Multilevel development of cognitive abilities in an artificial neural network

Significance Multiple biological mechanisms support the unique ability of the brain to develop complex cognitive abilities. Nevertheless, it remains unclear which mechanisms are necessary and sufficient. We propose a neurocomputational model of the developing brain spanning sensorimotor, cognitive, and conscious levels. The model solves three tasks of increasing complexity: from visual recognition to cognitive manipulation and maintenance of conscious percepts. Results highlight two fundamental mechanisms for the multilevel development of cognitive abilities in biological neural networks: 1) synaptic epigenesis, with Hebbian learning at the local scale and reinforcement learning at the global scale; and 2) self-organized dynamics, through spontaneous activity and balanced excitatory/inhibitory ratio of neurons. We emphasize how these core features of human intelligence could guide future development in artificial intelligence.


Introduction
Understanding the human brain remains the major challenge of biological sciences and has become the focus of considerable attention from both neurobiology and computational sciences (1)(2)(3). As a consequence of the recent success of neuro-inspired algorithms, including artificial neural networks and reinforcement learning, the fast-developing field of machine learning continues to look to neuroscience for inspiration. For instance, recent reflections on the connectomic implications of brain Hominization in the course of evolution and brain development have underlined the importance of a still under-evaluated notion of multilevel processing in the brain from sensory processing -and its elementary "local" neuronal circuitsto higher brain functions (4). Several recent theories emphasize the hierarchical relationship between local and global processes in the making of higher-level cognition. Among these, the Global Neuronal Workspace (GNW) (5) is exemplary. A hierarchical relationship between local and global processes is also outlined in Kahneman's functional distinction of "System 1" handling fast and nonconscious cognitive processes, and "System 2" handling cognitive tasks requiring slower and more concerted conscious effort (6). While current algorithms in machine learning are now tackling System 1 tasks at the human performance level, a major milestone for modern artificial intelligence would be to provide models capable of approximating System 2 cognition (7).
To a large degree, the strides made in artificial intelligence research during recent decades can be explained by the overwhelming success of error backpropagation. Nevertheless, its biological plausibility remains a matter of considerable debate (8)(9)(10). Unlike artificial deep networks, and due to their discontinuous nature, spiking neurons are thought to learn by other means. Concretely, electrophysiological methods have indicated two principal mechanisms of learning in the human brain: Hebbian learning (11) and reinforcement learning (12). The mechanism of epigenesis by synapse selection (or synaptic pruning) is also known to play a significant role in both learning and biological development. Moreover, it has been demonstrated that neurogenesis can occur in adults (13) and that astrocytes may be implicated in synaptic modulation during learning (14). Despite solid evidence from joint anatomical, physiological and molecular investigations in the course of nervous system development, such mechanisms have been underexploited in brain modeling and computer sciences (15,16). Recent work in neuroscience continues to uncover novel mechanisms at play during learning, which are far from being integrated into computer science. These and other mechanisms may provide important links, both within and between levels of organization, in diverse contexts -ranging from genes networks to long-range neuronal connectivity of the brain (17).
If detailed key mechanisms of learning in the brain have already been investigated, there is as yet no theoretical consensus on how these varied learning mechanisms interact in the brain (18)(19)(20)(21). Yet it is known that the brain operates constantly through variation-selection mechanisms at multiple timescales (17) and thus supports the development of a multiscale and dynamical perspective on cognition and learning (22,23). In artificial intelligence, new approaches have adopted such dynamical and multiscale aspects using attention (24) or social interaction between multiple agents (25). Basic developments inspired by neuroscience have also highlighted the importance of better capturing hierarchical relationships (7,26,27). Finally, the idea of biological learning without any inductive bias, typical of ANNs, has been criticized.
On the contrary, biological investigations point dramatically to the embodied quality of cognition, perception, and action, as well the important role of embeddedness in natural and social milieus (28,29).
In this paper, we present a framework for biologically plausible learning, using simulations of synaptic epigenesis that combine multiscale architecture with STDP and dopamine signaling.
