From Predictive Coding to EBPM: A Novel DIME Integrative Model for Recognition and Cognition

We introduce Experience-Based Pattern Matching (EBPM) as a biologically grounded alternative to Predictive Coding (PC) for rapid recognition and action. EBPM posits that familiar inputs trigger direct reactivation of multimodal engrams (sensory–motor–interoceptive–cognitive), avoiding continuous generative simulation and thereby reducing latency and compute. We provide a method-by-method comparison between EBPM and adjacent frameworks—attractor/Hopfield models, Complementary Learning Systems, hippocampal indexing (pattern completion/separation), HTM/sequence memory, Sparse Distributed Memory, memory-augmented neural networks (NTM/DNC), episodic control, metric-learning/prototypical networks, and PC/Active Inference—explicitly identifying (i) shared mechanisms, (ii) additions brought by each prior theory, and (iii) what EBPM adds beyond them. We further present DIME (Dynamic Integrative Matching and Encoding), a hybrid architecture that exploits EBPM’s one-shot recall on familiar inputs and falls back to PC under uncertainty or novelty via a lightweight controller. Across ANN benchmarks (MNIST, Fashion-MNIST, CIFAR-10) and a virtual robotics obstacle-course, EBPM consistently yields the lowest inference latency on familiar inputs, PC remains more robust to noise/novelty, and DIME adapts between them without sacrificing task performance. Finally, we formalize Algorithm 1 (EBPM) and map each step to prior art (sparse coding and assemblies; lateral inhibition/WTA; attractor-like selection; salience/LC-NE; Hebbian/STDP; hippocampal indexing), clarifying how EBPM integrates classic mechanisms into a unified, multimodal engram framework. The results motivate EBPM/DIME as neuro-plausible, energy-aware blueprints for hybrid AI and real-time robotic control.