey I decided to give up on this for some time because I was starting to get too comitted and to talk about things that affect me too much in a personal level but I want to put some things clear.
First, all the discussion before about Howhy’s commentary resumes in that saying, as he does, that the marginal likelihood (he doesn’t even say likelihood…) serves to normailze the posterior probabilities to a value between 0 and 1, if you think it carefully, doesn’t even makes sense. What dividing by the marginal likelihood serves for is to obtain the posterior probabilities.
I already understood when looking about variational bayes which is applied in predictive processing that it serves to approximate the probability distribution for the posterior (without having to compute the marginal likelihood). And if you read the paper I posted at the begginig of this conversation about learning and inference you will see that MAP esimates (estimating the mode) is not called inference in this context but only estimation, what they call inference is to approximate the PDF for the posterior. So as I was saying again and again, in order to make (correct) inferences the human cognitive and perceptive systems need to know (at least an aproximation) of the true posterior, what means that it is like if we divided by the marginal likelihood (but maybe not so exact)
About that of the agency that we talked below, I would say that propioception and sensory attenuation does actually have a lot to do with sense of agency and not ownership. This is something related to forward models and corollary discharges or sensory predictions, but in the concrete case of the discrimnation between exteroception and mental imagery and maybe episodic memory I don’t exactly know how this is explained.
Thanks in general for your answers you helped me to understand some things a lot better.
About hallucinations in general there are some sort of explanations that talk about noisy sensory prediction errors and powerful priors, as you can see here but in the case of auditory hallucinations in schizphrenia they usually have to be with inner speech so they are better explained through aberrant precision of propioceptive prediction erros due to a loss of sensory attenuation which also leads to delusions of agency, as you can see here. If hallucinations and agency are always related to forward models, I don’t truly know, but as you can see psychiatric explanations in general doen’t stand out for their precision.
In this paper
https://www.sciencedirect.com/science/article/abs/pii/S001002772030189X
(if you are able to download it, maybe Z-library would be helpful) they talk about what I said and they specifically say
“Free energy is a concept used in statistics to approximate inference
via variational Bayes. The free energy principle thus applies a
particular approximation of BDT, namely variational inference via free
energy minimization, to explain all adaptive behavior.
In order to see how this fits in with the discussion of Bayesian inference
thus far, consider that,due to the difficulty of computing p(e),
directly selecting the hypothesis that would maximize the posterior
probability is not a feasible objective. Progress could be made, however,
if we were to make additional assumptions about the phenomenon
of interest. Let us assume, for example, that there is a family of distributions,
Q, that (ideally) includes the true conditional distribution p
(h∣ e) that interests us. Instead of searching over all possible posterior
distributions, we can then approximate p(h∣ e) by finding a member of Q
that is, in some sense, closest to p(h∣ e)”
and more things that maybe you would find interesting. I’m sorry for the does and the do this is something I don’t get to internalize and I always have to think about when to use them.I hope you post my moderated commentary (and this one too) I don’t know why wouldn’t you?
I would understand if you don’t post my comment about hallucinations because it is a mess how I write here and there and no one would ever be able to keep track of this whole discussion (despite I guess it is in moderation because of the html links). Tbh I don’t even want to scroll down and see if you answered my last comments about psychiatry because it is something that actually hurts me when I say how most people completely dissaprove my vision of the psychiatric practice. I don’t hate them (the psychiatrists), I do even think their work is useful and necessary, but they have hurt me a lot and I feel compelled to express my opinion when I get the opportunity, but it hurts me a lot too when I see most people don’t even care about what I say.
Hey pharmakoi,
thank you again for all your comments and paper suggestions. Unfortunately, I have so many unread books and papers on my desk (mostly concerning other fields such as recent developments in deep learning) that I can’t promise you that I’ll be able to read those papers on hallucinations and psychosis any time soon, although they seem interesting.
Please also note that I’ve written this article on predictive processing almost three years ago, so I’m not really in the loop on this topic anymore, even though I’ve found revisiting it due to your comments very refreshing.