We hypothesize that synaptic epigenesis unfolds differently at local and global scales and that the requisite conditions for solving complex tasks associated with the global neuronal workspace include not only local but global epigenesis. We first identify necessary and sufficient conditions for proper learning of perceptual, cognitive, and conscious tasks. Next, we analyze how the identified factors influence performance in these tasks. After introducing the neurocomputational model, results are presented and the key role of dopamine and inhibitory neurons in the learning of higher cognitive tasks is discussed. Finally, we delineate future perspectives for the field of computational cognitive neuroscience and neuroAI.

Figure 1.
T h  e  f  u  l  l  n  e  t  w  o  r  k  a  r  c  h  i  t  e  c  t  u  r  e  c  o  m  p  r  i  s  e  s  S  e  n  s  o  r  i  m  o  t  o  r  ,  C  o  g  n  i  t  i  v  e  ,  C  o  n  s  c  i  o  u  s  ,  a  n  d  M  e  t  a  c  o  g  n  i  t  i  v  e  l  e  v  e  l  s  .   T  h  i  s  i  s  a   s  c  h  e  m  a  t  i  c  r  e  p  r  e  s  e  n  t  a  t  i  o  n  o  f  t  h  e  f  u  l  l  n  e  t  w  o  r  k  a  n  d  i  t  s  c  o  m  p  o  n  e  n  t  s  .  C  o  l  o  r  s  o  f  a  r  r  o  w  s  d  e  p  i  c  t  d  i  f  f  e  r  e  n  t  t  y  p  e  s  o  f  c  o  n  n  e  c  t  i  o  n  s  :  l  i  g  h  t   b  l  u  e  -e  x  c  i  t  a  t  o  r  y  l  o  n  g  -r  a  n  g  e  d  (  f  i  x  e  d  :  b  e  t  w  e  e  n  V  C  a  n  d  P  F  C  a  n  d  S  T  D  P  :  b  e  t  w  e  e  n  P  F  C  a  n  d  M  C  ) , e  x  c  i  t  a  t  o  r  y  (  S  T  D  P  )  ,  y  e  l  l  o  w  -e  x  t  r  a  c  e  l  l  u  l  a  r  d  o  p  a  m  i  n  e  m  o  d  u  l  a  t  i  o  n  (  a  f  f  e  c  t  s  d  o  p  a  m  i  n  e  m  o  d  u  l  a  t  e  d  S  T  D  P  ) .

Overview of the Model and its neuronal components
We propose a computational model of an elaborate multilevel neural network able to pass cognitive tasks of increasing difficulty (Fig. 1). The task solving develops at three nested levels of hierarchical organization, where synaptic epigenesis may proceed at both local and global scales. The three different levels of structural organization show the increasing complexity of connectomic architecture with each nested level through a continuous progression. The first level, which we refer to as the "sensorimotor level", deals with local sensory processing and classification of visual information: it requires local synaptic epigenesis. At the second level, or "cognitive level" (30), the network successfully passes a delay conditioning task: it mobilizes multiple cortical areas and their integration requires long-range axonal connections. The third level referred to as the "conscious level" is able to carry a trace conditioning task using a similar architecture as the cognitive level yet, with the addition of the necessary contribution of inhibitory interneurons.  (11). Each level may include several differentiated territories, or sub-networks, such as, for instance, the visual cortex or the motor cortex in the lower sensorimotor level. Every part of the network is modeled using neuronal assemblies except for the first two layers of the local network of the visual cortex. The state of each neuron is determined by the leaky integrate-and-fire model (31) and each assembly includes not only internal connections but external connections to other assemblies of the whole network. An important feature of the model is its evolution with time. In the course of its elaboration, the network connections are subject to epigenesis by selective stabilization of synapses (10). At critical stages of development, stable connections emerge from an initial growth process with overproduction of synaptic contacts with maximal diversity and variability, then a selectionstabilization regulated by the state of activity of the network takes place, together with the elimination -or pruning -of the un-selected ones (10). We examine, at all three levels of organization, how this development happens, both at local and global scales, and how internal and environmental factors affect its evolution.
