A general theory of the brain should in principle help us to understand all features of the brain, at all levels from molecules to cells to systems to behavior, in analogy to the way that Darwin’s theory helps us understand all aspects of biology. It should predict which designs are better, meaning healthier and more intelligent, in a given environment. It should therefore help us to design intelligent machines, and to recognize which machines really are intelligent (ChatGPT is not). A general theory should also help us to understand mental phenomena and their relation to the physical. I believe that my theory addresses all of these and is therefore ‘general,’ although it is not as complete as it should be (and some of it is not yet published).
My theory states that the brain, and every neuron in it, is designed to maximize evidence for its own future existence. From this grand objective we can explain perceptual phenomena, such as illusions, and bottom-up and top-down attention, and the properties of neurons, including the signaling of prediction errors, pattern generation in motor neurons, Hebbian and anti-Hebbian learning rules, and reinforcement learning. The theory makes numerous testable predictions, from the general organization of information in the brain to the biophysical properties of synapses and ion channels in each neuron (see below). It has also generated several patents related to machine learning.
My theory is distinguished from other theories of brain function by its generality, its conception and mathematical measurement of evidence, its emphasis on “the neuron’s point of view,” and its biophysical detail. It combines objective Bayesian probability theory with three theoretical frameworks that were already established in neuroscience: Bayesian brain theory, information theory, and reinforcement learning.
All three frameworks revolve around information and prediction, but they have traditionally been studied in parallel, with almost no overlap. I do not think that any of them have an adequate definition and measure of information. In addition, none of the three theoretical frameworks arose from within physical science, and attempts to map them onto neural processes have typically been ad hoc and poorly justified, especially in the case of Bayesian brain theory.
Below I distinguish descriptive and prescriptive components of my theory, I summarize a particular study that tested one prediction of the theory, and I discuss a highly novel hypothesis that I would like to see tested experimentally.
Description and Prescription
It is always important to distinguish what is from what should be. In biology we must first describe biological structures as they are, and then we must prescribe what they should be if they are to survive. A general theory of the brain must have descriptive and prescriptive components. It must describe any brain, while also distinguishing a healthy from a pathological brain, and explaining how a brain (or machine) should be designed to be intelligent and thereby survive.
The descriptive component of my theory necessarily comes first. It says that any thing is evidence about its external environment and its own future, and it therefore predicts the state of its external environment and future. The ideal way to describe this is to mathematically measure the evidence in a cognophysical model of a thing with probabilities. This was part of my theory from the start (Fiorillo, 2008), but the mathematical method I originally proposed was naive. One of my primary goals in recent years has been to develop a more realistic cognophysical model of an observer, together with the correct mathematical measure of evidence.
When we use terms like “predict” and “evidence” and “knowledge,” we often mean accurate predictions based on a large amount of evidence/knowledge. When I say that any thing is evidence and predicts, I do not mean that it is a lot of evidence or makes accurate predictions, but only that it is at least some evidence that predicts something. This aspect of my theory is purely descriptive, like a physicist stating that each particle has a velocity. When I say that we can measure evidence with probability, this is analogous to a physicist who says we can use calculus to find the trajectory of a particle given initial conditions. None of this concerns what the particle should be doing, or what the evidence should be. It is purely descriptive.
The prescriptive component of my theory says that more evidence is better for survival. Survival requires that a thing has sufficient evidence, and thus makes accurate predictions. However, it does not need to have evidence about every aspect of its environment. It only needs evidence for its own future existence (“future reward” in the language of reinforcement learning). The brain acquires sensory evidence about each of many local aspects of the external world, and transforms it into evidence for its own future existence. The latter evidence is most concentrated in motor systems.
The prescriptive component of the theory makes numerous predictions, some of which are described below. The descriptive component of the theory provides the foundation for the prescriptive component, but it is not testable in the usual sense of the term. It just says that a thing is evidence, and that we can measure evidence mathematically with probabilities. The concept of “evidence” has no recognized physical definition, and it is not observable in a physical sense. We can argue about how to define it conceptually, and how to measure it, but we cannot design an experiment to test my definition and mathematical measure of evidence, or to test my proposal that a physical thing is evidence. However, we may eventually see whether my proposal and mathematical measure proves useful in explaining and predicting our observations, similar to the way that the concept of “force,” and the math of calculus, have proven useful in physics (note that a “force” is not observable, and if one were to redefine and “operationalize” it so that it is observable, it would lose all its explanatory power and become a redundant and trivial description).