Concerning Hohwy, I don’t have a copy of his book at hand, so I can’t check what he’s saying for myself, but to maybe take a step back: In what way do you find this technical detail about the marginal likelihood particularly relevant for the larger picture here?
Concerning mental illnesses, your criticism of psychiatry seems too generalizing and too subjective for me to respond. I don’t even know what country you’re from or what illnesses you’re talking about. For example, what evidence do you have in mind about the case of a woman suffering from chronic depression or an eating disorder having “poor civil rights” in Austria?
I´m gonna requote what I wrote way before because maybe you are not in the mood of going back through all this, this is something I didn’t understand at the moment I posted it but now I do and I see my intuitions about he importance of dividing by p(e) in order to make inferences were kinda correct, even if we can only do “as if” by applying variational bayes (I think it doesn’t only serve to choose the hypohtesis that would maximize the posterior but to get confidence intervales and the variance and maybe another things that I tried to explain in a less formal way. Despite it can be that this method is used here because MAP estimates are not correct for the kind of functions used in this approach or something like that, since predictive processing doesn’t use cost functions, but this is sommething that is still out my reach).
” Making inferences about causes (e.g. the
most likely cause, or how certain we are that the cause falls
within some interval) depends on some representation of the
relative probabilities of values the causes can take. This
entails representing the probability distribution or density of
the causes. The key density is the conditional or posterior
density that summarises the likelihood of any cause given the
input”
and
“Representational learning is framed in terms of estimating
probability densities of the causes. This is referred to as
posterior density analysis in the estimation literature and
posterior mode analysis if the inference is restricted to
estimating the most likely cause. The mode of a distribution
is the location of its maximum. Although density learning is
formulated at a level of abstraction that eschews many
issues of neuronal implementation (e.g. the dynamics of
real-time learning), it provides a unifying framework that
connects the various schemes considered below. It is
important to appreciate the distinction between simply
estimating the most likely cause (i.e. mode) and the broader
problem of inference. Inference entails estimating the
conditional density not just its mode”
https://www.fil.ion.ucl.ac.uk/~karl/Learning%20and%20inference%20in%20the%20brain.pdf
That is, as you can see you are right that in general terms MAP estimates are inferences, but in the last sentence Friston says that ” inference entails estimating the conditional density not just its mode”, so as I said before, in the predictive processing framework and under the free energy principle (which is what we are talking about here) not all estimations are inferences, but only those that are about the PDF of the posterior.
I don’t know why you don’t approve my commentary about hallucinations and delusions of agency I think it was kinda interesting and I took my time to insert the html links to those two interesting papers on the topic, I think I’ll try to post it below were we started the conversation about that, maybe one day you will let it be.
I don’t find that of Howhy important for the larger picture but you tried to defend him even by saying things that were not correct at all, so I guess you were acting in accordance to the argument from authority what I dislike a lot even if I’m only “another anonimous crazy guy on the internet” and Howhy is an important figure on the field.
About mental illnes and psychiatric practice I said somthing very specific that could apply to any country in the world, that about the lack of insight and how such a subjective diagnose serve to strip you out of your rights. I don’t know, I told it, it is impossible for me to enumerate each case as you can imagine and there are very extreme situations in which the repressive partenalism of psychiatry is actually justified, but in most cases it is not, at all.
By the way I’m from Spain, I don’t know if for you that is a third world country or something like that, but I think civil rights here are not so different from Austria.
I’ll try to explain you too about the involuntary internment and how little possiblities you have to fairly defend your civil rights to not be interned in those situations, which is something very similar across developed countries although maybe with subtle differences. I’ll do it another day because when I get out of the very technical fields I use to read in English it is very difficult for me to find the correct words to express myself.But it is important to note that psychiatirc diagnoses and practices and the theory behind them are in a lot of cases very poorly objective and very imprecise, the most they have there are some empirical facts more or less validated and a lot of messy, imprecise, especulative and sometimes incoherent theorization. But even tho, the word of the psychiatrists is always the law and they can diagnose whatever they want with just a blink, they don’t need to give biological proofs or even to use the poor psychometric tools they have because they don’t even know almost any vaild biological marker for their diagnoses, and because, as I said, their word is the law and in the best case you are only gonna get the assitance of a lawyer to defend you but not of an external and impartial psychiatrist minded to apply all the best measure tools (wich are still very poor)that could refute or at least give some validity to your diagnose.