The intrinsic spontaneous activity of the component neurons is a singular functional component introduced in the model proposed to occur at all its hierarchical levels. Also, the network is not exclusively composed of excitatory neurons as in most standard networks. At the highest level, we introduce inhibitory interneurons and test how the interneurons vs excitatory neurons ratio affects the performance. Long-range connections (5) which interconnect widely distinct territories of the global network are introduced at the higher cognitive and conscious levels.
Last throughout this work we use concomitantly different types of learning. At the local level, the network evolves under Hebbian learning with STDP (32). However, all the global connections are controlled by Reinforcement learning (12). To do so, we introduce the dopamine reward system (33), which is modulated by interactions with the external environment ( Fig. 1).

Description of the tasks
The three selected tasks challenge differentially the sensorimotor, cognitive, and conscious levels.
The sensorimotor level is common to all tasks. At this level, we use the MNIST (Mixed National Institute of Standards and Technology) dataset of handwritten digits as visual input to the Visual Cortex (34). The network performs a number recognition task (35). Following image presentation, the network processes spikes arriving within the first 250ms, followed by a relaxation period of the same length. Thus, one image is shown every 500ms during the learning period. At the higher levels, when the full network receives an input from the Visual Cortex, the network is expected to respond to the question posed by the experimenter: "Is this number larger than x? If yes, push a button.". For the case with only 0 and 1 in the input data, we took x equal to 0.
Two other tasks delay and trace conditioning (36), have been used to challenge the cognitive and conscious levels. In delay conditioning, the unconditioned stimulus (US) immediately

Local network of the Visual Cortex
The architecture of this network (Fig. 2a), inspired by the work of Masquelier and Thorpe (37), consists of three layers representing the simplified model of the visual pathway. The first layer imitates the lateral geniculate nucleus and performs a convolution operation with filters of various angles. Then, in the second layer, only the strongest orientation reaches the V1 area.
And in the last layer (V4 area) the patterns of digits are learned by mean of STDP. In the text we use the notation of 'complex' and 'simple' cells, thus the overall architecture can be written as S1-C1-S2 (see 37). A detailed description of each layer can be found in the Materials and Methods section.
To train the model and observe the local epigenesis of STDP regulated synapses we used the MNIST dataset of handwritten digits (Fig. 2b). These inputs consist of images of handwritten digits from 0 to 9. MNIST is one of the fundamental datasets in Machine Learning, which makes it reliable and easy to use, as well as an informative benchmark for non-traditional learning methods. The images are introduced to the network continuously with equal relaxation periods in between. These relaxation phases are needed for the system to eliminate the temporal effect on STDP coming from the previous stimulus, and to bring all membrane potentials back to rest.
After a number of images are presented, the network's plasticity allows it to evolve from a homogeneous state to one that has encoded various patterns by way of Hebbian learning.

Full Network
After the local network has undergone the process of Hebbian learning, the weights of C1-S2 connections are frozen and this network is considered mature. Then, a new architecture is

Tasks solving by the Model
The proposed model of an elaborate multilevel neural network (fig1) is able to pass behavioral tasks of increasing difficulty. The three levels of structural organization exhibit increasing connectomic complexity with each nested level.

Stabilization and pruning
At this level, the neuronal network -described above as the "visual cortex local network" -can perform a recognition task of numbers, using handwritten digits as inputs. This image processing task mobilizes nonconscious perception, by leveraging local synaptic epigenesis as well as the spontaneous activity of the network. Fig. 3 a-d shows the computed multistep evolution of this performance.
The initial state of the LGN-V1 synaptic population is highly homogeneous. The synaptic weights are normally distributed around the mean of 70% with the standard deviation of 2%.
When the network is presented with the constant input stream of images the synaptic population starts to evolve. The evolution of the synaptic strength of the whole synaptic population is presented in Fig. 3b. This graph shows the evolution of the percentiles for the population with a step size of 10, from 0 to 100. It can be seen that all synapses do not behave in the same way.
