ARotting/neural-process-pocket

Neural Process Pocket

Neural Process Pocket is a Conditional Neural Process trained across a distribution of sine functions. Five unordered context observations are encoded into a task representation; a probabilistic decoder predicts the mean and standard deviation at arbitrary target coordinates.

Evaluation covers RMSE, Gaussian negative log likelihood, and empirical 90% interval coverage on unseen functions. A fixed-kernel RBF Gaussian Process is the non-neural few-shot control.

Verified local result

The 12,866-parameter CNP reached 1.093 RMSE, 1.242 Gaussian NLL, and 88.76% coverage for nominal 90% intervals across 500 unseen five-context-point tasks. The fixed-kernel Gaussian Process reached 1.259 RMSE, 1.328 NLL, and 73.63% coverage.

uv run python projects/neural-process-pocket/train.py
uv run pytest tests/test_neural_process_pocket.py