💻 computer science

Mutual Lateral Prediction for Locally Trained Spiking Neural Networks

This paper introduces MPC-SNN, a novel convolutional spiking neural network that achieves state-of-the-art accuracy on the DVS128 Gesture benchmark using strictly local, backpropagation-free learning by enabling neurons to predict the feedforward inputs of their peers through a learned lateral weight matrix.

Original authors: Nitesh Kamanuru Purushotham, Pranav Kulkarni

Published 2026-08-19

📖 4 min read☕ Coffee break read

CC BY 4.0

Original authors: Nitesh Kamanuru Purushotham, Pranav Kulkarni

Original paper licensed under CC BY 4.0 (https://creativecommons.org/licenses/by/4.0/). ✨ This is an AI-generated explanation of the paper below. It is not written or endorsed by the authors. For technical accuracy, refer to the original paper. Read full disclaimer

In the quest to build machines that see and think with the efficiency of the human brain, researchers have long looked to a special kind of artificial network called a spiking neural network. Unlike the standard computer programs that process information in a steady, continuous stream, these networks mimic the brain's own electrical language: brief, sharp bursts of activity known as spikes. This approach is particularly well-suited for new types of hardware designed to mimic biology, which can perform complex tasks while using a tiny fraction of the energy required by conventional computers. However, a major hurdle has remained. The most accurate way to teach these networks to recognize patterns currently relies on a method that requires the computer to send error signals backward through every layer of the network, a process that is impossible to run on the energy-efficient hardware itself and has no equivalent in the biological brain. Scientists have been searching for a way to train these networks using only local rules, where each part of the network learns from its immediate neighbors without needing a global signal from the top.

A team of researchers has now introduced a new system that solves this problem by teaching the network to predict what its neighbors are about to experience. In this new architecture, called MPC-SNN, every neuron does not just react to the signals it receives; it also tries to guess the signals that its peers in the same layer will receive. If the prediction is correct, the neuron stays quiet. If the prediction is wrong, the neuron fires a spike to signal the surprise. This mechanism allows the network to learn entirely from local interactions, using a specific rule where a neuron updates its connections based on its own error and the recent activity of a neighbor. The researchers tested this system on a dataset of hand gestures recorded by a special camera that sees only changes in light, a task that perfectly matches the strengths of this type of technology.

The results of the study show that this approach works remarkably well. On the test set of 288 gesture samples, the network achieved an accuracy of 95.83 percent, a performance that matches the best existing methods that do not use global error signals and significantly outperforms systems that have already been deployed on physical neuromorphic chips. Perhaps more importantly, the network became much more efficient as it learned. During the training process, the number of spikes it emitted dropped by more than 60 percent, meaning the final system uses far less energy to make a decision. This efficiency was not forced by a specific rule to be sparse; rather, it emerged naturally because the network learned to fire only when it was truly surprised by the input.

A key discovery in the study was that the mechanism allowing neurons to predict their neighbors' inputs was not just a theoretical idea but an active, growing part of the system. When the researchers started the training, the connections responsible for these predictions were nearly zero. As the network learned, these connections grew significantly, with the strongest layer increasing its predictive power by more than twenty times. This growth happened without any labels or external guidance, proving that the network was genuinely learning to anticipate the activity of its peers. The study also found that the system was highly efficient with the information it was given. Even when the researchers reduced the amount of labeled data used for training to just one-quarter of the original amount, the network retained nearly three-quarters of its accuracy, suggesting that the initial local learning phase had already built a robust understanding of the visual world.

While the system performed exceptionally well in simulation, the researchers were careful to note that the results are currently limited to computer models and have not yet been tested on physical neuromorphic chips. They also observed that the system struggled to identify gestures that were structurally very different from the training data, indicating that its ability to detect unusual inputs is still limited. Nevertheless, the work provides a concrete demonstration that a network can learn complex tasks using only local rules and predictive coding, offering a viable path toward building intelligent machines that operate with the low energy and high speed of biological systems. The success of this approach suggests that the future of efficient artificial intelligence may lie not in sending massive amounts of data backward through a network, but in teaching every part of the system to listen to and predict its neighbors.

Technical Summary: Mutual Lateral Prediction for Locally Trained Spiking Neural Networks

Problem Statement
Spiking Neural Networks (SNNs) are the natural computational model for neuromorphic hardware (e.g., Intel Loihi-2, SpiNNaker2), offering low-energy inference for event-based data like that from Dynamic Vision Sensors (DVS). However, the most accurate SNNs currently rely on Backpropagation Through Time (BPTT) with surrogate gradients. This approach requires propagating error signals backward across layers and through time, a mechanism incompatible with neuromorphic chips (which lack inter-layer gradient broadcast mechanisms) and biological plausibility (which lacks comparable backward pathways). While predictive coding offers a route to strictly local learning, existing spiking formulations typically allow neurons to predict only their own feedforward drive, neglecting the drive received by their lateral neighbors.