Around two-thirds of population synapses immediately weaken and are as a consequence eliminated. At the same time, other synapses are strengthened and selected (see 10). The selected synapses are those responsible for the transfer of the signals related to the input information. Others are gradually eliminated, such that they exert a lesser impact on the final layer. However, this process is not linear and learning patterns change over time. Fig. 3d represents the evolution of the states within the synaptic population. In this schematic representation, a synapse belongs to the green ("selected") area if the synaptic strength (or synaptic weight) is greater than 75%. Similarly, the red ("eliminated") area accounts for synapses with a strength of less than 25%. All others between 25% and 50% fall into the gray ("undetermined") area. This zone contains the entire population at the beginning of the simulation.
Due to non-linear behavior, we have outlined two distinct stages of learning depending on the The second stage of learning can be characterized as the fixation and reinforcement of the changes that happened before. The fraction of eliminated synapses monotonously reaches a plateau level with a value of around two-thirds of all synapses. The group of selected synapses, however, experiences "overshooting" and, after 200 images, starts to slowly decrease before reaching a plateau. This behavior can be described as a period when the S2 features become clearer, eliminating redundancy between them and within the spontaneous intrinsic activity.
In addition to synaptic epigenesis, the intensity of spontaneous intrinsic activity is among the factors affecting the performance of the network. In our model, each of the C1 neurons is able not only to produce spikes caused by the S1 layer but also to generate spontaneous intrinsic activity in a form of random Poisson distribution with a fixed rate. We can see that high performance (with the accuracy of classification > 90%) can only be achieved in a specific range of values (Fig. 3d). We observe no or very weak classification with values of spontaneous intrinsic activity close to zero; the same behavior is spotted for large values as well (> 1.0 Hz). 2. Cognitive level -nonconscious delay conditioning task.

Long-ranged connections and reward
In this section, we examine which network global architecture is needed for the performance of a delay conditioning task (Fig. 2) again using digits as visual stimuli for the unconditioned and conditioned stimuli. We discover that both the presence of long-range connections and their epigenesis are required. The initial state of the network is homogeneous: the local connections are set to 1% of the maximum weight and the long-range connections between the Prefrontal Cortex and the Motor Cortex are 50% for all synapses (Fig. 4).
The network uses long-range connections to transmit stimuli between Visual and Motor Cortices via a hub of excitatory neurons, as hypothesized in the Global Neuronal Workspace (GNW) model (5). We show that the network can perform a delay conditioned task (Fig. 4a represents the temporal relation between the stimulus and the trigger) without conscious processing. Fig.   4b depicts the activity of the Prefrontal Cortex at the beginning and the final stages of the learning process. The stimulus is erased from memory as soon as the input is gone, meaning that it is not processed in any way between Visual and Motor Cortices. However, depending on the initial parameters, our network can start firing constantly in a positive feedback loop since it contains only excitatory neurons. In this case, the network is not able to hold representations either. The epigenesis process transpires rather quickly, on the scale of ~15 shown images, and all the synapses responsible for the representations of the digit 1 get selected while the representations of the digit 0 get eliminated (Fig. 4c).  E  p  i  g  e  n  e  s  i  s  o  f  t  h  e  l  o  n  g  -r  a  n  g  e  c  o  n  n  e  c  t  i  o  n  s  n  e  t  w  o  r  k  b  e  t  w  e  e  n  P  r  e  f  r  o  n  t  a  l  a  n  d  M  o  t  o  r  C  o  r  t  i  c  e  s  w  h  i  l  e  p  e  r  f  o  r  m  i  n  g   b  i  n  a  r  y  c  l  a  s  s  i  f  i  c  a  t  i  o  n  u  n  d  e  r  t  h  e  d  e  l  a  y  c  o  n  d  i  t  i  o  n  i  n  g  .  a  .   D  e  s  c  r  i  p  t  i  o  n  o  f  t  h  e  d  e  l  a  y  c  o  n  d  i  t  i  o  n  i  n  g  t  a  s k .
