Online-trained two-layer neural predictor (backprop through a tanh hidden layer) driving a binary arithmetic coder, NNCP-style. The network learns as it compresses; the decoder replays the identical learning trajectory, so no weights are transmitted.
| Property | Value |
|---|---|
| Category | Compression Algorithms |
| Sub-category | Neural Network |
| Security status | 🎓 Educational Only |
| Complexity | Expert |
| Inventor | Educational Implementation |
| Year | 2019 |
| Origin | 🌐 International |
| Source | algorithms/compression/neural-compression.js |
Status: 🎓 Educational Only
No vulnerabilities are recorded for this implementation.
4 vectors ship with this algorithm and run in the test suite. Byte values are hexadecimal.
Vector 1 — Empty input
| Field | Value |
|---|---|
input |
(empty) |
expected |
00000000 |
Vector 2 — Single byte
| Field | Value |
|---|---|
input |
41 |
expected |
010000003d00 |
Vector 3 — Simple repetition
| Field | Value |
|---|---|
input |
4141 |
expected |
020000003d0c |
Vector 4 — Pattern recognition
| Field | Value |
|---|---|
input |
61626361 |
expected |
040000005beb2c4a |