In the field of the social welfare I think there must be bigger differences between countries. I’m sure in some places you can still get to have a decent life when being diagnosed of a mental illness, what in Spain becomes very difficult.
About that of the involuntary internment I know there are some countries where, after being internated for like some months, it is required the evaluation of an external psychiatrist. But he probably only has to give his clinical opinion, and I think there is no country where you can be assisted by an external psychiatrist in the moment of beinig internated.
My opinion about psychiatry is that it is still in the state of what Khun calls prescience (you know, they have such prescience that they can even predict with which foot are going to wake up their patients every fucking day) and psychiatrists shouldn’t have the authority that they have over their patients. Anyway insofar we are talking about human freedom I don’t know if we should think about that as an essentially scientific problem. At least nowadays we don’t say sexual orientations are mental illnesess, unlike 40 or 50 years ago for homosexuality and like, how much? 5 years? for transexuality… it would be hilarious if it wasn’t so sad and harmful.
About if the discussion around Howhy’s commentaries is important for the larger picture I would say, however, that I don’t exactly know what do you mean whit larger picture, but what is clear to me is that dividing by p(e), or estimating the posterior probabilties (what to some extent is the same) is not irrelevant at all, as Hohwhy seems to say. Indeed I think the free energy principle is derived from the application of variational bayes (reducing free energy is the way to approximate the PDF of the posterior in variational bayes) so maybe it is not only important but essential for the larger picture.
Anyway even if that is not important for the larger picture, as they use to say, devil is in the details.
Don’t take me wrong with that of the absence of biomarkers, there are experimental findings that correlate strongly with different psychiatric conditions, but they are not valid for diagnose. They serve mostly to theorize about the possible causes of each mental illness, to justify yourself for what you do to your patients (sometimes kinda sadistic), to appear in your local newspaper like if you were a great researcher for making an experiment that have been replicated like a million of times or to put a big poster in your consultation to try to impress your patients, between others.
Well tbh I remember in my involuntary internment , which was totally innecesary and somehow sadistic, because I had decided to not protest anymore after being actually devastated (this is devastating and not the mental illness itself as some sensatioalists researchers like to say at the beggining of their papers about mental illness) by the (mis)treat I received because of asking them to let me go out and sitting in front of the door as a sign of protest. So as I was saying, I remember when the process for the involuntary internment (the psychiatric said to me, with no little satisfaction in her face, “so we are gonna make things well, I’ll make you an involuntary internment”) a forensic psychiatris came to see me and she asked me nothing in about five minutes, I told her I have had some crazy tantrums where I did hurt (not tried to) no one and I had thought in suicide (and because of that and a panic attack and the curiosity and the loneliness and the infinite ingenuity I had I decided to go and internate myself) and she did put that kind of face like “oh, I’m so sad for this guy, he is so bad” (this kind of hypocresy and autocomplacency I do actually hate) but didn’t asked me at any point if I wanted to get internated.
So the psychiatrist said in her inform to the jury (just another big lie) I had become violent (violent me?… they were actually violent, that’s the true and what I said to the jury, who just smiled and firmed the papers he had to firm). And she didn’t say me I could ask for the assistance of a lawyer when I had said repeated times that I wanted to denounce them while they dragged me over the floor of the hall and tied me to the bed (just for sitting in front of the door…). Even more, to justify all she did she said I had lack of insight, when I told her from the beggining that I was psychotic and had delusions and identified my delusions as such, what created me no little legal problems that ended in a real suicide attemp…after that she said one time that she did bad, the she didn’t know the kind of patient I was and that she didn’t control the legal issues…(so why you make me an involuntary internment and say I have lack of insight and lie and say I was violent and seem to get a sadistic satisfaction in doing all that to myself, like when she said in a permission to go out the weekend “If you try to escape and don’t come back I’ll call the police”…with no little sadistic satisfaction in her face)
So now I actually have devastating negative symptons, because of what they did to me, because of asking for help and receiving just repression, violence and hate. That’s what psychiatry means to me, a big lie, an enormous shit that I don’t even want to hear about.
