Sakana AI’s Error Diffusion Trains Dale-Compliant Dual-Stream Networks, Reaching 96.7% MNIST and 61.7% CIFAR-10 Without Backpropagation
Sakana AI presented an error diffusion method that trains dual‑stream excitatory/inhibitory networks respecting Dale’s principle, achieving 96.7% accuracy on MNIST and 61.7% on CIFAR‑10 without using backpropagation.
- Error Diffusion trains Dale‑compliant dual‑stream networks without backpropagation.
- The method achieves 96.7% MNIST and 61.7% CIFAR‑10 accuracy, a notable result for backprop‑free learning.
- It aligns neural training with biological constraints, offering potential benefits for neuromorphic hardware.
Sakana AI announced a novel training approach called Error Diffusion, which eliminates the need for traditional backpropagation by routing modulo errors between excitatory and inhibitory pathways. The method respects Dale’s principle, ensuring each neuron is either purely excitatory or inhibitory, a constraint more aligned with biological neural circuits.
Using this technique, the researchers trained dual‑stream networks on the MNIST digit classification task, reaching 96.7% accuracy, and on the more challenging CIFAR‑10 dataset, achieving 61.7% accuracy, both without any gradient‑based weight updates. The paper also explores extensions to reinforcement learning scenarios and provides ablation studies that highlight the importance of task‑specific error routing.
The work positions itself as a step toward more biologically plausible learning algorithms, potentially reducing the computational overhead associated with backpropagation and opening new avenues for neuromorphic hardware implementations.
Provides a new training paradigm that could simplify hardware requirements for neural networks.
Shows innovative research that may lead to future AI hardware or algorithmic products.
Introduces a biologically inspired learning method useful for coursework and research projects.
Demonstrates progress toward more brain‑like AI learning without costly backpropagation.
- Dale's principle
- A neurobiological rule stating that a neuron releases the same type of neurotransmitter at all its synapses, making it either excitatory or inhibitory.
- Error diffusion
- A training technique that propagates modulo errors between network streams instead of using gradient backpropagation.
AI ResearchGoogle Cloud’s Always-On Memory Agent Replaces RAG and Embeddings With Continuous LLM Consolidation on Gemini 3.1 Flash-Lite
Integrating omics and artificial intelligence in pediatric environmental health: tools, challenges, and cohort-based insights - Nature
AI ResearchCan a Semantic Cache Become Your Primary Retrieval Layer?
Artemisia, Matteo Basilé and artificial intelligence trained to dream - osservatoreromano.va
Joanna Stern: An open letter to OpenAI’s secret gadget - NBC News
BusinessThe Pentagon's new AI playbook treats slow adoption as a bigger risk than "imperfect alignment"
The US Department of the Navy has signed a strategy to integrate AI into its operations, prioritizing speed over perfection.
Palantir Stock Slipped 35% From Its Peak. Is the Artificial Intelligence (AI) Software Leader a Safe Buy for the Second Half of 2026? - Yahoo Finance
Palantir's stock has slipped 35% from its peak, sparking concerns about its safety as a buy for the second half of 2026.
Chinese chip start-up Biren bets on light-based ‘supernodes’ to match Nvidia - South China Morning Post
Chinese chipmaker Biren is developing light-based supernodes to compete with Nvidia's performance. This strategy aims to bypass current hardware limitations.
BusinessAnthropic slashes Claude Fable 5 limits in Max and Team Premium and pushes Pro users toward API pricing
Anthropic has reduced the limits for Claude Fable 5 in its Max and Team Premium plans, while offering Pro users a one-time credit and API pricing.
Why Colorado replaced its AI discrimination law with a transparency requirement that the feds might challenge anyway - Yellow Scene Magazine
Colorado has replaced its AI discrimination law with a transparency requirement, which may face federal challenges.
Claude make Fable 5 permanent
Anthropic is making Claude Fable 5 a permanent fixture in its Max and Team Premium tiers starting July 20, offering it at 50% of usage limits.