A 200-clause Tsetlin Machine MNIST digit classifier implemented entirely on a Tang Nano 9K FPGA board, running inference through a UART interface. The design receives 98 raw image bytes and outputs a single predicted digit (0-9) with 100% accuracy on a 100-sample test batch.
Researchers demonstrate that multi-agent systems with test-time communication substantially outperform independent parallel agents on challenging problem-solving tasks, with benefits compounding as team size scales. A team of k communicating agents matches performance of 4k independent agents, and communication enables solving problems that individual agents cannot, with gains validated on benchmarks like ARC-AGI-3, polyomino packing, and MNIST classifier compression.
A browser-based spiking neural network inspired by Drosophila connectome statistics classifies handwritten digits using sparse connections and event-driven computation. The network achieves ~112× estimated energy savings compared to dense networks by using only addition operations, running entirely via ONNX Runtime Web.
Researchers demonstrated a binarized neural network achieving 99.42% accuracy on handwritten digit recognition running on a stock 1975 Commodore 64 with 64 kB RAM and no multiply instruction. The model uses extreme quantization and confidence-gated early exit to reduce inference time by 60% while maintaining accuracy on real hardware.