We can easily name brain functions, and we are well informed about brain structure. However, it is not easy to bridge the gap between the two. Part of the problem is that simple circuit mechanisms do not directly give rise to high-level functions. Yet, they already implement simpler forms of information processing, a sort of “neural assembly language.”
In our Journal of Neuroscience paper, in collaboration with Andrea Brovelli (Marseille), we track such primitive operations of processing using metrics from information theory and benchmarking them on simulations of functional models (i.e. neural network models with architectures designed to emulate specific cognitive operations (working memory, selective attention, etc.). Information processing leading to different different functional computations is thus shown to have different « primitive flavors », i.e. different mixes of raw low-level computations of information storage, transfer and modification.
We thus transform descriptions of neuronal dynamics into descriptions of how these dynamics specifically propagate and modify information, into ad hoc models but also in more complicated non-human-primate connectome-based models, where we find that feed-forward and feed-back connections have different computing roles (the first devoted primarily to transfer while the second to modification).
To know more:
- Voges, N., Lima, V., Hausmann, J., Brovelli, A., and Battaglia, D. (2023). Decomposing neural circuit function into information processing primitives. J. Neurosci., JN-RM-0157-23. 10.1523/jneurosci.0157-23.2023.