The M4 project : Memory Mechanisms in Man and Machine
Humans appear to be able to recognize visual and auditory stimuli that have not been experienced for decades. Thus, someone born in the 1950s might recognize the theme from an old TV series (e.g. “The Lone Ranger”) not seen since the early 1960s. What sort of biological mechanism could allow such extremely long-term memories to remain intact in the absence of reactivation? Until now this question appears to have been ignored, but I recently proposed a mechanism involving the formation of highly selective neurons that are totally silent unless the original stimulus is presented again. By remaining totally silent, the cells can keep their original selectivity for decades. The basic idea assumes a learning rule based on STDP (Spike-Time Dependent Plasticity, Masquelier and Thorpe, 2007) where the strength of synaptic inputs to a neuron only change if the neuron fires. We have already used simulations to show that such a rule, when combined with temporal coding (an idea that I have been developing for over two decades, see Guyonneau et al., 2005), allows simple integrate-and-fire neurons to become highly selective to repeatedly presented stimuli, even after only a few tens of presentations. The claim is that if a neuron is selective enough, it should never fire at all unless the original stimulus is presented again. And since synapses only change when there is a spike, this would allow the neuron to keep the trace stored during the original learning, even after several decades.
The idea is highly controversial. It implies both the existence of highly selective “grandmother cells” in the neocortex, and the idea that large numbers of cortical cells may simply never fire at all, constituting a sort of cortical “dark matter”. There is virtually no direct evidence for either. However, it would appear that no other hypothesis exists that could explain how memories could remain intact for so long without reactivation. The M4 project aims at testing and extending this radical hypothesis, and is funded with support from the European Commission FP7 , since May 2013.
We are now conducting a program of research to prove that memories can indeed persist for several decades without reactivation – currently, this is still a matter of conjecture. In particular we are investigating the precise conditions required in order for these extremely long-term memories to be formed, using a range of techniques including functional imaging and even single unit recording from epileptic patients to understand how brain structures respond during the reactivation of extremely remote memories. In parallel, we are also working on modelling studies to understand how temporal coding and STDP combine to allow very long term storage and extreme selectivity. In collaboration with other groups, we will use the results of the modelling and simulation studies to develop bio-inspired hardware systems that can reproduce the learning capacity of the brain using novel memristor-based technologies.