subbyte-quantization

Sub-Byte 4-Bit (s4) & Codebook LUT Quantization

Demonstrates ultra-low Flash quantization techniques for microcontrollers with extreme storage constraints using embedded-nn.

---

Quantization Capabilities

1. Signed 4-Bit (s4) Linear Packing:
   • Packs two 4-bit signed weights [-8..=7] into a single byte (pack_s4_pair / unpack_s4_pair).
   • Cuts Flash weight storage by 50% relative to standard INT8 CMSIS-NN formats.
   • Zero-allocation direct execution via fully_connected_s4 and convolve_s4.
2. Nonlinear 16-Entry Codebook LUT (s4_lut):
   • 4-bit nibbles index into a 16-entry codebook centroids table.
   • Preserves non-uniform and bell-curve weight distributions with negligible accuracy loss.

---

Running the Example

cargo run --package embedded-nn --example subbyte_quantization

Or from within this directory:

cd examples/subbyte-quantization
cargo run