Modelling collective phenomena in neuroscience

This paper shows how concepts and tools borrowed from statistical physics might be useful for modelling biological systems. This is illustrated by specific examples taken from the study of neural dynamics. The relevance for neuroscience of the notion of 'collective phenomena', for which statistical physics gives appropriate modelling tools, is discussed. Short-term memory is probably the simplest form of human (and not only human) memory. It also shows remarkable features that demand explanation, among which are: • associativity: what has been stored in memory is retrieved from partial or related information. Unlike standard computer memory, human memory is thus a content-addressable memory (CAM) – interestingly, since the early 1990s, CAM hardware has been under development for internet routers, allowing for very fast processing 1 • learning: working memory is permanently storing new events/objects, and this without getting doomed by its (necessarily finite) maximal capacity. This implies the ability to forget • persistent neural activity: as shown by electrophysiological recordings in monkeys during delayed response tasks, 2 on short timescales a neural activity can be maintained while keeping its specificity to a stimulus despite the noise in the network dynamics. In the late 1940s, Donald Hebb gave general qualitative accounts of neural organisation, dynamics and mechanisms that could explain the behaviour of human memory at the psychological level. 3 Hebb is mostly quoted for his views on the learning mechanism, the 'associativity' at the neural level resulting from strengthening of the coupling between cells that are simultaneously activated. 4 This, made explicit as a conjecture on synaptic plasticity, 5 has led to the experimental search for the 'Hebbian synapse', with the discovery of the strengthening (long-term potentiation, LTP) and weakening (long-term depression, LTD) of synapses, and of how these synaptic adaptations depend on pre-and post-synaptic activity. 6 On the theoretical side, a large amount of work has been done in order to understand if and how such learning dynamics can lead to learning associations. Efficient variants of the original Hebb proposal, such as the covariance rule and other closely related rules, in particular the so-called BCM rule, have been shown to be compatible with experimental data. 7 But Hebb did more than just suggest the existence of an adaptive mechanism modifying how neural cells interact. He made this proposal within a global analysis trying to explain human memory behaviour from the organisation and dynamics at the neural level. According to Hebb …