The combination of Hebbian and predictive plasticity learns invariant object representations in deep sensory networks

Abstract

Discriminating distinct objects and concepts from sensory stimuli is essential for survival. Our brains accomplish this feat by forming meaningful internal representations in deep sensory networks with plastic synaptic connections. Experience-dependent plasticity presumably exploits temporal contingencies between sensory inputs to build these internal representations. However, the precise mechanisms underlying plasticity remain elusive. We derive a local synaptic plasticity model inspired by self-supervised machine learning techniques that shares a deep conceptual connection to Bienenstock-Cooper-Munro (BCM) theory and is consistent with experimentally observed plasticity rules. We show that our plasticity model yields disentangled object representations in deep neural networks without the need for supervision and implausible negative examples. In response to altered visual experience, our model qualitatively captures neuronal selectivity changes observed in the monkey inferotemporal cortex in-vivo. Our work suggests a plausible learning rule to drive learning in sensory networks while making concrete testable predictions.

Sequence anticipation and spike-time-dependent-plasticity emerge from a predictive learning rule

“…From this optimization problem we derived a learning rule which gave rise to experimentally observed STDP mechanisms. Our results, together with previous studies [69, 70, 71], suggest that STDP is a consequence of a general learning rule given the particular state of the system, the stimulation protocol and the specific properties of the input. As a consequence, several STDP learning windows which are described by other phenomenological rules are predicted by our model, as well as the dependence on synaptic strength and depolarization level.…”

Local minimization of prediction errors drives learning of invariant object representations in a generative network model of visual perception

“…Networks trained on the static frames of the sequences, in which activity was reset after each frame also lacked a block-diagonal structure (Fig S2), illustrating the role of continuous motion in the training paradigm, which is to provide the necessary temporal structure in which subsequent inputs can be assumed to be caused by the same objects. The Hebbian learning rule thus groups together consecutive inputs in a manner reminiscent of contrastive, self-supervised methods (34,35,58) that explicitly penalize dissimilarity in the loss function. Here, the higher-level representation from the previous timestep provides a target for the consecutive inputs reminiscent of implementations of supervised learning with local learning rules (59–61).…”

“…At the same time, the model acquired generative capacity that enables reconstruction of partially occluded stimuli, in line with retinotopic and content-carrying feedback connections to V1 ( (13,36), see also (3) for a review of predictive feedback mechanisms). Other neuronlevel models of invariance-learning (28,29,31,34) neither account for such feedback nor experimentally observed explicit encoding of mismatch between prediction and observation (10,35) and used considerably more complex learning rules requiring a larger set of assumptions (34). By solving the task of invariance learning in agreement with the generativity of sensory cortical systems, the claim for predictive coding circuits as fundamental building blocks of the brain's perceptual pathways is strengthened We argue that the model generalizes predictive coding to moving stimuli in a biologically more plausible way than other approaches (16,42,46) that rely on non-local error backpropagation (41) or backpropagation through time (46).…”

“…Indeed there is evidence for the importance of temporal stimulus continuity for learning of transformation-tolerance in early visual areas of chicks (32) and area IT of monkeys (33). Based on this principle of representing consecutive inputs similarly, Halvagal and Zenke (34) recently showed that a more intricate learning rule with additional variance maximization leads to disentangled high-level representations. However, all of these models process information in a strictly feedforward manner or limit the role of feedback connections to a modulatory function, in contrast to evidence on retinotopic, content-carrying feedback connections in the brain (35)(36)(37).…”

Forward Learning with Top-Down Feedback: Empirical and Analytical Characterization

“…Several training rules for artificial neural networks (ANNs) have been proposed to break the weight symmetry constraint required by BP (Lillicrap et al, 2016b;Nøkland, 2016;Liao et al, 2016;Nøkland & Eidnes, 2019;Frenkel et al, 2021;Hazan et al, 2018;Kohan et al, 2018;Clark et al, 2021;Meulemans et al, 2021;Halvagal & Zenke, 2022;Journé et al, 2022). The Feedback Alignment (FA) algorithm (Lillicrap et al, 2016b) replaces the transposed forward weights W in the feedback path with random, fixed (nonlearning) weight matrices F , thereby solving the weight transport problem (Fig.…”