Testing the Theory
There is a great deal of experimental evidence that is relevant to the theory, at all levels from psychology to systems to cells to molecules. I discuss some of this evidence in Fiorillo (2008) and Fiorillo et al (2014). My students and I confirmed certain predictions of the theory concerning the role of voltage-gated ion channels in sensory and motor regions (specifically T-type calcium channels in sensory and motor thalamus; Hong et al, 2014; Kim et al, 2015). Below I summarize another study that is exceptionally strong evidence because it shows that the theory alone was able to predict biophysical data with unusually high accuracy (Kim and Fiorillo, 2017).
Prediction Error and Learning
The theory states that every impulse of neural activity (spike) signals “prediction error” (though this is less obvious in neurons of motor systems; Fiorillo, 2013; Fiorillo et al., 2014; Kim et al., 2015). At every moment a neuron has a prediction of how much synaptic excitation it will receive from its external environment, and it only generates a spike when that excitation exceeds its expectation (Fiorillo et al., 2014). Therefore we can describe a spike as a “prediction error.” Prediction error is the new information that is needed for learning and communication, since these are only necessary when information is lacking and predictions fail. If we all know everything that there is to know about our environments, our predictions are perfectly accurate and there is nothing to learn or communicate (though this ideal is unattainable).
Spikes cause learning, and learning corresponds to a change in the number of ion channel proteins of a particular type in the neuronal membrane (here a “type” could correspond to a specific type of protein, or to a set of channels located at a specific synapse). Hebbian learning rules select ion channels (and thus synapses) that promote prediction errors, since prediction errors provide new information. But only certain types of ion channels are controlled by a Hebbian rule. Other types are controlled by an anti-Hebbian rule, which selects those ion channels (and synapses) that minimize prediction error. The two learning rules work in parallel to maximize the total amount of evidence that the neuron has about its external environment.
Minimizing prediction error allows the neuron’s spike output to maintain sensitivity to its excitatory synaptic input. The same principle applies to any well designed sensor (measuring device), such a balance scale for measuring weight. We want the angle of the arm to be highly sensitive to the unknown weight we wish to measure. We adjust the reference weight until the arm appears to be perfectly horizontal. The reference weight we choose to put on one side of the scale is our prediction of the unknown weight that will be on the other side. The angle of the arm then signals prediction error, telling us that the unknown weight was more or less than our prediction. Then we adjust our prediction by choosing another reference weight and repeating the process. A neuron does essentially the same thing. There are certain types of ion channels (most of which are inhibitory) that function like the reference weight, and they are controlled by an error-minimizing (anti-Hebbian) learning rule. There are others that function like the weight to be measured, and they are controlled by an error-maximizing (Hebbian) rule.
The theory says that neurons in a healthy animal, receiving typical statistical patterns of synaptic input, should make accurate predictions of their synaptic excitation. When predictions are highly accurate, there will be a positive prediction error in half of the instances (as in the balance scale analogy), which means that half of excitatory events will cause spikes (Fiorillo et al., 2014). We tested this prediction of theory by performing computer simulations of a neuron and mimicking its natural patterns of synaptic activity. This allowed us to predict the amplitude and decay time of synaptic inhibition that is optimal for counterbalancing synaptic excitation according to theory (Kim and Fiorillo, 2017). We then compared those predictions to experimental data, and found that the two were closely matched, as illustrated and described further below.
Analogy to Darwin’s Theory of Evolution
My theory proposes that every neuron has the same basic function, maximizing its evidence for its own future existence according to more or less the same learning rules. This may appear inconsistent with the tremendous diversity of neurons that we observe, but in fact my theory is analogous in this respect to Darwin’s theory of evolution. His theory says that all life could conceivably have originated from the same single cell, and that it was shaped by the same process of natural selection over time. Despite proposing this simple and unified origin, Darwin’s theory can nonetheless explain the diversity of life we see. His theory relies on just two sources of variability to explain diversity, one temporal and one spatial, and these two sources are inevitably present. There is variability in the form of organisms that is created with the passing of time (particularly due to sexual reproduction), and there is variability across local environments. Natural selection will inevitably shape this process, with the result that the surviving species tend to be well adapted to their local environments.
