Abstract
We propose a solution to the weight transport problem, which questions the biological plausibility of the backpropagation algorithm. We derive our method based upon a theoretical analysis of the (approximate) dynamics of leaky integrate-and-fire neurons. Our results demonstrate that the use of spike timing alone outcompetes existing biologically plausible methods for synaptic weight inference in spiking neural network models. Furthermore, our proposed method is more flexible, being applicable to any spiking neuron model, is conservative in how many parameters are required for implementation and can be deployed in an online-fashion with minimal computational overhead. These features, together with its biological plausibility, make it an attractive mechanism for weight inference at single synapses.
Introduction: The Weight Transport Problem
In modern artificial intelligence, backpropagation is the gold standard for training neural networks. However, it faces a significant hurdle regarding biological plausibility known as the Weight Transport Problem.
Standard backpropagation requires that the weights used in the feedback (error) pass are identical to those used in the forward pass. While this is trivial in digital computers, biological synapses are unidirectional and physically distinct. There is no known biological mechanism that allows a feedback synapse to precisely "copy" the strength of a distant forward synapse.
To bridge the gap between high-performance AI and biological reality, we need mechanisms that allow synapses to infer these weights using only local information—signals available at the synapse itself.
Methods: Spike-Timing-Dependent Weight Inference (STDWI)
The authors introduce Spike-Timing-Dependent Weight Inference (STDWI), a learning rule derived from the physics of a Leaky Integrate-and-Fire (LIF) neuron.
Theoretical Derivation
The method begins with a deterministic analysis of how a single input spike affects the "time-to-spike" of a post-synaptic neuron. To ensure robustness in the noisy environment of the brain, this complex physics is simplified into a rule based on eligibility traces.
The Fast and the Slow Traces
Each neuron maintains two distinct traces of its activity:
- Fast Trace: Captures the instantaneous firing rate, representing immediate activity.
- Slow Trace: Captures the average firing rate, serving as a baseline.
The Learning Rule
When a pre-synaptic neuron spikes, the synapse updates its weight estimate based on the difference between the post-synaptic neuron's fast and slow traces. If the neuron fires earlier than its average "schedule," the synapse infers a strong positive connection. This approach is event-driven and operates online, learning continuously without the need for batch processing.
Experimental Evaluation
The study evaluated STDWI using simulated spiking neural networks (SNNs):
- Network Architecture: 100 input neurons connected to 10 output neurons using LIF models.
- Stimulation Protocols: Tested under "Sparse" (20% input activity) and "Dense" (100% activity) conditions to evaluate signal extraction from noise.
- Baseline Comparison: STDWI was compared against a mathematically optimal Bayesian baseline using a Wiener process model.
Key Insights
- Timing vs. Rate: While many previous methods relied on firing rates, this research demonstrates that the precise timing of spikes provides much richer information for synaptic weight inference.
- Bayesian Approximation: The simple two-trace rule effectively approximates complex Bayesian optimal updates, providing a computationally efficient way to perform sophisticated inference.
- Biological Mirroring: The rule resembles a mirrored version of traditional Spike-Timing-Dependent Plasticity (STDP), suggesting that the brain might utilize such mechanisms specifically to solve the weight transport problem.
- Model Agnostic: STDWI relies only on spike timing rather than internal state variables like voltage, making it applicable to a wide variety of spiking neuron models.
Results
- Superior Accuracy: STDWI outperformed state-of-the-art methods like Akrout (rate-based) and RDD (Regression Discontinuity Design) in both sign accuracy and Pearson correlation with true weights.
- Robustness and Efficiency: Unlike RDD, which degrades under dense stimulation, STDWI remains effective. Furthermore, it is highly efficient, requiring four times fewer parameters per synapse than RDD.
Conclusion and Future Work
STDWI provides a biologically plausible solution to the weight transport problem by leveraging local spike timing. While currently univariate—focusing on single input-output relationships—future work will explore:
- Multivariate Extensions: Accounting for interactions between multiple simultaneous inputs.
- Deep Networks: Testing the rule's ability to support hierarchical learning in multi-layer architectures.
- Neuromorphic Implementation: Deploying the event-driven rule on specialized hardware for energy-efficient AI.
Summary updated