To study how properties of predictive learning rule may relate to consciousness processes, we created a recurrent neural network. It had 420 input units, 50 hidden units, and 10 output units as illustrated in Figure 2A. The network was trained on a hand-written digit recognition task MNIST (LeCun et al., 1998), with 21 × 20 pixels from center of each image given as input to the network. The details of network training are described in a study (Luczak et al., 2022). First, network is presented with only an input signal and the activity starts propagating throughout the network until it converges to a steady-state, when the neurons’ activity stops changing, as depicted in Figure 2B. This is repeated for 1,600 randomly chosen stimuli. During this phase, we also trained a linear model to predict the steady-state activity. Specifically, for each individual neuron, the activity during the five initial time steps (x₍₁₎,…,x₍₅₎) was used to predict its steady-state activity at time step 20: x₍₂₀₎, such that: x(20)⁢≈⁢x∼=λ(1)*x(1),+⋯+λ(5)*x(5)+b, where x∼ denotes predicted activity, λ and b correspond to coefficients and offset terms of the least-squared model, and the terms in brackets correspond to time steps (Figure 2B). Next, a new set of 400 stimuli was used, where from step 8, the network output was clamped at values corresponding to image class (teaching signal). For example, if the image of number 5 was presented, then the value of the 5th output neuron was set to “1,” and the values of the other 9 output neurons was set to “0,” and network was allowed to settle to the steady-state. This steady-state was then compared with predicted steady-state activity, which was calculated using the above least-squared model. Subsequently, for each neuron, the weights were updated based on the difference between the actual (x_j) and its predicted activity (x∼j) in proportion to each input contribution (x_i), as prescribed by the predictive learning rule in Figure 1 (Matlab code for a sample network with our predictive learning rule is provided in Supplementary Material).

The network using predictive learning rule showed a typical learning curve, with rapid improvement in performance in the first few training epochs, and with plateauing performance during later training epochs (Figure 3A). Notably, this shape of learning curve is also typical for skill-learning in humans, where, initially at the novice level, there are fast improvements, and it takes exponentially more time to improve skills at, for example, elite athlete level (Newell and Rosenbloom, 1981). However, what is new and interesting here, is how a surprise (i.e., the difference between actual and predicted activity) evolved during network training (Figure 3B) and how it compares to the stages of “skill consciousness,” as explained below.

It was observed that learning involves the four stages of “conscious competence” (Broadwell, 1969; Das and Biswas, 2018; Figure 3C): (1) Unconscious incompetence – where individual does not know what he/she doesn’t know, and, thus, that individual is not aware of his/her own knowledge deficiencies (e.g., foreigner may not know about certain local traffic regulations); (2) Conscious incompetence – where the individual recognizes his/her own lack of knowledge or skills but does not have those skills (e.g., a car passenger who does not know how to drive); (3) Conscious competence – where the individual develops skills but using it requires conscious effort (e.g., beginner car driver); (4) Unconscious competence - where due to extensive practice, the individual can perform learned tasks on “autopilot” (e.g., driving car on the same route every day).

Here we illustrate how the above stages of conscious competence could be recapitulated by the network with our predictive learning rule. We used the neuronal surprise as a proxy measure of consciousness, which is motivated by previous theoretical (Friston, 2018; Waade et al., 2020) and experimental work (Babiloni et al., 2006; Del Cul et al., 2007), which will be discussed in later sections. We calculated the surprise for each neuron j as: < |xj-x∼j|>, where |…| denotes absolute value, and < … > denotes average across all 400 images presented in a single training epoch. The neuronal surprise was defined as mean surprise across all of neurons. To better illustrate the network behavior, we also plotted accuracy (a.k.a. competence) vs surprise (a proxy of consciousness) (Figure 3D; model details and code to reproduce presented figures are included in Supplementary Material). Initially, when the network was presented with an input image, the neurons in the hidden layer could almost perfectly predict what will be the steady-state activity after the output units are clamped (Figure 3B, first few epochs). This is because the network starts with random connections and the signal coming from 10 output units is relatively week in comparison to the signal coming to the hidden layer from a much larger number of input units: 420. Thus, the steady-state activity, which neurons learn to predict when only the input image is presented, is not much different from the steady-state activity when input image is presented together with clamped outputs. This is like the “unconscious incompetence” stage, as the network is almost completely “not aware” of the teaching (clamped) signal (Figure 3D, blue line). However, as the activity of the output neurons is mostly correlated with any discrepancy between the actual and the predicted activity in hidden layer neurons, thus, the synaptic weights from output neurons are most strongly modified. Consequently, as the learning progresses, the hidden neurons are more and more affected by the output units, and their surprise: the discrepancy between actual and predicted activity, increases. This is analogous to the “conscious incompetence,” where the network becomes “aware” of the clamped teaching signal, but the network has not yet learned how to classify images correctly (Figure 3D, light blue). In result, as magnitude of surprise |xj-x∼j| increases, then other synaptic weights also start changing more, as prescribed by the predictive learning rule in Figure 1. Those synaptic updates made the activity driven by the input image, closer to the desired activity as represented by the clamped output units. This could be characterized as “conscious competence,” where the surprise signal allows the network to learn and to become more competent on that task (Figure 3D, yellow line). Finally, as network learns to predict the image class with high accuracy, then the surprise (the difference between predicted and clamped teaching signal) is diminishing, which is analogous to an expert who achieved “unconscious competence” (Figure 3D, green line). This, that the neuronal surprise recapitulates the stages of conscious competence, by first increasing and then decreasing during learning, was a general phenomenon across different datasets and across diverse network architectures (Supplementary Figure 1).

