Romain Brette

@romainbrette.bsky.social

857 Followers 130 Following

I think about the biology of cognition in brains and in protozoa. New book: The Brain, in Theory. http://romainbrette.fr/

dy/dt = -a*y + K[x(t) - y(t-tau)]
Anticipatory relaxation dynamics (ARD) for a smoothed random signal. a) Section of the signal to be predicted, x, and the predicted ARD signal, y. The rescaled signal λy, with λ a scaling factor, is shown, too.
Second, strict identification as in “organisms are machines” or “neurons compute” is a great source of confusion. In what sense are living organisms machines? Are they made of parts? Assembled? Are they mechanical? Are they lawful? Are they engineered? These are very different claims. If one needs to carefully explain in what exact sense organisms are machines, and if different people pick different features, then organisms are not actually machines. They are somewhat similar, and somewhat different. This acknowledgment is crucial for conceptual clarification.
The machine view of life is so engrained in scientific culture that it seems very difficult for many scientists to imagine that living organisms could be something else than machines, or that brains could be something else than computers – “what else could it be?” seems to be a valid scientific objection (van Gelder, 1995). As I pointed out in the introduction, molecular biologists and cognitive scientists appear to have replaced God by the engineer. The Church has disappeared, but otherwise there is a stunning intellectual continuity: God was a watchmaker, and now life is clockwork. Is there really nothing in the world apart from human artifacts and magic? Couldn’t life be its own thing, rather than a kind of clock?
Reconsider the quote above: “an experience of pure darkness is what it is by differing […] from […] other possible experiences”. The claim is superficially appealing: we appreciate the particular quality of darkness in comparison with different experiences we have had before, such as lightness. But this works if we already had previous experiences. In other words, it presupposes that we are able to experience and to remember. One cannot define consciousness by presupposing both consciousness and memory. Thus, the fact that Integrated Information Theory is a panpsychist theory of consciousness is not an insight of the theory, but an assumption: the theory presupposes that there is a cognitive agent behind every state configuration, without explaining either cognition or agency. Both information by reference and information as difference define information as a potentiality: what an observer can infer from a signal. These are homuncular views, which take as prerequisite what they are supposed to explain, since inference requires prior knowledge (of external things, or of alternative possibilities). One cannot explain knowledge by assuming knowledge.
These phenomena have been termed “representational drift” or “dynamical representations”, but with the same logic, it could be claimed that the decimals of π represent the decimals of e, only with a representation that changes from digit to digit (Figure 5.7). It becomes rather unclear what is “representational” at all about such time-varying correspondences. Thus, the systematic analysis of the activity of neural populations has confirmed that “neural representations” are associations between experimental dimensions and neural activity that are stimulus-dependent, context-dependent, task-dependent, relevance-dependent and finally, time-dependent across various timescales, from milliseconds to seconds to days. In fact, the “represented” dimensions also depend on the dimensionality reduction method (Langdon et al., 2023). At this stage, the representational and coding terminology starts sounding like epicycles.
In Greek mythology, Theseus managed to find his way out of the Minotaur’s labyrinth thanks to Ariadne, who gave him a thread he attached to the door. To escape the labyrinth, he just had to follow the thread back to the entrance. Out of the labyrinth, one could solve the path integration problem by tightening the thread, so that the thread takes the shape of the shortest path to the starting point, a straight line. Solving this problem involves no computation. Whatever Theseus’ position, the action that solves the problem is always the same: pulling on the thread. The procedure is neither computation nor information processing. Instead, it relies on interacting with the environment. In neuroscience and cognitive science, this is related to the concept of situated or embedded cognition. The interesting point about this example is that everything we have said earlier about the computational structure of behavior remains valid. It is still the case that the homing vector is obtained by iterating the operation x(t+dt)=x(t)+v(t)dt, and yet neither the brain nor the organism computes the homing vector. This is not just a question of boundary, i.e., of whether cognition occurs within the boundaries of the skull or is extended (Clark and Chalmers, 1998), because the thread does not implement the vector update operation either.
