NeuronSoup: Evolving Asynchronous, Shared-Neuron Temporal Graphs without Backpropagation

2026-07-16 · Source: Machine Learning · Field: Technology & Digital — Artificial Intelligence & Machine Learning · Depth: Expert, quick

Summary

NeuronSoup introduces a neural computation architecture that replaces synchronous layer-by-layer processing with asynchronous, delay-mediated signal propagation through a pool of shared neurons. Paths route continuous-valued signals, where shared hidden neurons create constructive or destructive interference based on signal polarity and arrival timing. The entire architecture, including topology, weights, delays, and connectivity, is co-evolved by a genetic algorithm operating on a flat real-valued genome of 14,602 genes. On 10-class MNIST digit classification using frozen ResNet18 features, the system achieved 85.9% test accuracy after 10,000 generations, utilizing 204 active paths through 266 hidden neurons (156 shared). The trained model occupies a compact 115 KB. This architecture requires no differentiable computation graph, adapts its computation depth per-sample, and discovers lateral interactions.

Key takeaway

For AI Architects designing novel neural architectures, NeuronSoup offers a compelling alternative to traditional backpropagation-based deep learning. You should explore its asynchronous, shared-neuron approach for developing compact, adaptive models that inherently discover lateral interactions. This method avoids differentiable computation graphs, potentially simplifying complex model design and enabling per-sample adaptive computation depth, which could be critical for resource-constrained or dynamic environments.

Key insights

NeuronSoup evolves asynchronous, shared-neuron temporal graphs without backpropagation, adapting computation depth per-sample.

Principles

• Shared neurons create interference based on signal timing and polarity.
• Genetic algorithms are optimal for evolving complex, non-differentiable architectures.

Method

A genetic algorithm co-evolves network topology, weights, delays, and connectivity on a 14,602-gene genome, enabling asynchronous signal propagation through shared neuron pools.

In practice

• Classify 10-class MNIST digits using frozen ResNet18 features.
• Generalize to new domains by swapping encoder and output structures.

Topics

• NeuronSoup
• Genetic Algorithms
• Asynchronous Neural Networks
• Temporal Graphs
• MNIST Classification
• Model Evolution
• Shared Neurons

Best for: Research Scientist, AI Scientist, AI Architect

Related on AIssential

• On De-Individuated Neurons: Continuous Symmetries Enable Dynamic Topologies — Continuous symmetries enable dynamic neural network topologies with real-time growth and pruning while preserving function.
• EvoForest: A Novel Machine-Learning Paradigm via Open-Ended Evolution of Computational Graphs — EvoForest evolves computational graphs and functions, guided by diagnostics, to discover optimal computations for structured prediction.
• Asynchronous AI cuts computing energy by orders of magnitude while learning continuously — Asynchronous Neural Turing networks (ANT) significantly cut AI energy by mimicking brain's selective neuron updates.
• Cost Accounting for Reactive Computational Graphs: Exhaustive Sweeps, Sequential Mutation, and the Backward-Locality Gap — Cost accounting for reactive computational graphs reveals non-constant speedups for sweeps and specific behaviors for sequential mutations.
• Beyond KV-Cache: Test-Time Training & Invariant Latent Topologies for ICL — New methods enable LLMs to learn during inference and generalize across diverse domains by modifying internal weight structures or identifying invariant neural subspaces.
• Google hints at New Topological Flat Transformer — A new transformer architecture based on continuous flow dynamics promises enhanced reasoning and memory efficiency.

See Counsel's argued verdicts on the open AI decisions leaders are weighing →

Open in AIssential →

Editorial summary, takeaway, and curation by AIssential. Original article published by Machine Learning.