Mojo0869/ABSL
ABSL or Adaptive Bitshift learning is a lerning method only using Integers for AI i made. The goal is to make a good Integer Neuronr neuron - Mojo0869/ABSL
5 stars · 0 forks · 0 watchers
README
##ABSL - Adaptive bitshft learning ( v.1.0.0 XOR) ABSL is an experimental, 100% integer-only learning algorithm for neural networks, written from scratch in Rust. By avoiding floating-point math entirely, ABSL doesn't need an FPU. That makes it interesting for low-power embedded systems, 8-bit/16-bit microcontrollers, and neuromorphic hardware. Instead of a fixed learning rate, ABSL scales weight updates using an adaptive bit-shift based on the integer error magnitude. Solving XOR without floats XOR is normally solved with continuous gradients. I wanted to see if a purely integer-based approach could get there too. Using a 2-4-1 architecture (2 inputs, 4 hidden neurons, 1 output) instead of the minimal 2-2-1 gives the network enough redundancy to work around the rounding errors that come from integer quantization. ##Benchmarks Metric---------------------- v5 Perfekt runs (4/4)---------- 98.7% Glaobal Accuracy------------ 99.7% Avg. correct cases / run---- 3.99 total failure (</=2 correct) 0.0% (look in results/ at the last dierctorys) ##Accuracy by Input
(0,0): 99.9% (0,1): 99.9% (1,0): 99.2% (1,1): 99.6%
How does it Work? I'm a bit lazy so look in src/learning_rule/absl.rs for the Code The shift amount is calculated from the bit-length (ilog2) of the error and weight, which keeps weights from exploding while staying fast. ##Roadmap
Try to get the best run possible: 99%+ done (99.7% global accuracy with v5) Try to scale it to MNIST
##About me I am Mojo 15 years old and live in north germany im currently in 10th grade and love Computers an AI My Goal is to study an land a good job and live in Japan ##License Costum License, look LICENSE