Derivation of the predictive learning rule in Figure 1 shows that the best strategy for a cell to maximize metabolic energy is by adjusting its synaptic weights to minimize surprise: |x-x∼|. However, this change in surprise does not need to take minutes or hours, as typically required for structural synaptic modification to occur (Xu et al., 2009). Neurons have adaptation mechanisms, which could serve to reduce surprise at a much faster time scale of tens of ms (Whitmire and Stanley, 2016).

Neural adaptation is a ubiquitous phenomenon that can be observed in neurons in the periphery, as well as in the central nervous system; in vertebrates, as well as in invertebrates (Whitmire and Stanley, 2016; Benda, 2021). Neuronal adaptation can be defined as the change in activity in response to the same stimulus. The stimulus can be a current injection into a single neuron or a sensory input like sound, light, or whisker stimulation. Usually, neuron activity adapts in exponential-like fashion, with rapid adaptation at the beginning, and then later plateauing at a steady-state value (Figure 4A). Typically, neuronal adaptation is shown as the decrease in activity in response to excitatory stimuli. However, neurons can also adapt by increasing its spiking ability when inhibitory stimulus is presented; for example, an injection of constant hyperpolarizing current (Aizenman and Linden, 1999). Thus, adaptation could be seen as change in neurons activity toward a typical or expected (predicted) level (x∼).

To investigate effect of adaptation on neuronal processing, we implemented a brain-inspired adaptation mechanism in the units in our network. For this, during the clamped phase from time step 8, the activity of each neuron was nudged toward the predicted state (Figure 4B). Specifically, the activity of neuron j at time step t was calculated as: xj,t=a*x∼+(1-a)*∑i(wi,j*xi,t-1), where 0 ≤ a ≤ 1 is a parameter denoting strength of adaptation. For example, for a=0, the adaptation is equal to zero, and the network activity is the same as in original network described in Figure 2. To update synaptic weights, we used the same learning rule as in Figure 1: Δ⁢wi,j=xi⁢(xj-x∼j), but here x_j represents clamped activity with added adaptation, which can also be denoted as x_A. Interestingly, networks with implemented adaptation achieved better accuracy than the same networks without adaptation (Supplementary Information). This could be due to the fact that if an activity in the clamped phase is much different from an expected activity without the clamp, then learning may deteriorate as those two network states could be in different modes of the energy function (Scellier and Bengio, 2017). Adaptation may reduce this problem by bringing clamped state closer to expected. To give an analogy, if part of a car is occluded by a tree, then, purely by sensory information, we cannot say what is behind that tree. However, based on our internal model of the world, we know what shape a car is, and, thus, we can assume that the rest of the car is likely behind the tree. Similarly, neuronal adaptation may allow a neuron to integrate the input information with predictions from its internal model, and then adjust its activity based on this combined information leading to a more appropriate response.

It is largely accepted that consciousness is a gradual phenomenon (Francken et al., 2021). It was also suggested that even a single cell may have a minimum level of consciousness, based on the complexity of behavior and complexity of information-processing within each cell (Reber, 2016; Baluška and Reber, 2019). For example, every single cell contains large biochemical networks, which were shown to make decisions and to perform computations comparable to electrical logic circuits (Supplementary Figure 2; McAdams and Shapiro, 1995). This allows for highly adaptive behavior, including sensing and navigating toward food, avoiding a variety of hazards, and coping with varying environments (Kaiser, 2013; Boisseau et al., 2016). For instance, single-celled organisms were shown to be able to “solve” mazes (Tero et al., 2010), to “memorize” the geometry of its swimming area (Kunita et al., 2016), and to learn to ignore irritating stimulus if the cell’s response to it was ineffective (Tang and Marshall, 2018). Moreover, single-celled microorganisms were shown to predict environmental changes, and to appropriately adapt their behavior in advance (Tagkopoulos et al., 2008; Mitchell et al., 2009). Those complex adaptive behaviors were proposed to resemble cognitive behavior in more complex animals (Lyon, 2015). This likely requires organism to build some sort of internal predictive model of their own place in the environment, which could be considered as a basic requirement for consciousness.

