Backpropagation-Based Recollection of Memories: Biological Plausibility and Computational Efficiency

Since the advent of the neuron doctrine more than a century ago, information processing in the brain is widely believed to follow the forward pre to post-synaptic neurons direction. Challenging this view, we introduce the backpropagation-based recollection hypothesis as follows: Cue-based memory recollection occurs when backpropagated Action Potentials (APs), originating in sparse neurons that uniquely activate in response to a specific trace being recalled (e.g. image of a cat), travel backwards. The resulting transient backpropagating currents follow the available open backward and lateral pathways, guided by synaptic weights or couplings. In doing so, they stimulate the same neurons that fired during the very first perception and subsequent encoding, effectively allowing a "replay" of the experience (e.g., recalling the image of the cat). This process is pervasive, seen in tasks like cue-based attention, imagination, future episodic thinking, modality-specific language understanding, and naming. After detailing our hypothesis, we challenge it against a thorough literature review, finding compelling evidence supporting our claims. We further found that gap junctions could be a plausible medium for such currents, and that cholinergic modulation, which is known to favour backpropagated APs and is crucial for memory, is a reasonable candidate trigger for the entire process. We then leverage computer simulations to demonstrate the computational efficiency of the backpropagation-based recollection principle in (i) reconstructing an image, backwards, starting from its forward-pass sparse activations and (ii) successfully naming an object with a comparable high accuracy as a state of the art machine learning classifier. Given the converging evidence and the hypothesiss critical role in cognition, this paradigm shift warrants broader attention: it opens the way, among others, to novel interpretations of language acquisition and understanding, the interplay between memory encoding and retrieval, as well as reconciling the apparently opposed views between sparse coding and distributed representations, crucial for developing a theory of consciousness and the mind. Significance StatementTry to mentally picture the image of a cat. In this process, the word "cat" acted as a cue, and the fragile and non-persistent retrieved mental image is a recollected memory. Similar cue-based generative activities are ubiquitous in our lives, yet the underlying neural mechanisms are still a mystery. Neuroimaging and optogenetic-based studies suggest that cue-based recollection of memories involve the reactivation of the same neural ensembles which were active during perception (encoding). However, the exact neural mechanisms that mediate such reactivation remain unknown. We elaborate a novel hypothesis explaining how this can be implemented at single neurons: we hypothesize that the very same neural pathways used for perception are used backwards for recall, thus creating similar impressions during retrieval.