DREAM: Dynamic Recall and Elastic Adaptive Memory on Machine Learning

DREAM is a novel neural architecture that enables real-time, in-situ parameter adaptation during inference. By integrating Active Inference and Spike-Timing-Dependent Plasticity (STDP), it overcomes the static constraints of Transformers and LSTMs. Core Mechanisms: Slow-Fast Weights: Resolves the stability-plasticity dilemma by separating long-term invariants from elastic contextual updates. Surprise Gate: A homeostatic regulator that triggers plasticity only upon detecting structural novelty, preventing overfitting to noise. Low-Rank Updates: Ensures computational efficiency for Edge AI by restricting adaptations to a compact feature subspace (r=16). Liquid Time-Constants (LTC): Dynamically adjusts information integration rates to instantly discard outdated context during anomalies. Key Results: On non-stationary audio tasks (LJ Speech), a 82K parameter DREAM model achieved a 99.9% error reduction, outperforming stasis-bound models (LSTM 893K, Transformer 551K) through millisecond-scale self-calibration