Abstract
If edge devices are to be deployed to critical applications where their decisions could have serious financial, political, or public-health consequences, they will need a way to signal when they are not sure how to react to their environment. For instance, a lost delivery drone could make its way back to a distribution center or contact the client if it is confused about how exactly to make its delivery, rather than taking the action which is "most likely" correct. This issue is compounded for health care or military applications. However, the brain-realistic temporal credit assignment problem neuromorphic computing algorithms have to solve is difficult. The double role weights play in backpropagation-based-learning, dictating how the network reacts to both input and feedback, needs to be decoupled. e-prop 1 is a promising learning algorithm that tackles this with Broadcast Alignment (a technique where network weights are replaced with random weights during feedback) and accumulated local information. We investigate under what conditions the Bayesian loss term can be expressed in a similar fashion, proposing an algorithm that can be computed with only local information as well and which is thus no more difficult to implement on hardware. This algorithm is exhibited on a store-recall problem, which suggests that it can learn good uncertainty on decisions to be made over time.
Introduction
As AI moves to "the edge"—powering autonomous drones, medical implants, and remote sensors—it faces two major hurdles: energy efficiency and uncertainty quantification (UQ). Traditional deep learning is power-hungry, whereas neuromorphic computing offers a low-power alternative by mimicking the brain's spiking nature. However, most AI models are "overconfident." If a delivery drone encounters a situation it wasn't trained for, it might blindly take a wrong action rather than signaling for help.
Training these networks on-chip (while they are deployed) is notoriously difficult because of the weight transport problem: standard backpropagation requires global information about the network that isn't physically available to individual neurons on a chip. This research seeks to bridge the gap by creating a learning rule that is both local (hardware-compatible) and Bayesian (capable of measuring uncertainty).
Methods: Local Bayesian Learning
The proposed approach builds upon e-prop 1, a learning algorithm for Spiking Neural Networks (SNNs) that solves the locality problem using "eligibility traces" and Broadcast Alignment (BA). BA replaces complex feedback weights with random ones to bypass the weight transport issue.
The core contribution of this work is the integration of Variational Bayesian Inference. Instead of training a single fixed value for every synaptic weight, the algorithm learns a distribution (a mean and a variance) for each weight. The authors mathematically demonstrate that the gradients required for this Bayesian approach can be computed using only local information. A neuron only needs to know its own state to update its "uncertainty" parameters. During learning, the network employs Monte-Carlo sampling, drawing different weights from these learned distributions to estimate the optimal path forward.
Experimental Setup: The Store-Recall Task
The researchers evaluated their approach using a "virtual mouse" Store-Recall Task. In this setup, a virtual agent perceives light through 20 spike channels (10 for "Up," 10 for "Down"). A signal is flashed briefly, followed by a long delay of up to 1.5 seconds. A subsequent "cue" signal tells the agent it must decide where the reward is. To succeed, the SNN must maintain a memory of the initial signal across the silent delay period.
To test uncertainty quantification, the authors presented a novel "Confusion" scenario where both "Up" and "Down" signals were triggered simultaneously—a situation the network never encountered during training.
Results and Discussion
The Bayesian SNN successfully learned to associate initial spikes with the correct later decision, effectively managing the temporal credit assignment problem. Crucially, adding Bayesian complexity did not compromise the locality required for neuromorphic hardware; the math for uncertainty fits within the physical constraints of on-chip learning.
In the conflicting signal test, the model's posterior distribution widened significantly. Unlike standard SNNs that might arbitrarily pick a direction, the Bayesian SNN showed high variance in its output, effectively "admitting" it didn't know the answer. While using Broadcast Alignment makes learning noisier than using perfect global information, the Kullback–Leibler (KL) cost still trended downward, proving the viability of on-chip Bayesian learning.
Limitations and Future Work
Current limitations include the use of minibatches, which may need adjustment for true "online" single-sample learning. Additionally, the sampling overhead (drawing multiple weight sets) impacts energy efficiency. While the algorithm is "hardware-ready," future work involves deployment on physical neuromorphic chips like Intel's Loihi, potentially utilizing multi-compartment neurons to run weight samples in parallel.
Summary updated