ETLP: Event-based Three-factor Local Plasticity for online learning with neuromorphic hardware

TL;DR

ETLP addresses online learning under locality constraints on neuromorphic hardware by combining a temporal gradient estimate with DRTP-based local updates, using a three-factor signal: pre-synaptic spike trace, post-synaptic surrogate gradient, and a teaching-label-driven trigger. It leverages ALIF neurons with adaptive thresholds and an event-driven update mechanism, enabling fully local, on-chip learning suitable for edge devices. Empirical results on N-MNIST and SHD show competitive accuracy with lower computational complexity than BPTT, with recurrence and threshold adaptation significantly boosting SHD performance. A proof-of-concept FPGA implementation demonstrates practical mapping of ETLP primitives to neuromorphic hardware, highlighting low latency and energy efficiency for real-time online learning at the edge.

Abstract

Neuromorphic perception with event-based sensors, asynchronous hardware and spiking neurons is showing promising results for real-time and energy-efficient inference in embedded systems. The next promise of brain-inspired computing is to enable adaptation to changes at the edge with online learning. However, the parallel and distributed architectures of neuromorphic hardware based on co-localized compute and memory imposes locality constraints to the on-chip learning rules. We propose in this work the Event-based Three-factor Local Plasticity (ETLP) rule that uses (1) the pre-synaptic spike trace, (2) the post-synaptic membrane voltage and (3) a third factor in the form of projected labels with no error calculation, that also serve as update triggers. We apply ETLP with feedforward and recurrent spiking neural networks on visual and auditory event-based pattern recognition, and compare it to Back-Propagation Through Time (BPTT) and eProp. We show a competitive performance in accuracy with a clear advantage in the computational complexity for ETLP. We also show that when using local plasticity, threshold adaptation in spiking neurons and a recurrent topology are necessary to learn spatio-temporal patterns with a rich temporal structure. Finally, we provide a proof of concept hardware implementation of ETLP on FPGA to highlight the simplicity of its computational primitives and how they can be mapped into neuromorphic hardware for online learning with low-energy consumption and real-time interaction.

Figures (10)

• Figure 1: Computational graph for a 5 time steps simulation on a 1-layer SNN using soft-reset LIF neurons and back-propagation through time. Where I, V and S are vectors representing the neuron input, voltage and output spike, respectively. $\text{S}^*$ is the target output and E is the error. The black arrows represent the forward pass of the data, while the curved orange arrows represent the gradient backward pass. The neuron resets is detached from the computational graph in the gradient calculation.
• Figure 2: Weight update mechanism on BPTT and DRTP. In the case of BPTT, and error is calculated using the output layer and the target $S^*$, the error is then propagated backward through the layers using the transpose of the synaptic weights. On DRTP, the error is only calculated at the output layer, the weights in the hidden layer are updated using directly the target, using random fixed matrix.
• Figure 3: Feedforward SNN with ETLP learning. The network consists of a hidden layer, an output layer and teaching neurons. The output layer is connected to the teaching neurons by both excitatory (green) and inhibitory (dashed red) synapses. The learning layer is also connected to the group of recurrent neurons via a random weight matrix B. When a class is active, the neuron in the learning layer starts to emit spikes at a certain frequency (e.g. 100Hz), thus modifying the synapses of the neurons according to the equations \ref{['eq:weight_update_hid']} and \ref{['eq:weight_update_out']}.
• Figure 4: Accuracy of BPTT, eProp and ETLP on N-MNIST using a feedforward network of LIF neurons. Fig. (a) shows a comparison in accuracy between the training and test datasets, Figs. (b) and (c) show the accuracy over all the epochs for training and test, respectively.
• Figure 5: Accuracy for different threshold adaptation on ETLP in SHD dataset with the recurrent network.