To build an electronic computer, stable states are constructed out of dynamical systems (flip-flop circuits), such that the dynamical nature of the underlying hardware can be entirely ignored: computation occurs as transitions between computational states, entirely shielded from the dynamics of electrons. Such shielding does not exist in brains. Thus, biological cognition cannot be reduced to elementary computations, supposedly implemented by neurons. Rather, computation is an elaborate form of cognition.
Efficient coding theory postulates that neurons encode stimuli such as images in such a way as to transmit the maximum amount of information (Barlow, 1961; Olshausen and Field, 2004). This works by reducing redundancy, as in data compression. For example, in the Lempel-Ziv-Welch (LZW) algorithm, repeated patterns are assigned to shortcut symbols. In the same way, spikes are postulated to encode common patterns. This might sound like a reasonable principle, but the LZW algorithm is useful only because the table of correspondence between shortcut symbols and data patterns is also communicated, in addition to the compressed data. Without the table, one cannot decode the compressed data. Unfortunately, neurons do not communicate tables together with the encoded messages.
The dynamical landscape of reproduction. The ball rolls down to a valley, and returns there after a perturbation.
A Paramecium with a spine reproduces. The spine eventually disappears.
a. A Paramecium doublet, with two mouths. b, Invariant reproduction of a singlet and doublet Paramecium.
Jumping spiders recognize mates by aligning their linear ret­inae with the mate’s legs (adapted from Land, 1969)
Thus, when we see a still image, the actual sensory input is not still at all. It is the still image that is a mental construction, and not the input, which is dynamic. In the same way, one of the properties of images such as photographs is translation invariance: translating a photograph does not change the perceived identity of the person on the photograph. This has motivated a powerful technical trick in artificial neural networks known as convolutional networks (LeCun and Bengio, 1998): connections between layers are enforced to be translation invariant (via “shared weights”), which drastically reduces the number of adjustable parameters. But translation invariance does not exist at the level of the retina, for the retina is not homogeneous. The eye has inhomogeneous optical aberrations, the fovea has a much finer spatial resolution than the periphery and different color sensitivity, and the visual field is sampled by a mosaic of cells of different types, meaning that two neighboring cells typically have very different properties (see (O’Regan and Noë, 2001)). There is no translation invariance at the interface between the nervous system and the optical environment. Again, this invariance appears at the psychological level: if the photograph is moved, one sees the same person. But this does not occur physiologically by having shared synaptic weights, or translation-invariant weights.
When the scientist wants to predict what happens when you reduce the volume of a container at fixed temperature, she uses the constraint of the ideal gas law to infer that the pressure will increase. Of course, we might want to explain the phenomenon in terms of the collisions of atoms, but these are narratives rather than science. Newton’s laws are also formal expressions of constraints. For example, the second law says that the acceleration of an object times its mass equals the total force exerted on that object. From this constraint, we can infer the trajectory of an object, for example in a gravity field. Dynamics is secondary to the constraint on acceleration. The various conservation laws of physics are examples of general constraints. To know whether an object would float or sink, we do not examine the motion of individual molecules in the fluid. Instead, we use Archimedes’ principle, which stipulates that an immersed object is subjected to a force opposite to the weight of the fluid that the body displaces. This is a global relational constraint (relation between the object and its environment), and yet, physicists are generally at ease claiming that the buoyant force causes the object to move up or down in the water.
Thus, the systematic analysis of the activity of neural populations has confirmed that “neural representations” are associations between experimental dimensions and neural activity that are stimulus-dependent, context-dependent, task-dependent, relevance-dependent and finally, time-dependent across various timescales, from milliseconds to seconds to days. In fact, the “represented” dimensions also depend on the dimensionality reduction method (Langdon et al., 2023). At this stage, the representational and coding terminology starts sounding like epicycles.