The results presented in Figure 3 suggest that the level of consciousness could be related to the amount of surprise. This is also supported by results from human EEG studies, where the neuronal signature of surprise: P300, closely reflects conscious perception (Del Cul et al., 2007; Dehaene and Changeux, 2011). Here, we propose that in a neuron, adaptation could be seen similarly to P300, as a measure of surprise, and thus, it could provide an estimate of the level of “conscious cellular perception.” Specifically, as described above, surprise could be defined as a difference between actual (x) and predicted activity (x∼). Because adaptation changes neurons’ activity toward a predicted activity level, thus, the size of adaptation (|x−x_A|) is directly related to the size of surprise: (|x-x∼|). Therefore, we propose to define the single-neuron correlate of consciousness (sNCC) as the magnitude of neuronal adaptation sNCC = |x−x_A|, (Figure 4B). Based on this, we hypothesize that single-cell predictive adaptation is a minimal and sufficient mechanism for conscious experience.

First, we will explain the main ideas using a simplified example, then later, we will present how it can be generalized. Let us have a two-dimensional environment, where at each location P, there is a certain amount of food. There is also an organism that wants to go to a location with the highest amount of food. That organism does not know exactly how much food there is at any given location, but based on past experience, the organism has an internal model of the environment to help with predictions. For instance, let us assume that the maximum concentration of food (m) is at point P_m, but the smell of food comes from the direction of point: P_s, where s stands for sensory evidence (Figure 5). However, the concentration of food in the past was highest in the North direction. The internal model combines this information and predicts the highest probability of food in the North East direction at point P_p, where p stands for predicted. Based on this, the organism adapts and moves toward P_p to location P_A, where _A stands for adaptation. When the organism arrives to P_A location, then it can compare the actual amount of food at that location with the predicted one, and update the internal model accordingly. Thus, by combining sensory information and internal model predictions, our organism was able to adapt its behavior more appropriately.

In the above-described case, we could say that our organism was quite conscious of its environment, as it made close to optimal decision. We can quantify it by measuring how close to optimal location an organism moved: d(P_A, P_m), as compared if it would move in reflex-like fashion to location that is purely determined by sensory stimulus: d(P_s, P_m), where d(.,.) denotes a distance between 2 points. Specifically, we can define organism consciousness of environment as C_e = d (P_s, P_m) – d (P_A, P_m), (Figure 5, insert). It is worth noting that if an organism has a good model of external environment to correctly predict location with maximum food, then: P_p ≈ P_m, and thus, the first term in C_e: d(P_s, P_m) ≈ d(P_s, P_p), where this distance d (P_s, P_p) between sensory evidence (P_s) and model prediction (P_p) is a description of surprise. The second term in C_e: – d (P_A, P_m), describes how far organism is from location with maximum food/energy (P_m) after it made adaptation (P_A). This could be seen as an error term, which could arise if predictive model is incorrect or if organism is unable to move exactly to the predicted location. Hence, according to the above definition of C_e, consciousness is equivalent to surprise, if error term is 0, which would be the case for an organism to able to perfectly adapt.

Although, we used here a two-dimensional environment as an example, this can be generalized to a high-dimensional sensory space. Let us consider a simple organism which can sense concentration of 10 substances in a deep ocean. As organism swims, it changes its position in 3D space, but more importantly, concentration of 10 substances indicating location of food and predators also changed with each movement. Thus, 3D space translates to 10D sensory space, which is more relevant to that organism behavior. Therefore, distances d in C_e = d (P_s, P_m) − d (P_A, P_m), may be more appropriately calculated in sensory space of that organism, instead of the standard 3D spatial coordinates. For example, we implicitly used idea of sensory space in case of neurons shown in Figures 3, 4. Neuron senses its local environment through variety of channels located especially in synapses. Activity of a neuron affects other neurons, which through feedback loops change synaptic inputs to that neuron, and thus, its sensory environment. Because neuron gets energy from blood vessels, which dilation is controlled by coordinated activity of local neurons, therefore, neuron may “want” to move in the sensory space corresponding to activity patterns resulting in the most local blood flow. Therefore, change in neuron activity is equivalent to a movement in a chemical sensory space, where different locations in that space correspond to different amount of energy obtained by a neuron. For that, the word “environment” in C_e refers to this highly dimensional sensory space rather than that of the typical 3D space.