For example, hummingbirds and certain flowering plants rely on one another, and they evolved together. As a result, the length of the bill in a particular species of hummingbird tends to match the depth of the nectar in the particular species of flowers upon which the bird feeds. In this way Darwin’s theory can explain the diversity of bill lengths (and the depth of nectar). Furthermore, when we observe this matching between the structure of an organism and the structure of its local environment, it is powerful evidence in support of Darwin’s theory.
My theory should predict and explain neuronal diversity in an analogous manner. It says that neuronal diversity (both within individuals and across species) ultimately arises from the fact that each neuron develops and learns in a distinct local environment within the brain. The drawing below shows the diversity of neurons within the cerebellum (kindly made for this website more than 100 years ago by Dr. Ramon y Cajal). The synaptic inhibition in Purkinje neurons (the largest neurons in the brain) decays in about 3 milliseconds, whereas it persists for over 10 milliseconds in cerebellar granular neurons (the smallest neurons in the brain). The study described next shows that my theory not only explains this diversity, but accurately predicts the specific decay time in each cell type.
The Theory Accurately Predicts Synaptic Decay Times
Fast synaptic inhibition (mediated by GABAA and glycine receptors) exemplifies neuronal diversity, as shown in the figure above. It decays with a time constant that varies across different types of neurons, covering a range of about 2-60 ms. For example, it lasts for 2 ms in neurons of the cochlear nucleus, 15 ms in midbrain dopamine neurons, and 60 ms in the inferior olivary nucleus. Jaekyung Kim and I tested the ability of the theory, concerning the meaning of spikes as described above (Fiorillo et al., 2014), to predict and explain this variability (Kim and Fiorillo, 2017).
Using computer simulations we found that the amplitude and decay time of synaptic inhibition that was optimal according to theory depended on the average rate of synaptic excitation. The greater the rate of synaptic excitation, the higher the optimal peak amplitude of synaptic inhibition, and the faster its decay. We then compared the amplitude and decay times that were optimal according to theory with experimentally observed amplitudes and decay times, and we found a close match. The figure below shows that the theory predicted the observed decay time of synaptic inhibition with remarkable accuracy across 21 distinct types of neurons (types selected because data was available that allowed us to estimate average in vivo rates of synaptic excitation). It is extremely rare in neuroscience that biophysical properties can be accurately predicted from theory alone.
The theory also predicts that the density of every type of ion channel is controlled by a Hebbian or anti-Hebbian rule (that maximizes or minimizes prediction error, respectively), and that the density of the GABAA and glycine-gated ion channels that mediate the synaptic inhibition of relevance to this study should be determined by an anti-Hebian rule (to minimize prediction errors). There are many subtypes of these ion channels, with different subtypes having different rates of decay (although other factors can also influence decay times). The theory says that each neuron learns which subtypes best predict and counterbalance its synaptic excitation. Computer simulations showed that neurons could indeed learn the amplitudes and decay times that were optimal according to theory and observed experimentally.
Our study is analogous to the example above in which the bill lengths of hummingbirds provided evidence supporting Darwin’s theory. In fact our study provided stronger evidence for our theory than most studies of evolution provide for Darwin’s theory. Our theory not only explained the diversity of synaptic decay times, but predicted their quantitative values from first principles alone. In contrast, there are many cases in which Darwin’s theory is used to explain observations only post hoc. This means the scientists first had their observations and then searched for an explanation, and they naturally turned to the only big theory in biology. Synaptic decay times were measured decades before my theory was developed, but those measured values were not used to construct my theory, nor was my theory designed for the purpose of explaining decay times or any other specific observation.