The most curious thing about the involuntary internments is that, even as painful as they can be, you end developing some kind of Stockholm syndrome and (to some extent because it is a very contradicting experience) enjoying being recluded. Indeed I internated myself voluntarily another time, but this time I knew how I had to act so they didn’t make me an involuntary internment (because the first time I internated voluntarily but then I wanted to go out and that was like opening Pandora’s box).
Now I know I would never internate myself voluntarily again. If you don’t do it you will probably never have to see the dark side of psychiatry (despite I have some serious doubts about the healthiness of some or most of the psychiatric medications; antipsychotics, for example, could produce orbital prefrontal syndrome and be some kind of chemical lobotomy, beside their well known and officially accepted side effects)
About how good is to be recluded in a psychiatric institution I can tell you that a woman who had been in jail too, told me once that it was better to be in jail, that she had more freedom and could do more things there.
I remember for example someone who went there with the hope of receiving help and maybe some kind of rehab, and how he complained about being drugged and sleeping all day (indeed the best you can do when you are there is to take your drugs with no complains and try to sleep for the most part of the day).
About antipsychotics I have to say that according to hodology, neurotransmission and mostly neuropsychology and empirical observation, it is clear to me that they produce some level of orbitofrontal deficit. What I’m not sure is if it is something symptomatic or if in the long term they are actually like a chemical lobotomy.
I appreciate you sharing your experience, pharmakoi. I don’t know how common experiences similar to yours are: you seem to describe a case where the voluntariness of internment is more intricate than I would naively imagine.
Personally, I know of a recent case that happened among my relatives, involving a suicide attempt with severe self-harm. After hospitalization, the person was released to live at home again, despite fervent pleads for temporary internment by close relatives; however, the law wouldn’t allow it. So at least in this case, the civic right of autonomy prevailed.
I think I get since E is the same for all causes the higher the joint probability, the higher the mode. But I think this wouldn’t be called inference in this context.
the highest joint probability will be the mode, the higher the joint probability, the higher the posterior. (mode is or it isn’t but can’t be an higher mode than another).
and this
https://www.ncbi.nlm.nih.gov/pmc/articles/PMC3637647/
• Descending messages in the somatomotor system are
therefore predictions of proprioceptive input and not
motor commands.
In the somatomotor system, however, proprioceptive predictions should
not be corrected but fulfilled, by the automatic peripheral
transformation of proprioceptive prediction errors into movement.
…
A proprioceptive prediction error can be generated
at the level of the spinal cord by the comparison of
proprioceptive predictions (from motor cortex) and
proprioceptive input. Sources of proprioceptive input
include muscle spindles (via Ia and II afferents), Golgi
tendon organs (via Ib afferents), and articular and
cutaneous receptors. The prediction error can then
activate the motor neuron to contract the muscle in
which the spindles—or other receptors—are sited: this
is the classical reflex arc (Figs. 1, 2). In short,
peripheral proprioceptive prediction errors are (or
become) motor commands.
page 4 or 614
Hey pharmakoi,
thank you for your comments and your paper suggestion!
In the paper, Adams et al. write, “In brief, under active inference, descending signals do not enact motor commands directly, but specify the desired consequences of a movement” (p. 615) and “The crucial point here is that active inference works by providing proprioceptive predictions (from a forward model) to reflex arcs (the inverse mapping), which automatically generate motor commands” (p. 635).
This, I think, best describes how active inference relates to motor commands. I concede that my use of ‘motor commands’ in the sentence you quoted from my article can be misleading and I’ve replaced the term with ‘motor signals’. My discussion about action as active inference, however, does seem to be well in line with the papers you mentioned.
