HERMES OPTIMUS v3: 24-Word Edition
Expanded 30-50-12 architecture with 12-bit output decoding achieving 90.40% stress-suite accuracy and ~98% realistic typo accuracy on 24-word vocabulary
30-50-12 with 12-Bit Binary Output Decoding (V3)
Click to enlarge - V3 shows 12 output bits mapped to 24 words via Hamming codebook
- Input Layer: 30 neurons
- Positions 0-3 (4 neurons): Sorted positional encoding (normalized values 0-1)
- Positions 4-29 (26 neurons): Binary presence flags for A-Z
- Hidden Layer: 50 neurons with sigmoid activation
- Output Layer: 12 neurons, interpreted as bit channels for 24-word binary decoding
V3 maintains the proven 30-neuron hybrid input encoding from V2 but reinterprets the 12 output neurons as bit channels. Each output is rounded to binary (0 or 1), then the resulting 12-bit pattern is matched against a 24-entry Hamming codebook to select the nearest word. This innovation doubles vocabulary capacity from 12 to 24 words without expanding the output layer, critical for TI-84 memory constraints.
Evolution: v1.0 → v2.0 → v3
v1.0: 4-60-12 Positional Baseline
- 4 input neurons (positional encoding only)
- 60 hidden neurons
- ~85% accuracy, poor on scrambled words
v2.0: 30-50-12 Hybrid Robustness
- 30 input neurons (hybrid: 4 positional + 26 binary presence)
- 50 hidden neurons (memory-optimized)
- 96.86% accuracy - robustness breakthrough
- 12-word direct output classification
v3: 30-50-12 with Binary Codebook (Current)
- 30 input neurons (same hybrid encoding)
- 50 hidden neurons (same memory footprint)
- 12 bit-channel outputs → 24-word Hamming codebook
- 90.40% stress-suite accuracy, ~98% realistic typo accuracy
- 24-word vocabulary (double capacity, same memory)
Memory Optimization
Weight Matrix [I]:
30×50 = 1,500 weights
Weight Matrix [J]:
50×12 = 600 weights
Biases:
50 + 12 = 62 values
Total Memory:
~22.5KB (1.5KB remaining)
The architecture was precisely calibrated to fit within the TI-84's 24KB RAM limit. Reducing from 60 to 50 hidden neurons was necessary to accommodate the expanded 30-neuron input layer while maintaining sufficient overhead for program execution.