Feature-Base-Jump-Encoding-for-Convolutional-Matrices

Some theorising

Base-Jump Feature Encoding for Neural Network Matrices

This repository demonstrates a novel technique for feature encoding in neural network (NN) matrices.

Summary

Instead of expanding matrix size (by adding channels) or increasing compute via multiplication, this method encodes the presence of features by incrementing matrix values by a fixed "base jump" whenever a feature is detected.
This can help compress feature maps for low-memory environments or creative ML experiments.

• Memory-efficient: No new channels, shape stays fixed.
• Lossless decoding: Easily retrieve both the original value and number of feature jumps.
• Simple logic: Encode and decode with integer arithmetic.

How it works

1. Input:

   • matrix — original data (2D numpy array)
   • feature_mask — same shape, 1 where feature is detected, else 0
   • base — a value larger than any possible matrix entry

2. Encode:

   • encoded = matrix + (feature_mask * base)

3. Decode:

   • jump_count = encoded // base
   • original = encoded % base

Example

import numpy as np
from base_jump import base_jump_encode, base_jump_decode

matrix = np.array([[2, 3, 6],
                   [1, 4, 5]])
feature_mask = np.array([[1, 0, 0],
                         [0, 1, 0]])
base = 10

encoded = base_jump_encode(matrix, feature_mask, base)
print(encoded)
# Output: [[12  3  6]
#          [ 1 14  5]]

jumps, original = base_jump_decode(encoded, base)
print(jumps)     # [[1 0 0], [0 1 0]]
print(original)  # [[2 3 6], [1 4 5]]

...

## License

MIT — see [LICENSE](LICENSE).

---

**Original idea by Ante Marić**  
If you use, fork, or modify, attribution appreciated!