E v  o  l  u  t  i  o  n  o  f  t  h  e  f  i  r  i  n  g  r  a  t  e   o  f  t  h  e  t  w  o  p  o  p  u  l  a  t  i  o  n  s  o  f  n  e  u  r  o  n  s  w  i  t  h  i  n  t  h  e  G  N  W  (  e  x  c  i  t  a  t  o  r  y  s  e  l  e  c  t  i  v  e  f  o  r  t  h  e  d  i  g  i  t  1  ,  e  x  c  i  t  a  t  o  r  y  s  e  l  e  c  t  i  v  e  f  o  r  t  h  e  d  i  g  i  t  0  )  .  1  5   n  e  u  r  o  n  s  p  e  r  p  o  p  u  l  a  t  i  o  n  a  r  e  s  h  o  w  n  ,  e  a  c  h  d  o  t  s  h  o  w  s  a  n  e  u  r  o  n  s  p  i  k  i  n  g  .  D  e  p  e  n  d  i  n  g  o  n  t  h  e  n  e  t  w  o  r  k  p  a  r  a  m  e  t  e  r  s  ,  i  f  t  h  e  i  n  t  e  r  n  a  The model turns out to be sensitive to the rate of spontaneous intrinsic activity. We have observed that, regardless of configuration, the model can learn how to perform the task only when the rate is set within a certain interval (Fig. 5). Moreover, while there is a specific range of values within which the network provides decent classification (accuracy > 95%), excessive activity can disturb the learning process, and the accuracy will remain around the chance level.
The performance does not change significantly as long as the spontaneous intrinsic activity is set in the right interval.   3. Conscious processing level -trace conditioning task.
The critical role of inhibitory neurons.
To date, the models associated with the Global Neuronal Workspace theory have focused on purely excitatory networks. Here we show the critical role of inhibitory neurons for conscious tasks. Additionally, we demonstrate how the performance on those tasks depends on the interplay between the ratio of excitatory and inhibitory neurons and spontaneous intrinsic activity.
Although the long-range connectivity itself is enough to perform the delay conditioning, it cannot achieve the trace conditioning task by itself (which is represented in Fig. 6a). On this level, the representation and erase it upon the presentation of a different stimulus (Fig. 6b). In this case, global epigenesis takes more time, on the order of ~150 shown images, because it requires the local epigenesis in the PFC to occur first (Fig. 6c). Therefore, the mechanism of learning trace conditioning is highly dependent on the ability of working memory to maintain a needed stimulus.   u  r  o  n  s  p  e  r  p  o  p  u  l  a  t  i  o  n  a  r  e  s  h  o  w  n  ,  e  a  c  h  d  o  t  s  h  o  w  s  a  n  e  u  r  o  n  s  p  i  k  i  n  g  .  D  e  p  e  n  d  i  n  g  o  n  t  h  e  n  e  t  w  o  r  k  p  a  r  a  m  e  t  e  r  s  ,  i  f  t  h  e  i  n  t  e  r  n  a  l   e  x  c  i  t  a  t  i  o  n  i  n  t  h  e  G  N  W  i  s  n  o  t  s  t  r  o  n  g  e  n  o  u  g  h  t  o  c  a  u  s  e  c  o  n  s  t  a  n  t  f  e  e  d  b  a  c  k  f  i  r  i  n  g  ,  t  h  e  n  e  t  w  o  r  k  g  r  a  d  u  a  l  l  y  l  e  a  r  n  s  t  o  s  u  s  t  a  i  n  t  h  e   r  e  p  r  e  s  e  n  t  a  t  i  o  n  o  f  a  s  t  i  m  u  l  u  s  i  n  t  h  e  G  N  W  .  W  h  e  n  t  h  e  n  e  x  t  s  t  i  m  u  l  u  s  i  s  p  r  e  s  e  n  t  ,  t  h  e  p  r  e  v  i  o  u  s  o  n  e  i  s  e  r  a  s  e  d  .  T  h  e  i  n  t  e  r  n  e  u  r  o  n  s   p  o  p  u  l  a  t  i  o  n  a  l  s  o  f  o  r  m  s  t  w  o  d  i  s  t  i  n  c  t  s  u  b  p  o  p  u  l  a  t  i  o  n  s  s  e  l  e  c  t  i  v  e  t  o  t  w  o  d  i  f  f  e  r  e  n  t  e  x  c  i  t  a  t  o  r  y  g  r  o  u  p  s  . c .  (Fig. 7a). In this case, we have identified a very specific range of spontaneous intrinsic activity (8)(9)(10)(11)(12)(13), as well as a narrow interval of the E/I ratio which is located between 70% and 95% (maximum accuracy achieved with ratio = 80% and rate = 12Hz).
