Abel Wajnerman Paz – Access consciousness as a key to understand neural coding in non-sensory systems

Access consciousness as a key to understand neural coding in non-sensory systems – Abel Wajnerman Paz (FONDECYT/UAH/UBA)


Studies on consciousness in cognitive neuroscience often focus on explaining conscious processes or states by characterizing their underlying neural mechanisms. The efficient coding tradition in computational neuroscience proposes an inversion of the explanatory relation between neural and cognitive phenomena. The main idea is that we can understand the organization and operations of a given brain structure by characterizing the cognitive task it performs. For instance, the specific response properties of early sensory cells have been explained by the demand of reducing the redundancy in incoming perceptual signals. I will propose to apply this approach to the relation between access consciousness (AC) and its underlying mechanisms. Based on Dehaene’s Global Workspace hypothesis, I will suggest that AC can be used to understand a puzzling aspect of neural coding in non-sensory areas.

Dehaene and colleagues postulate the existence of a global network or a ‘global workspace’ (GWS) constituted by a set of cortical neurons that send and receive projections to many distant areas through long-range excitatory axons. These GWS neurons break the (relative) modularity of the cortex by maximizing the ability to exchange information between processors. Information that is encoded in GWS neurons can quickly be made available to many (or even all) brain systems (e.g. Dehaene & Changeux 2004). Although the long-range axons of GWS representations imply that these have no sharp anatomical delineation, they mostly originate from the pyramidal cells of layers 2 and 3 that are particularly dense in prefrontal, parieto-temporal and cingulate associative cortices (Dehaene, Changeux & Naccache 2011). Dehaene hypothesizes that the entry of inputs into this global work space constitutes the neural basis of access to consciousness.

It has been shown that the anterior areas where GWS neurons are denser implement a localist code. In populations with localist codes information can be decoded from single cells. The implementation of this coding strategy is puzzling because it has been shown that it is metabolically very expensive: the same information could be represented by a much smaller population if a distributed code (in which information cannot be decoded from single units) was implemented. I will propose that this could be explained by the demand of maximizing the accessibility of information imposed by the GWS.

The obvious way to maximize accessibility (i.e., the number of brain structures to which I can send a given piece of information) is to maximize the redundancy of GWS representations (i.e., the number of cells that carry the same information). Each additional redundant cell can send the same information to an additional brain location. In turn, an obvious way to maximize redundancy is by maximizing the density of GWS representations (i.e., the number of active cells that constitute a given representation). By increasing the number of cells that constitute a representation we can increase the number of projections it has to different brain areas (i.e., make its content more accessible). However, Dehaene’s model imposes a significant limit to density maximization. GWS representations are supposed to be sparse (a very small proportion of GWS cells is active at any given time). This is because different sets of GWS cells must represent different pieces of information that compete for acquiring conscious access. Only the winning representation will be active.

Instead of maximizing density, an alternative coding strategy to maximize redundancy is increasing what we can call the ‘compression’ of information in a representation. This is the minimum number of cells from which information can be decoded. If information is maximally compressed in a given representation, then it could be obtained by decoding any of the units that constitute that representation. In this compressed representation accessibility is maximized because each unit would be sufficient to send the relevant information to a different brain structure. As it is obvious, in these maximally compressed representations the information is encoded in a localist manner. Therefore, the informational demands and constraints imposed by the GWS, could potentially explain the implementation of a localist regime.