An Untested Hypothesis: Hebbian Control of Voltage-gated Ion Channels
All of a neuron’s ion channels provide information to its membrane voltage. The purpose of learning is to select those ion channels that are most informative given the spatiotemporal patterns of its synaptic excitation. There is no theoretical rationale, nor any mechanistic justification, for the common view that ion channels at synapses are subject to Hebbian learning rules, but others are not. A well designed neuron should use associative learning rules for all of its ion channels (at least during an early stage of its development). My theory proposes that every type of ion channel should be under the control of Hebbian or anti-Hebbian learning rules, so that a neuron maximizes the total evidence in its membrane voltage, and thereby learns to make the best predictions that it can, given the set of ion channel subtypes and synapses that it has available to select from (Fiorillo, 2008).
There are two broad classes of ion channels. Ligand-gated ion channels are those that are gated (opened) by the binding of a chemical, usually a synaptically released neurotransmitter. Voltage-gated ion channels are those that open in response to a change in the voltage across a neuron’s membrane. There has long been a theoretical rationale, and experimental evidence, for the view that ligand-gated ion channels, such as AMPA-type glutamate receptors, are under Hebbian control. But there has been almost no consideration of the possibility that voltage-gated ion channels are also under the control of such learning rules.
Each synapse, and its associated ligand-gated ion channels, collects evidence from a distinct region of space. The learning rules thus select the region of space that provides the neuron with the most evidence. My theory ascribes the same basic function to the learning rules that regulate voltage-gated ion channels, except that the diversity of these channels is in the temporal rather than spatial domain.
Distinct types of voltage-gated ion channel have distinct kinetic properties, and therefore they carry information from a distinct period of the past. In other words, each type of voltage-gated ion channel has a distinct memory of the past membrane voltage of the neuron, and indirectly, the past state of the external world that caused the past membrane voltage. Some periods of the past are better predictors of the present and future than others, and which period is the best predictor depends on the particular patterns of synaptic excitation that a neuron receives. Hebbian and anti-Hebbian learning rules are proposed to select those types of voltage-gated ion channels that make the best predictions of synaptic excitation, and indirectly, the external environment.
The K+ (potassium) channels are the most diverse type of voltage-gated ion channel, with about 100 different genes and a much larger number of distinct ion channel subtypes. In most cases their function is to counterbalance synaptic excitation, and for that they should be under anti-Hebbian control, like the GABAA and glycine receptors at the inhibitory synapses described above and in Kim and Fiorillo (2017).
An anti-Hebbian rule is a form of negative feedback and acts to minimize prediction errors. If a subtype of K+ channel is open just before a spike, then it correctly predicted the synaptic excitation that caused the spike, but it was not powerful enough to prevent the spike. Therefore it should have provided stronger inhibition. An anti-Hebbian rule will make it stronger by inserting more K+ channels of that type into the membrane (and perhaps by other mechanisms as well, such as phosphorylation). If instead this channel type is open but there is no spike, it contributed to causing a negative prediction error. That is evidence that its influence was too strong, and it will be weakened by an anti-Hebbian rule (by removing channels from the membrane, for example). If instead these channels were closed, then they have no responsibility for either the presence or absence of a spike, and their strength should not change (or it should very slightly weaken, since ion channels that almost never open are not useful).
For there to be Hebbian or anti-Hebbian control over voltage-gated ion channels would require initiation and expression mechanisms analogous to those for synaptic plasticity. There would need to be a “coincidence detector” to detect the coincidence between channel configuration and a subsequent spike. This would need to be a voltage-sensor in physical contact with the channel. There would also need to be an expression mechanism, such as phosphorylation or insertion of channels into the membrane, that is activated following a coincidence.
One way to test this aspect of the theory is by looking for learning directly, and the other is to look at how K+ channel expression relates to temporal patterns of sensory-related excitation. By “look directly” I mean to do the same types of experiments that have long been done on excitatory synapses to look for long-term potentiation and depression, or ideally, its spike-timing dependent analogue. However, this is more challenging with voltage-gated ion channels, because the analogue of “presynaptic activity” is the configuration of the ion channels themselves. If the channels are open, then that is analogous to presynaptic activity (spike-induced neurotransmitter release). Ideally we want to have independent control over channel configuration and membrane voltage, but this is difficult since the channels themselves are voltage-gated. It is not impossible however, especially given that the analogue of “postsynaptic activation” should be a spike, and the causal relationship between channel configuration and spike or no spike necessarily involves a delay.