This is what an engineer would typically do: to build a robot, one would design perception modules, typically one per sensory modality, planning modules, and motor control modules (but see Brooks (1986) for an alternative way of building robots). The reasons for this modularity are cognitive and social. Cognitively, this allows one to focus on one simple problem at a time, following Descartes’ rule of analysis (Descartes, 1637): to divide the problem into smaller, more intelligible, pieces ("Diviser chacune des difficultés que j'examinerais, en autant de parcelles qu'il se pourrait et qu'il serait requis pour les mieux résoudre"). Socially, it also allows work to be divided between different people, or even different companies: one in charge of each module. These modules can then be reused in different contexts, different machines that combine modules in new ways. For this to work, the input-output properties of modules are well-specified, for example by stating that a module’s output represents sound direction. What the output represents should not depend on the context in which the module is embedded. Thus, there are excellent reasons for engineers to use modular design. But brains are not made by people, let alone by human societies. However appealing the functional argument for computational modularity might be, it is hard to reconcile with the known physiology and anatomy of nervous systems.
Of course, the rhetorical trick of active inference is precisely to call the goal a prediction. So, the runner predicts that she ­ will arrive quickly at the finish line. In this way, the person acts so as to meet her “prediction.” But this is a rhetorical move with no explanatory power. It can work only as long as one decides to name as “prediction” exactly the goals of the person and forgets about all the other predictions that she might make. For example, we might grant that the runner predicts that she ­ will win the race, although that is obviously false for many self-aware mediocre runners. But she would also predict that the race ­will start soon, that she might fall while ­ running, that she might lose the race. The runner might also see the clouds and predict that it ­ will rain. In any situation, you can predict many ­ things: some events that you want to happen, ­others that you want to avoid, and yet ­ others that you might not care about. Therefore, stipulating that you act so as to meet your predictions cannot work, since ­ there are several predictions at any given time.
Consider the following thought experiment inspired from the TV series Bewitched (Brette, 2019b). In that series, Samantha, the ­ house­ wife, is actually a witch. She can twitch her nose and every­ one freezes except her. Then she twitches her nose and every­ one unfreezes, without noticing that anything happened. For them, time has efectively ­ stopped. What happens during that time is not a continuous experience of the same ­ mental state. Rather, ­ there is no experience, no ­ mental image at all, and it is as if time had jumped from one instant to another. This is because an experience is not a ­ thing but an activity (Erdin and Bickhard, 2018; Zahnoun, 2020). Despite the common terminology, a “­mental state” is nothing like a physical state.
Cascade by MC Escher. A perspective illusion where water appears to run perpetually around a closed path with a cascade.
This is more than an unfortunate manner of speaking. Linguists have pointed out for a long time that language shapes our thought, in par­ticular­ metaphors (Lakof and Johnson, 1980). In this case, conceptualizing information as a ­ thing hides a fundamental feature of perception: it is a constructive­ process, not a ­ matter of extracting and filtering concepts from the environment. Consider for example a nose. A nose can be big or small, pointy, hairy, and so on, but what kind of physical object is a nose? A nose has no clear molecular delimitation, and it even includes holes. The delimitation is one made by our cognitive systems. More generally, objects are cognitive concepts­ shaped by our interactions with the environment. Therefore, contrary to Marr’s claim, to perceive the world is not a ­ matter of finding out what is out­ there, of inferring a preexisting objective description hidden by our ­limited sensory apparatus or degraded by “noise.” Rather, perceptual objects are cognitive constructions of subjects interacting with the world. This fundamental point dis appears in the mainstream view of information: information by reference.
Linguistically, talking about “information” as if it ­ were a ­ thing is a case of hypostatization, as Zahnoun has pointed out for the related concept of representa tion (Zahnoun, 2020): we use the grammatical category of nouns for something that is not substantive, like an action or a property, as when we say: “I am ­ going for a walk.” ­ There are no such ­ things as walks in the world. In the same way, we say that we “acquire information”: it simply means that we inform ourselves, not that ­ there actually is some kind of epistemic phlogiston that we pick up in the world and put in our brain. Surely, you need letters to read, just like you need a ground to walk on. But the ground is not a walk.