Hey thank you for your answers!
I have seen this about posterior expectations and prediction errors
In neural network terms, equation (A 3) says that error-units compute the difference between expectations at one level and predictions from the level above .Conversely, posterior expectations are driven by prediction errors from the same level and the level below.
here https://www.ncbi.nlm.nih.gov/pmc/articles/PMC4387510/
Tbh I don’t understand the mathematical formulations at all xD, but from what I saw in another schema I think that the thing is that in the mesagge passing from one level to the above posterior expectations are used as sensory input from the lower to the upper level (except in the first level where sensory input is actually sensory evidence). I’m not totally sure about this but thank you again for your answers you made a good work whit this resume of predictive processing.
I thinks this clarify what I was saying
“Prediction error is simply the difference between (bottom-up) sensory input and (top-down) predictions of that input. This minimization rests on recurrent neuronal interactions between different levels of the cortical hierarchy in which bottom-up signals relay prediction error to higher levels, which respond by changing expectations to provide better top-down predictions (optimization of the posterior expectations or beliefs). This is a relatively simple and neuronally plausible scheme for which there is a large amount of circumstantial evidence.8 This evidence suggests that superficial pyramidal cells in the upper layers of cortex (green triangles) compare expectations at each hierarchical level (or sensory input at the lowest level) with top-down predictions from deep pyramidal cells (black triangles) of higher levels.”
https://www.ncbi.nlm.nih.gov/pmc/articles/PMC4166359/
But you explained it well, it is just that every level is influenced by the lower levels and influence the levels above.
and the opposite haha, there is descending and ascendind influence between levels.
I’d like to see if this image appears, it is about the message passing bettween levels
Here we see that posteriors from one level are passed to the error neurons on the level above, and the comparison with the prior of this level yields a PE. This is, the comparison between priors and posteriors is between levels and not within the same level or the same prediction unit (deep pyramidal cell).
For if image doesn’t work, it is here
https://www.semanticscholar.org/paper/How-Prediction-Errors-Shape-Perception%2C-Attention%2C-Ouden-Kok/8b6d76e63c4dea9f85f42dbd71ab5b92b1523e4d/figure/1
“Agency arises from a good enough fit between your self-model’s predictions and exteroceptive input.12 If the prediction error is too high, you feel like you’re not in control of your actions.”
“Compare this to Andy Clark’s conceptualization of agency: “When proprioceptive prediction error is highly weighted yet suitably resolved by a stack of top-down predictions (some of which reflect our goals and intentions), we feel that we are the agents of our own actions” (2016, Surfing Uncertainty, p. 158).”
About this, first, propioception is not the same as exteroception, and second, I think you should read about sensory attenuation and how it relates to delusions, hallucinations and beliefs about agency. In pages 218-219 he says something about that (7.9 Disturbing the Network (Sensory Attenuation).
About the dark room problem I think what you say is true but I think having in acount the infomax principle and the mutual information maximization is enough to see that an organism with a tendency toward free-energy minimisation would never like to be in such a predictable enviroment. In mutual information if the entropy of one variable is 0 the mutual information is 0 too. And such enviroment would have 0 entropy.
Anyway I have been reading most of your resume and it is great, I have no doubt you know much more about this topic than I do.
I have been looking at the paper that you recommend about source monitoring and psychosis, seems pretty in interesting.
I think the only thing that is not correct about what I quoted before is that what is needed for agency to arise is a good fit between your self-model’s predictions (corollary discharges) and propioceptive input, not exteroceptive input.
Naturally, proprioception is not the same as exteroception. This is why footnote 12 starts with “Compare this to Andy Clark’s conceptualization of agency…” in order to highlight that there could be a disagreement here between Braun et al. (2018) and Clark (2016), since the former link a sense of agency also to “external sensory consequences” (p. 6). However, I grant you that my definition of ‘agency’ in the article is incomplete: it likely requires both exteroceptive and proprioceptive signals, whereas ‘ownership’ is focused more on the latter.
