10th Grader Builds Integer-Only Neural Net in Rust, Solves Quantization Dead Zone

A high school student developed Green-AI, a Rust-based project exploring energy-efficient AI training using only integer arithmetic instead of floating-point operations. During development, they discovered a 'dead zone' problem where weight updates silently stalled once errors became too small, rounding down to zero under a fixed bitshift operation. To fix this, they devised Adaptive Bitshift Learning (ABSL), which dynamically adjusts the shift amount based on error magnitude — amplifying small updates and damping large ones. Benchmarks comparing ABSL against a fixed-shift approach and stochastic rounding showed ABSL achieved the highest accuracy and faster per-step performance than stochastic rounding. The project's source code has been published on GitHub, and the student is seeking feedback from those experienced in quantized or fixed-point neural network training.