Machines are defined by their functional specification, the relations be- tween their parts, in­ de­ pen­ dently of their material existence. This is a key aspect of computationalism: cognition and consciousness are determined by the relations between functional states, at an abstract level, in­ de­ pen­ dently of the material “implementation” of ­ those relations. In other words, the “hardware” plays no role in the specification of the “software,” it only supports its execution. But as we have seen in chapter 2, this is not the case of living systems, ­ because the normativity of living systems is grounded in the physics of this world: it is ­ because of thermodynamics that living systems need to feed, and therefore that they have goals, that they take actions, that they anticipate; in other words, that ­ there is such a ­ thing as “be­hav ior.” Therefore, Marr’s “computational level” is in fact grounded in the material level, which is then improperly called “implementation.”
An elementary example is a clock. We can certainly say that a clock implements a computation, namely counting. But this is not what makes it a clock. What makes it a clock is that each tick takes exactly one second, and this is not obtained by computation but from the physical properties of the device. For example, in a traditional clock (figure 4.10), it is the dynamics of the pendulum that sets the tick frequency.
The machine view of life is so ingrained in scientific culture that it seems very difficult for many scientists to imagine that living organisms could be something else than machines, or that brains could be something else than computers—“what else could it be?” seems to be a valid scientific objection (van Gelder, 1995). As I pointed out in chapter 1, molecular biologists and cognitive scientists appear to have replaced God with the engineer. The Church has disappeared, but otherwise there is a stunning intellectual conti- nuity: God was a watchmaker, and now life is clockwork. Is there really noth- ing in the world apart from human artifacts and magic? Couldn’t life be its own thing, rather than a kind of clock?
The engineer starts from the desired functionality, then designs and tunes a structure that fulfills this functionality. In contrast, evolution is a blind historical process: a structure gets created by chance, and might then be conserved and possibly refined if it serves a useful functionality; or some existing structure might change and take on a new functionality. A central argument of Darwin in The Origin of Species (Darwin, 1859) is precisely that we find striking anatomical similarities across different species despite the diversity of lifestyles, which only makes sense if we assume a common ancestor: “What can be more curious than the hand of a man, formed for grasping, that of a mole for digging, the leg of the horse, the paddle of the porpoise, and the wing of the bat, should all be constructed on the same pattern, and should include the same bones, in the same relative positions?”.
Yet, if we were to explicitly ask a modern neuroscientist whether the brain is actually an engineered device, she would certainly strongly object. Brains are not the result of intelligent design. This is a religious view of life that has been discredited by Darwinism. Why then are we to “reverse-engineer” brains, if brains were not engineered in the first place? This terminology is typically excused by adding that brains are engineered by evolution, not by God. But Darwin’s insight is precisely that evolution is not a case of engineering. Engineering is the use of knowledge to solve technical problems. It presupposes an external mind that plans and assembles machines according to a preexisting goal. But evolution has no goals, plans or knowledge; in other words, it is not an engineer.
Remember that classical neurocomputational models of visual perception transform an input image into a mental symbol, such as a face label. This is because psychologically, we see a stable image and we are led to believe that there must be a stable physical entity in the brain that stands for that image, a “neural representation”, which is then further processed (Ramsey, 2007). But when we looked at the biology of seeing, we realized that there is nothing like an image in neurons of the retina: the visual field is dynamic and the activity of neurons is transient, made of electrical spikes. The stability of the visual world is a property of perceptual activity, not an intrinsic property of raw perceptual material.