The same study has been conducted for the delay conditioning task to assure that the model can achieve good performance on the same parameter values (see Supplementary Figure 1). It is a crucial step towards biological plausibility: the network can not perform the trace conditioning task without being able to work with delay conditioning beforehand. In our case, we have identified the region where both tasks are performed with high accuracy.

Discussion
In this paper, we address the issue of how the progressive complexification and integration of an artificial neural network gives rise to cognitive abilities. The current model is based on three hierarchical levels of information processing: the sensorimotor level, the cognitive level corresponding to the global nonconscious processing of information across different brain regions; and last the conscious level corresponding to the autonomous and lasting processing of information, even in the absence of externally applied sensory stimulation. This partition is consistent with the former distinction by Descartes and Kant of three levels of functional processing "Sensibility, Understanding and Reason" (see 31,32). More recently, Daniel Kahneman proposed the dichotomy of a fast "System 1", instinctive and emotional, and a slow "System 2", which is more deliberative and logical (6). In light of the need for grounding cognitive models in biology and validating their genuineness across levels, we followed successive mechanisms in known structural and functional organizations, identified through both the development and evolution of the human brain (42).
The first step proceeds from local to global integration of information; a second one moves from nonconscious to conscious processing. This leads to the three structurally and physiologically grounded levels: "perception," with the learning of invariants through local synaptic epigenesis, "cognition" with the learning of nonconscious integration through the long-range connectivity and reward, and finally, conscious processing with the learning of maintenance of representation online through interneurons and spontaneous activity. The model expresses key insights across the three levels of analyses: first, synaptic epigenesis, modeled here by selection and stabilization of synapses, is a critical mechanism at all levels, from perception to consciousness; second, dopamine is necessary for cognitive tasks to achieve proper credit assignment despite temporal delay between perception and reward; thirdly, interneurons allow the maintenance of self-sustained representation within the GNW in the absence of sensory input, thus enabling the system to solve conscious tasks. Finally, our results show how balanced spontaneous intrinsic activity facilitated epigenesis at both local and global scales, the balanced excitatoryinhibitory ratio increased performance. Those observations, in addition to emphasizing general principles at play in the human brain, are in surprising accordance with empirical observations.
For instance, without including synaptic pruning, the temporal evolution of synaptic weights leads to the three classical phases of growth, maximal variability, and stabilization (38). More unexpectedly, the optimal neurobiological parameters predicted for the conscious level are 20% of inhibitory neurons, which is predicted for balanced networks (43). Lastly, the "optimal" 12 Hz of spontaneous activity discovered in our artificial network coincides with the actual recordings of spontaneous spiking in the human cortex and is also strongly reminiscent of the highly conserved and omnipresent alpha rhythm in the cerebral cortex (44).
