Network Architectures

TinyMind provides a range of neural network architectures, all as header-only C++ templates with both fixed-point and floating-point support.

Architecture Memory (Q8.8, trainable) Key Advantage
LSTM, GRU & Liquid (LTC/CfC) 952 / 808 bytes Sequential data, temporal patterns; continuous-time cells for irregular sampling
Kolmogorov-Arnold Networks 1,192 bytes Learnable activation functions
Conv & Pooling Layers 1,825 bytes (1D pipeline) 1D time-series + 2D spectrogram/image feature extraction, MobileNet-style separable blocks
Linear Self-Attention ~6 KB (mid-range) Sequence dependency modeling without O(N^2)
FFT Layer 768 bytes (64-pt Q8.8) Frequency-domain feature extraction
Quantized Networks 128 bytes (packed binary) 32-64x weight compression
Int8 Affine Quantization int8 weights + int32 accum TFLite/CMSIS-NN style post-training int8 across dense/conv/pool/BN/LN/softmax/RNN/attention/FFT
Mixed Precision int8 + fp16 + bf16 bridges qbridge converters between int8 affine / QValue Q-format / float / fp16 / bf16
SIMD Backends n/a (perf, not capacity) ISA-capability gates: NEON / SVE / Helium / AVX2 / AVX-512, byte-identical to scalar
Mixture of Experts router + N resident experts Decouples compute from capacity: 1 of N experts runs per inference (top-1), or k-of-N softmax-blended

Table of contents


Back to top

Dan McLeran — danmcleran@gmail.com — MIT License

This site uses Just the Docs, a documentation theme for Jekyll.