Samantha in the TV series Bewitched
Consider the following thought experiment inspired from the TV series Bewitched (Brette, 2019c). In that series, Samantha, the housewife, is actually a witch. She can twitch her nose and everyone freezes except her. Then she twitches her nose and everyone unfreezes, without noticing that anything happened. For them, time has effectively stopped. What happens during that time is not a continuous experience of the same mental state. Rather, there is no experience, no mental image at all, and it is as if time had jumped from one instant to another. This is because an experience is not a thing but an activity (Erdin and Bickhard, 2018; Zahnoun, 2020). Despite the common terminology, a “mental state” is nothing like a physical state.
Thus, in neuroscience, information is typically conceptualized as a sort of epistemic phlogiston: animals know because they have “information” in their brains, which enters through the eyes and moves up to higher cognitive centers. This is a fundamental misconception based on a reification of a cognitive process (Zahnoun, 2020). That is, what is actually a process (getting informed, understanding) is framed as the acquisition of some epistemic phlogiston that we call “information”.
A computation is essentially an effective method in a formal domain (as opposed to a domain of interactions with the world), generally in the domain of numbers, which can be applied by a human with just pen and paper – we will see in chapter 8 the significance of this detail. Clearly, a recipe is not a computation. […] A computation operates on data (not pasta). It has an input and output, and it implements a particular mapping from input to output by a sequence of steps that belong to a finite set of elementary operations on symbols. The latter restriction is crucial for the notion of computation. Otherwise, the concept of computability would be totally meaningless. For example, Alan Turing demonstrated that there is no program that can take another program as input and calculate whether that program eventually stops: the halting problem is not computable. Obviously, this theorem would be false if we were allowed to choose any kind of instruction (choose as an instruction the mapping that assigns a stopping / non-stopping label to programs). Therefore, a process can only be meaningfully called “computation” when it is composed of a finite set of elementary instructions. Otherwise, the process is just a mapping between different kinds of data, not a computation.
Thus, there are different computational concepts that need to be disentangled and discussed separately. First, a computer is a programmable machine (Brette, 2022a). As we have seen, living organisms are organized, but otherwise they are not machine-like. There is also no useful sense in which brains are programmable (4.1). But do brains compute? Here we need to be more explicit about what we mean by “compute” (Figure 4.1) (4.2). If we mean that some behavior can be described as programs in the broad sense of a sequence of predetermined actions, as in the “TV program”, then that is certainly often the case. If we mean that we can follow algorithms, in the broad sense of a sequence of fully specified elementary actions to solve a problem, then that is sometimes the case, as when we follow a cooking recipe. But if we mean computation in the sense of computer science, then we are referring to something much more particular, which operates in the formal domain with a finite set of rules. What we call “information processing”, the transformation of input data into output data, includes computation but also non-computational ways of solving data problems. When we examine the normal behavioral repertoire of organisms, we find that a large part is interactional and dynamic, including perceptual processes (4.3). Neither dynamics (4.4) nor interaction (4.5) can be meaningfully characterized as computation, unless by “computation” we mean “anything that solves a problem”. This undermines computational reductionism, the claim that all of cognition is ultimately decomposable into cognitive atoms, which are elementary computations. Nonetheless, an important part of behavior does not rely on interaction, and it is especially in those cases that neurons are said to “compute”. Unfortunately, we will see that the classical concept of “neural computation”, by which brain processes are said to implement a computational model, is flawed beyond repair (4.6).
Thus, in scientific practice, models are established from two types of constraints: the general laws of physics, the fundamental ones being microscopic, and constraints about the object of study, which operate at the scale of the object. Therefore, even when a macroscopic phenomenon is said to be “reduced” to fundamental laws, modeling practice is never entirely bottom-up: there are fundamental microscopic laws as well as macroscopic constraints. The very idea of reductionism is itself highly fallacious: of course, all physical phenomena cannot be reduced to the same few fundamental laws, since they are different phenomena. They can be reduced to the same general laws plus specific constraints. Often, these constraints are inferred from theories established at the scale of the object of interest.