Neural Predictor

The NeuralPredictor codec is an opt-in NNCP-style (Neural Network Data Compression) entropy coder that uses a neural network to model byte probabilities.

Theory

Unlike fixed-context or linear mixing models, this codec uses a deep neural network to predict the probability distribution of the next byte based on the previous bytes. The predicted distributions are then fed into a standard entropy coder (like arithmetic coding) to emit the compressed bitstream.

Neural predictors can capture complex, high-order correlations that traditional algorithms miss, yielding unprecedented compression ratios on complex data. However, this comes at the cost of significantly slower encoding and decoding speeds.

Usage

Because of its high computational cost, this codec is not included in the default trial-encode dispatch menu. It must be explicitly selected via the codec_menu. It requires a minimum plane size of 4 KiB to amortize its per-bit overhead.

References

• J. Schmidhuber & S. Heil, "Predictive Coding with Neural Nets: Application to Text Compression" (1996). Link
• Fabian Giesen, "Neural Network Data Compression Program (NNCP)".