This generalization to sensory space also allows to see notions introduced earlier in Figures 3, 4 as special cases of environmental consciousness C_e. For example, when organism has the perfect model of external environment, then it can correctly predict the location with maximum food, thus, P_p = P_m, as we have explained before. However, if that organism can also move exactly to predicted location such that: P_A = P_p, then, also P_A = P_m. In such case, an adaptation error d(P_A, P_m) becomes 0, and thus, C_e = d (P_s, P_m). Considering the above case that P_m = P_A, C_e can also be expressed as C_e = d (P_s, P_A), which is a distance by how much an organism moved or adapted. Thus, in case of the neuron described in Figure 4, C_e=d (P_s,P_A) ≈ d |x, x_A| = |x − x_A| = sNCC. Similarly, as mentioned earlier, C_e becomes equivalent to surprise if organism perfectly adapts (P_A = P_p = P_m). In such case, adaptation error is zero, and we can write Ce=d⁢(Ps,Pp)≈|x-x∼|, which is the distance between the stimulus-evoked activity and the model prediction, which we used to quantify the skill consciousness in Figure 3. Thus, C_e is a function of surprise and ability of organism to adapt to minimize that surprise.

Note that surprise and adaptation could be considered as contributing to C_e on different timescales, with synaptic changes gradually minimizing surprise over a long period of time, and with neuronal adaptation changing neuronal firing rapidly within 10–100 ms. When an organism is learning a new skill, then activity driven by bottom-up signals is different from activity provided by top-down teaching signals, which results in a higher surprise term. However, if neurons cannot adapt their activity accordingly (e.g., when biochemical processes mediating adaptation within a neuron are blocked), then adaptation error will be as large as the surprise term, resulting in C_e = 0 and, thus, in no conscious experience. Therefore, the surprise term could be interpreted as “potential consciousness,” meaning the maximum possible consciousness to a given stimulus. Synaptic strength gradually changes over a period of learning, resulting in slow changes in “potential consciousness.” However, when a stimulus is presented, and neurons rapidly adapt their activity toward the predicted level, it reduces the adaptation error term and results in C_e > 0, and, thus, in conscious perception within a fraction of a second.

A hypothesis, by definition, should generate testable predictions. Our main hypothesis is that the neuronal adaptation is a neuronal correlate of consciousness. This implies that neurons and, thus, brains, without adaptation cannot be conscious. Therefore, our hypothesis predicts that any mechanism which affects neuronal adaptation will also affect consciousness. This prediction was shown to be correct for a diverse group of neurochemicals involved in sleep and anesthesia, which also alter the neuronal adaptation. For instance, levels of multiple neuromodulators in the brain such as serotonin, noradrenaline, and acetylcholine are significantly different between waking and sleeping in REM or non-REM stages (España and Scammell, 2011). Whole-cell voltage-clamp recordings in vitro in the pyramidal neurons have demonstrated that all those neuromodulators also affect neuronal adaptation (Satake et al., 2008). Similar results were obtained when testing various substances used for anesthesia, such as urethane (Sceniak and MacIver, 2006), pentobarbital (Wehr and Zador, 2005), and ketamine (Rennaker et al., 2007). Moreover, it was shown that a large variety of anesthetics, including butanol, ethanol, ketamine, lidocaine, and methohexital are blocking calcium-activated potassium channels, which mediate neuronal adaptation (Dreixler et al., 2000). Interestingly, considering a broad spectrum of molecular and cellular mechanisms affected by different anesthetic compounds, there remains significant uncertainty of what is the single mechanism underlying anesthesia (Armstrong et al., 2018). Our theory suggests that what all anesthetics could have in common is the ability to disturb neuronal adaptation. Thus, our theory clearly provides testable predictions, which could either be invalidated or validated by using pharmacological and electrophysiological methods (see also “Limitation” section for more discussion on this topic).

Important consequence of a neuron adapting its activity toward a predicted level is that it allows neurons to exchange information about its predictions. Thus, neuron output activity is not exclusively driven by its synaptic inputs, but it is also a function of its internal predictive model. Below, we will briefly describe a few of the most prominent studies, as well as the theories of consciousness [for in-depth reviews see Francken et al. (2021) and Seth (2021)]. We will particularly focus on outlining the differences and similarities to our theory of predictive adaptation, and how it may provide a theoretical basis for connecting diverse theories of consciousness.