Contrastive-Signal-Dependent Plasticity: Self-Supervised Learning in Spiking Neural Circuits

TL;DR

This work addresses the challenge of designing neurobiologically motivated schemes for adjusting the synapses of spiking networks and proposes contrastive signal–dependent plasticity, a process which generalizes ideas behind self-supervised learning to facilitate local adaptation in architectures of event-based neuronal layers that operate in parallel.

Abstract

Brain-inspired machine intelligence research seeks to develop computational models that emulate the information processing and adaptability that distinguishes biological systems of neurons. This has led to the development of spiking neural networks, a class of models that promisingly addresses the biological implausibility and {the lack of energy efficiency} inherent to modern-day deep neural networks. In this work, we address the challenge of designing neurobiologically-motivated schemes for adjusting the synapses of spiking networks and propose contrastive-signal-dependent plasticity, a process which generalizes ideas behind self-supervised learning to facilitate local adaptation in architectures of event-based neuronal layers that operate in parallel. Our experimental simulations demonstrate a consistent advantage over other biologically-plausible approaches when training recurrent spiking networks, crucially side-stepping the need for extra structure such as feedback synapses.

Figures (8)

• Figure 1: Componential layout of one layer in the CSDP SNN recurrent circuit. A component diagram of a leaky integrate-and-fire (LIF) cell group/vector. The neuronal components include the electrical current $\mathbf{j}^\ell(t)$, the membrane voltage potential $\mathbf{v}^\ell(t)$, the spike emission $\mathbf{s}^\ell(t)$, the activation trace $\mathbf{z}^\ell(t)$, and the cross-layer homeostatically-constrained threshold $v^\ell_{thr}(t)$. Black (dashed) arrows with solid triangle heads indicate the flow of values within an LIF group's dynamics (e.g.., current $\mathbf{j}^\ell$ is input to dynamics of $\mathbf{v}^\ell(t)$), open-circle heads indicate excitation/additive pressure, and solid square heads indicate inhibitory/subtractive pressure.
• Figure 2: Neuronal layer within a CSDP-trained recurrent SNN. Depicted is the neural computation underlying one layer of our recurrent spiking circuit as well as its synaptic adjustment, induced via contrastive-signal-dependent plasticity. Note that the supervised variant of CSDP uses class-modulation synapses $\mathbf{B}^\ell$ whereas the unsupervised variant does not. To facilitate contrastive learning, we integrate another biochemical component $\mathbf{\delta}^\ell(t)$ into our leaky integrator cell model. Once electrical current $\mathbf{j}^\ell(t)$ has been injected into the cells, triggering an update to their membrane potentials $\mathbf{v}^\ell(t)$, possibly leading to emission of action potentials $\mathbf{s}^\ell(t)$, their glutamate traces contribute to a goodness modulator $\mathcal{C}^\ell(t)$ which then deposits a local synaptic adjustment signal to each cell's $\delta^\ell_i(t)$ component. Solid lines indicate synaptic pathways; those with open-circle heads indicate excitation/additive pressure, solid square heads indicate inhibitory/subtractive pressure and diamond heads indicate modulation/multiplicative pressure. Dashed lines (with solid triangle heads) indicate non-synaptic pathways (where no transformation is applied).
• Figure 3: Positive and negative modes of CSDP plasticity in a recurrent SNN. Shown are the two parallel 'modes' of plasticity that our recurrent spiking circuit undergoes -- a positive measurement mode is induced when a sensory input and its corresponding (if available) target class are used to drive inference and plasticity (this entails neural layers increasing the probability values assigned to a sensory pattern), and a negative measurement mode is induced when a negative / out-of-distribution sample and its corresponding negative target class are used to drive the same underlying neural calculations (this entails neural layers lowering probabilities that they ultimately assign). Solid square arrows denote inhibitory pressure whereas open circle arrows denote excitatory pressure. Colors for different synaptic pathways denoted by solid lines, all of which are plastic, are provided to visually distinguish between top-down class modulating synapses (green), bottom-up sensory driving synapses (purple), and lateral synapses (black).
• Figure 4: Latent activity patterns acquired by CSDP-trained SNNs. t-SNE visualizations of the latent space induced by recurrent spiking networks learned with the CSDP process. Note that t-SNE coordinate units are technically dimensionless units and we thus denote them as "tSNE units"). Rate codes are visualized for: (A) 'CSDP, Sup' (supervised CSDP-trained SNN) on MNIST, (B) 'CSDP, Unsup' (unsupervised CSDP-trained SNN) on MNIST, (C) 'CSDP, Sup' on K-MNIST, and (D) 'CSDP, Unsup' on K-MNIST.
• Figure 5: Reconstruction ability of a CSDP-trained SNN. Reconstructed samples produced by the SNN when learned with CSDP. The top row (A) presents intercalated original images taken from the MNIST database (randomly sampled) with CSDP model reconstructed images; columns alternate between original and reconstruction patterns (starting with original images in the leftmost column). The bottom row (B) presents original images and reconstructed values for the K-MNIST database (formatted in the same way as sub-figure A).