An experiment could be done with a cell-attached patch recording of one or more voltage-gated ion channels, and a second electrode (intracellular or extracellular) controlling spike generation. Activation of K+ channels in the on-cell patch followed within a couple of milliseconds by a spike should increase the number or strength of channels of that same type in the patch, and K+ channel activation followed by no spike should weaken that type of channel in the patch. However, since the recorded patch of membrane is small and contains few ion channels, a change in current amplitude due to learning would presumably be a rare and stochastic event. Therefore many pairings would be necessary to test the hypothesis of anti-Hebbian learning.
The same type of experiment could be done on voltage-gated calcium channels. They are interesting because the theory (Fiorillo, 2008), as extended in Fiorillo et al (2014) and Kim et al (2015), predicts that they can be under Hebbian or anti-Hebbian control, depending on whether they are in motor or sensory neurons, respectively. For example, T-type calcium channels in visual thalamus (LGN) should be under anti-Hebbian control (Hong et al., 2014), whereas T-type calcium channels in motor thalamus (Kim et al., 2015), and L-type calcium channels in striatal medium spiny neurons and spinal motoneurons, should be under Hebbian control.
In motor neurons, the calcium channel subtype in which activation is most synchronous with synaptic excitation should be selectively strengthened. A Hebbian rule implements positive feedback and it would reinforce coincidence of synaptic excitation and calcium channel activation just as it reinforces coincident activation of two or more excitatory synapses. The result of Hebbian control of calcium channels is to strengthen those channel subtypes that most effectively predict and amplify synaptic excitation to cause spikes.
A very different means of testing this aspect of the theory is to examine the relation of channel expression to temporal patterns of sensory excitation. The auditory system is attractive for such studies because sound exhibits a great diversity of temporal patterns. A simple prediction of the theory is that there is greater functional diversity (and presumably genetic diversity) of K+ channels in the auditory system than in other sensory systems. This could be tested using K+ channel protein and mRNA expression data from the Allen Brain Atlas.
A stronger and more quantitative test of the theory can be done by comparing its predictions for the specific relation of K+ channel subtypes to patterns of synaptic excitation. For example, in hair cells and in early auditory regions, there are tonotopic maps featuring neurons arranged topologically according to the frequency tuning of the neurons. As predicted by theory, K+ channel expression is known to vary systematically across the tonotopic spectrum, at least in some regions. For example, BK potassium channel expression varies across the tonotopic spectrum in turtle hair cells, with the BK channels in high-frequency hair cells having faster kinetics that are matched to the frequency of sound to which that cell is tuned (Fettiplace and Fuchs, 1999). This confirms the prediction of my theory.
However, it would be nice to see a more direct and quantitative comparison of theory and data, analogous to that described above for synaptic inhibition (Kim and Fiorillo, 2017). It would also be nice to see the same approach taken in other early auditory regions. In many such regions, membrane voltage is not phase-locked to the neuron’s preferred sound frequency, but neurons tuned to higher frequencies would nonetheless be expected to receive synaptic excitation with faster temporal dynamics. This is because smaller objects tend to move faster than large objects, creating higher frequency sound with a higher rate of amplitude modulation. Neurons tuned to higher frequencies are therefore likely to express faster K+ channels to counterbalance this faster excitation. But that is just my guess of what the theory would predict. The theory can make much more precise and quantitative predictions given data on the natural statistics of synaptic excitation in a particular neuron, which according to the theory, the neuron should learn and use for prediction.
Finally I will briefly mention that all of this depends on the voltage-dependence of the channels, which I neglected above for simplicity. The temporal patterns in the synaptic excitation, and in the external sensory stimulus, are not directly relevant to these ion channels. These channels directly sense only the membrane voltage, and only over a particular range, depending on the channel’s voltage-dependence. Therefore it is only the temporal pattern of the membrane voltage in that range that matters to an ion channel and the learning rules that regulate it. Some K+ channels sense and predict EPSPs, whereas others sense and predict action potentials. A single neuron needs multiple types of K+ channels to deal with these different voltage ranges, and particularly to deal with the multiple temporal patterns present even over a particular voltage range.