Computational neuroscience has increasingly relied on multi-scale brain models (45,46) to explain basic cognitive functions (47), active, top-down, and prospective memory retrieval (48,49), and even syntactic processing for the production of language (50). Some of these models, like ours, have employed spiking neural networks with STDP to study synaptic epigenesis during learning of sensory representation for tactile (51) and visual (37) perception, or to demonstrate how biases in natural statistics can influence population encoding and downstream correlates of behavior (52). These models, however, have tended to focus on a single learning mechanism, applied to a specific task. Here we combine both Hebbian and reinforcement learning across different canonical tasks. While the tasks considered here may appear minimalistic compared to recent sensational breakthroughs coming from artificial intelligence (53,54) our goal was not to achieve behavioral complexity but to uncover general biological principles through natural plausibility. Moreover, while it is certainly possible for artificial systems to mimic the behavior of a conscious agent, it is far less trivial to demonstrate the genuineness of its semantic understanding and subjective experience (55). From a purely pragmatic viewpoint, the Turing test (56) relies only on the judges being unable to tell if the agent is a human or a machine (57). We chose to view the problem in another way, by grounding the cognitive architecture in neurobiology and from there building the most parsimonious, realistic model capable of solving both perceptual and conscious tasks. This allowed delineating necessary and sufficient biological mechanisms for cognitive abilities in an artificial neural network (58). Future works should probe the role of other important biological mechanisms such as oscillations (59) or distributed coding for task value (60), and may likewise wish to take into account additional subcortical structures (beyond the striatum), such as the thalamus (61) and the basal ganglia (62). A promising avenue is also to explore the combination of the GNW architecture with predictive processing (63); the top-down feedback required for ignition and conscious access then becomes entangled with the prediction signal of the internal generative model of the world (64). Finally, the GNW architecture is also opening new venues in deep learning: on the one hand, by creating high-level inductive biases that improve out-of-distribution generalization (65,66), and, on the other, by providing an amodal latent space where alignment of representations across modalities become automatic (67).
Recently, a separate distinction was made between basic "conscious' processing and "the selfmonitoring of those computations, leading to a subjective sense of certainty or error" (68). This metacognitive dimension of consciousness is also linked to the hypothesis of the social origin of consciousness, which proposes the sense of self as a vestige of the evolutionary skills initially developed for understanding others (69). In this sense, the hominization of primates seems connected to molecular changes associated with social cognition (70) and, in line with our results, with the increase of long-range connectivity and regulation of the balance between excitatory and inhibitory neurons (71). Beyond these features, the prolonged postnatal brain development with a proper cascade of critical periods (72,73) leverages the multiple nongenetic interactions with the physical, social, and cultural environment, ultimately, giving rise to categorically human-specific cognitive abilities including the recursivity of language (71,74,75).
A key perspective is thus to further investigate the role of cultural embedding, adding explicitly a social dimension to the global neuronal workspace (17,76). Some dyadic computational models (77,78) have already been proposed using either continuous dynamical systems or discrete symbolic representation. Anchoring future computational models in both biological and social realities will not only continue to shed light on the core mechanisms underlying cognition, but it will also help to provide a unique bridge to artificial intelligence towards the only known systems with advanced social consciousness: the human brain (79). convolution is again resized to 128×128 pixels, thus creating an S1 layer with the size of 128×128×6.

Local Network
C1 layer. Each cell of this layer processes spikes from S1 cells in a 7×7 square (receptive field size of a given C1 cell). A C1 cell propagates only the single strongest orientation from its receptive field at a time. The receptive fields of two adjacent C1 cells are shifted by 6 pixels, meaning that they have an area of intersection of 7×1 pixels. This leads to a total C1 layer size of 25×25×2 (25×25 for the propagated values and 25×25 to store the orientations). However, the C1-S2 connections need to be set in a way that helps to identify not only the intensity of the propagated signal but the orientation. To implement this in a more biologically plausible way, the C1 layer has to be transformed. Each element on the 25×25 grid would now contain 6 C1 cells (one for each orientation) with the winner-takes-all condition: only one out of 6 cells can fire at a given time. Now, if we want to trace the signal back we can determine not only its spatial location but also its orientation. This modification leaves the C1 layer with a size of 25×25×6 with only 25×25 cells firing each time. We have also implemented the lateral inhibition procedure in a way it is done in the Masquelier and Thorpe model (37). The C1-S2 connections were governed by STDP.