Contextual Activation Functions for Temporal Systems: A Neuromodulated Approach

This repository contains the full implementation and experiment scripts associated with the paper “Contextual Activation Functions for Temporal Systems: A Neuromodulated Approach.”The work introduces Contextual Activation Functions, a class of differentiable, neuromodulated mechanisms that adapt each neuron’s nonlinearity according to explicit system-state signals and short-term history summaries.The implementation includes the System-Aware Temporal model evaluated on both synthetic benchmarks and the UCI Household Power Consumption dataset. Comparative experiments with LSTM and GRU baselines demonstrate consistent improvements in prediction accuracy (≈ 5 % reduction in test MSE) and interpretable gating dynamics reflecting learned context-dependent modulation. The codebase provides: PyTorch implementations of LSTM, GRU, and Contextual (System-Aware Temporal) models Training, evaluation, and visualization scripts Reproducible preprocessing pipelines for UCI datasets Example notebooks for gate analysis and visualization The paper and this repository together form a complete research artifact supporting reproducibility and further exploration of neuromodulated activation mechanisms.