The backpropagation-based recollection hypothesis: Backpropagated action\n potentials mediate recall, imagination, language understanding and naming

Zied Ben HouidiarXiv 2021 · DOI · arXiv

Abstract: Ever since the advent of the neuron doctrine more than a century ago, information processing in the brain is widely believed to mainly follow the forward pre to post-synaptic neurons direction. Challenging this prevalent view, in this paper, we emit the backpropagation-based recollection hypothesis as follows: weak and fast fading Action Potentials following the (highest weight) post to pre-synaptic backward pathways, mediate explicit cue-based memory recall. This includes also the tasks of imagination, future episodic thinking, language understanding and associating names to various stimuli. These signals originate in highly invariant neurons, which uniquely respond to some specific stimuli (e.g. image of a cat). They then travel backwards to reactivate the same populations of neurons that uniquely respond to this specific stimuli during perception, thus recreating "offline" an experience that is similar. After stating our hypothesis in details, we challenge its assumptions through a thorough literature review. We find abundant evidence that supports most of its assumptions, including the existence of backpropagating signals that have interesting properties. We then leverage simulations based on existing spiking neural network models with STDP learning to show the computational feasibility of using such a mechanism to map the image of an object to its name with the same high accuracy as a state of the art machine learning classifier. Although not yet a theory, we believe this hypothesis presents a paradigm shift that is worth further investigating: it opens the way, among others, to new interpretations of language acquisition and understanding, the interplay between memories encoding and retrieval, as well as reconciling the apparently opposed views between sparse coding and distributed representations.

Introduction

For over a century, the "neuron doctrine" has dictated our understanding of the brain: information flows forward, from pre-synaptic to post-synaptic neurons. However, this prevalent view struggles to explain how the brain suddenly "re-lives" a visual memory or associates the sound of the word "cat" with a furry, four-legged image. This paper addresses the "signifier-signified" problem: how the brain links a label (the word "cat") to a concept (the mental image of a cat) such that one can trigger the other. Traditional models often rely on complex supervised "top-down" feedback loops that lack biological evidence. We propose a simpler, unsupervised mechanism where the same pathways used for perception are repurposed—in reverse—for retrieval.

The Backpropagation-Based Recollection Hypothesis

We propose that explicit memory recall and imagination are mediated by weak and fast-fading action potentials following the highest-weight post-to-pre-synaptic backward pathways. These signals originate in highly invariant "source pointer neurons"—selective cells (analogous to "Jennifer Aniston" neurons) that act as hubs. When triggered by a cue, these neurons send signals backwards to reactivate the exact population of neurons that fired during the original experience, recreating an "offline" experience that is similar to the original perception.

This mechanism offers a "middle ground" in the debate between sparse and distributed coding: memories are stored in distributed sensory ensembles but indexed by sparse pointer neurons. Furthermore, the "vividness" of memory is explained by the fading nature of these backward signals, which produce a shadow of the original sensory input. Neuromodulators like acetylcholine may act as a biological "switch" to amplify these signals during intentional recall or imagination.

Methodology and Computational Feasibility

To test the computational feasibility of this hypothesis, we utilized a Spiking Neural Network (SNN) with Spike-Timing-Dependent Plasticity (STDP). The model involves three distinct phases:

  1. Feature Learning: An unsupervised phase where the network learns to recognize visual features (lines, shapes) from raw input.
  2. Co-occurrence Learning: A semi-supervised phase where an image and its "name" (a category spike) are presented simultaneously, strengthening the synaptic connections between the two.
  3. Naming (The Test): When shown an image, the network uses backpropagating signals from the "root" layer to "vote" on the correct category.

The study primarily utilized the Caltech Motorbike/Face dataset for binary classification and referenced the MNIST dataset to ensure the feature extraction layers performed at state-of-the-art levels for unsupervised spiking networks.

Key Results

The backpropagation-based naming mechanism achieved 95.7% accuracy on the Caltech test set, performing on par with a standard Support Vector Machine (SVM) classifier (96%). Notably, in extreme "one-shot" learning scenarios (learning from a single image pair), the backpropagation model significantly outperformed the SVM, reaching 96.2% accuracy compared to the SVM’s 84.5%. Literature review confirms that backpropagating action potentials (BPAPs) are a measured biological reality and are known to be stronger when a neuron is already firing—a necessary condition for this hypothesis.

Discussion and Future Work

While the hypothesis presents a compelling paradigm shift, several hurdles remain. Experimental proof is needed to confirm that BPAP strength is strictly proportional to pre-synaptic weights across multiple "hops" in a neural chain. Additionally, while the model explains how we identify a name, it does not yet account for the motor coordination required to produce speech. Future work will focus on scaling this mechanism to complex scenes with multiple overlapping objects and accounting for the stochastic nature of real-world memory retrieval. Despite these limitations, the hypothesis opens new avenues for understanding language acquisition and the fundamental interplay between memory encoding and retrieval.

Summary updated

Abstract

Ever since the advent of the neuron doctrine more than a century ago,\ninformation processing in the brain is widely believed to mainly follow the\nforward pre to post-synaptic neurons direction. Challenging this prevalent\nview, in this paper, we emit the backpropagation-based recollection hypothesis\nas follows: weak and fast fading Action Potentials following the (highest\nweight) post to pre-synaptic backward pathways, mediate explicit cue-based\nmemory recall. This includes also the tasks of imagination, future episodic\nthinking, language understanding and associating names to various stimuli.\nThese signals originate in highly invariant neurons, which uniquely respond to\nsome specific stimuli (e.g. image of a cat). They then travel backwards to\nreactivate the same populations of neurons that uniquely respond to this\nspecific stimuli during perception, thus recreating "offline" an experience\nthat is similar. After stating our hypothesis in details, we challenge its\nassumptions through a thorough literature review. We find abundant evidence\nthat supports most of its assumptions, including the existence of\nbackpropagating signals that have interesting properties. We then leverage\nsimulations based on existing spiking neural network models with STDP learning\nto show the computational feasibility of using such a mechanism to map the\nimage of an object to its name with the same high accuracy as a state of the\nart machine learning classifier. Although not yet a theory, we believe this\nhypothesis presents a paradigm shift that is worth further investigating: it\nopens the way, among others, to new interpretations of language acquisition and\nunderstanding, the interplay between memories encoding and retrieval, as well\nas reconciling the apparently opposed views between sparse coding and\ndistributed representations.\n
BibTeX
@misc{houidi2021backpropagation,
  author = {Zied Ben Houidi},
  title = {{The backpropagation-based recollection hypothesis: Backpropagated action\textbackslash\{\}n potentials mediate recall, imagination, language understanding and naming}},
  year = {2021},
  eprint = {2101.04137},
  archivePrefix = {arXiv},
  url = {https://arxiv.org/abs/2101.04137},
}
Figure