Yes you are right, it requires discriminating between exteroception and mental imagery too and propioception might be more related to ownership than to agency.
What I would like to say, in a more personal level, is that from my perspective (which is that of a mentally ill) social activism defending the social wellfare and the rights of mentally ill people would do more for us than psychiatry has done in the last 200 years or is gonna do in the next 200. Indeed the clinical obsession with supressing the positive symptons overall do more harm than good. And if you don’t show too much concern about the suffering these symptons must produce you from the perspective of the clinician you are said to have lack of insight and all your rights as a citizen are stripped or taken away fron you.
The civic rights of mentally ill people are an intricate issue which, I opine, would have to be discussed in the context of specific events, diagnoses, and polities.
Yes but their civil rights are still very poor, you just seem to be avoiding the question, as most people do since they like to think that we live in a very fair world and that psiquiatrists allways know very well what they do.
About the comment before, I wanted to say at any cost, not overall.
And a new insight into the topic:
professionals from the fields of mental illness (from clinicians to investigators) would do better attending more to the phenomenological descriptions that their patietns give about their own experiences, and not just trying to impress their colleagues whith their big equations and experiments and their bigger bank accounts. But of course, precision about that source of information is set very low in their cortical hierarchies. Would be to hard for their self-esteem to realize how little they learned in the career, and how little they know about the reasons that move their patients to act as they do.
“Yes but their civil rights are still very poor, you just seem to be avoiding the question”
Unless you believe that every person with any mental illness is treated unfairly by the legal system of any country in the world, I am not avoiding the question.
Hehe ok, but it doesn’t seem so intrincated to me, and what you say is actually true, all mentally ill are treated unfairly by the legal system, even if most times they don’t have to directly face this unfairness.
Tbh, I think mentally ill should have more civil rights than normal people and not less. In one hand (and I will explain why this is still not ethic, but would be better than the actual situation) they should have specificic civil rights designed in order to protect them from the specific violations of their civil rights that they have to face and which are not scientifically justified. But even I would say more, they should have exactly the same civil rights as all people and face the same limitations on their freedom that normal people have, that is, only when they actually commit delictive behaviors. Other thing wouldn’t be enough justified scientifically and even if it was it would not be ethic to limit their freedom based on this (we don’t force people to receive the vaccine for covid or to give up smoking, and we don’t do this because it is no ethically justified).
In the other hand they should receive another specific civil rights, those intended to attend their need of social care programs. They don’t need to be punished in order to justify the application of that social care programs.
I’m not gonna try to enumerate each specific case and how to resolve it, because it would be impossible and it is a problem that must be resolved by those dedicated to it when facing specific situations at the field. In this sense you have reason that it is an intrincate problem that depend on specificic diagnoses and policies and so on. But I think I have outlined how most of those problems arise from a more general mistake in the way we conceive the mentall illness and its required treatment.
Well don’t aprrove my last tries they appear way down, but you are not so right, propioception is very important for agency, and if you post my comment were I link those papers (specially that about the computational anatomy of psychosis) maybe someone interested could understand why.
This nested structure is kinda crazy when you want to answer a comment at some points
.”sense of agency (SoA) refers to the experience of initiating and controlling an action”
Braun et al. (2018)
That is exactly what they explain in the paper I mention and it has a lot to be with propioception and sensory attenuation.
better here maybe haha I will never see that comment posted I guess so why not try again…
About hallucinations in general there are some sort of explanations that talk about noisy sensory prediction errors and strong priors, as you can see here but in the case of auditory hallucinations in schizophrenia they usually have to be with inner speech so they are better explained through aberrant precision of propioceptive prediction erros due to a loss of sensory attenuation which also leads to delusions of agency, as you can see here. If hallucinations and agency are always related to forward models, I don’t truly know, but as you can see psychiatric explanations in general don’t stand out for their precision.
I want to see at least whre does it appears in this nested structure, since you don’t want to approve it ??