Methodology: MPC-SNN
The authors introduce MPC-SNN (Mutual Lateral Predictive Coding SNN), a convolutional SNN trained entirely with local, backpropagation-free rules. The architecture and learning process are defined by three core components:

Architectural Innovation (Lateral Prediction):
Unlike previous architectures where dendritic predictions rely solely on feedforward presynaptic spikes, MPC-SNN introduces a learned lateral weight matrix, $W_{lat}$ . Each neuron predicts the feedforward input of its same-layer peers using spikes emitted by those peers at the previous timestep.
- The dendritic estimate $\hat{I}[t]$ is calculated as: $\hat{I}[t] = W_d s_{pre}[t] + W_{lat} z_{layer}[t-1]$ .
- $W_{lat}$ is implemented as a learned $1\times1$ convolution over channels at the same spatial position.
- The prediction error $\epsilon[t]$ is the deviation between the actual presynaptic spike vector and this lateral/forward estimate.
Strictly Local Learning Rules:
The network utilizes three update rules derived from first principles, requiring no global error signals or external targets:
- Somatic Weight ( $W_s$ ): Updated via a three-factor Hebbian rule using prediction error, eligibility trace, and a neuromodulatory signal.
- Dendritic Weight ( $W_d$ ): Updated to minimize the feedforward component of the prediction error.
- Lateral Weight ( $W_{lat}$ ): The novel component, derived by differentiating a local objective (half the squared prediction error). The update rule $\Delta W_{lat} = \eta_{lat} \epsilon[t] e_{lat}[t]^\top$ requires only the neuron's own prediction error and the preceding spike of a same-layer peer. This is strictly more local than methods like e-prop (which broadcasts output errors) or LTTS (which requires external target spike trains).
Two-Phase Training Procedure:
- Phase 1 (Unsupervised Pretraining): 50 epochs with no class labels or global loss. Each layer updates its weights independently using local prediction errors. This phase aims to learn reusable representations and verify the activity of the lateral mechanism.
- Phase 2 (Supervised Fine-tuning): 400 epochs where a linear readout is appended, and the network is optimized end-to-end using cross-entropy loss and label smoothing.

Key Results
Evaluated on the 11-class DVS128 Gesture benchmark, MPC-SNN demonstrates the following:

Accuracy: The model achieves 95.83% test accuracy. This statistically matches the strongest local-learning baseline (DECOLLE, ~95.5%) and significantly exceeds deployed neuromorphic implementations (Loihi at 89.64% and Spike Gating Flow at 87.50%).
Mechanism Activity: During Phase 1, the lateral weights ( $W_{lat}$ ) grow from near-zero initialization ( $\approx 0.001$ ) by up to 21.1 $\times$ in the first layer, confirming the lateral rule is mechanically active without label-driven signals.
Label Efficiency: The model exhibits high label efficiency. With only 25% of the training labels, it retains 73.50% accuracy. This yields an Information Retention Ratio (IRR) of 2.94, meaning each labeled example provides nearly three times the accuracy gain of linear scaling.
Sparsity: During Phase 2, the spike count per inference decreases from ~730,000 to ~143,000, demonstrating emergent sparsity without explicit sparsity penalties in the loss function.
Ablation: Removing the lateral component (Condition B) results in a slight accuracy drop (94.79% vs 95.83% best) and a 5.8% increase in spike count, though these differences did not reach statistical significance on the 288-sample test set across five seeds.

Significance and Claims
The paper claims MPC-SNN is the first SNN to combine same-layer lateral predictive coding with a locally derived prediction-error rule. Its significance is framed in three distinct contributions:

Architectural: It introduces a learned same-layer lateral weight matrix to predict peer feedforward input, a construction not previously reported in SNN literature.
Theoretical: The lateral update rule is derived from a prediction-error objective rather than being heuristic, establishing a learning rule that is strictly local (requiring no inter-layer coordination or external targets).
Empirical: The model achieves competitive accuracy on a standard neuromorphic benchmark while demonstrating substantial label efficiency and active lateral weight growth during unsupervised pretraining.

Limitations and Honest Negatives
The authors explicitly note several limitations:

Statistical Significance: While the lateral mechanism shows large effect sizes (Cohen's $d=1.01$ ) in ablation studies, the small test set (288 samples) prevents these results from reaching statistical significance. A larger test set is required to definitively prove the lateral weight's reliability.
Convergence Issues: The prediction error convergence during Phase 1 was non-monotonic in the middle layer (conv2), suggesting the local objective may struggle with features that are neither purely local edges nor fully abstract.
Out-of-Distribution (OOD) Detection: An experiment testing spike count as an anomaly score for OOD detection yielded an AUROC of 0.578 (not significant). The authors report this as an "honest negative," attributing the failure to the semantic similarity between held-out and training classes rather than a fundamental flaw in the detection logic.
Deployment: The study is simulation-only; the efficiency of mapping Phase 1 training onto physical neuromorphic chips remains untested.

In conclusion, the paper presents MPC-SNN as a concrete, backpropagation-free realization of lateral predictive coding that bridges the gap between biological plausibility and high-performance neuromorphic computing, while acknowledging the need for further statistical validation and hardware deployment studies.

Technical Summary: Mutual Lateral Prediction for Locally Trained Spiking Neural Networks

More like this