Training CNNs Without Backpropagation Achieves SOTA Results
burkov · x · 2026-07-24
A convolutional neural network trained without backpropagation has achieved state-of-the-art results for this class of algorithms. The model reached 96.7% accuracy on MNIST and 61.7% on CIFAR-10.
The findings support continued investment in biologically constrained training methods, suggesting potential gains for neuromorphic hardware and built-in sparsity.
More from Research
- A 30-year graph theory conjecture was refuted after a GPT 5.6 Pro-assisted search — burny_tech · 2026-07-24
- A new atlas maps 2,226 coding tasks across 11 AI benchmarks — zainhas · 2026-07-24
- Poolside open-sources its eval data as its CEO says coding is the path to AGI — eliebakouch · 2026-07-24
- GEPA team releases omni, a meta-optimizer that combines multiple LLM optimizers — iamrobotbear · 2026-07-24
- Adhiraj Ghosh says task-adaptive batch sampling delivered a 3.33x pretraining compute multiplier — pratyushmaini · 2026-07-24
- Science study maps single-cell 3D genome rewiring in Alzheimer’s with a transformer model — jmuiuc · 2026-07-24