Year: 2026
Type: preprint
Abstract: Many time-series anomalies of practical consequence are not sharp spikes but weak deviations sustained across many consecutive samples. Standard deep-learning encoders, such as convolutional, recurrent and Transformer architectures, were not designed to distinguish whether faint evidence is transient or persistent. We introduce the Immunological Synapse Network (ISN), a differentiable architecture that maps four T-cell recognition mechanisms onto anomaly detection: a learned receptor scores local signal tokens by binding affinity; a kinetic proofreading cascade suppresses short-lived matches by requiring evidence to survive multiple sequential states; ITAM-like gates impose a learned threshold on each receptor; and a synapse-inspired spatial lattice concentrates surviving activations before classification. On a controlled synthetic benchmark whose label is recoverable only through temporal persistence, ISN is the only model that does not collapse to chance when its proofreading cascade is disabled, and it outperforms all baselines in full-model comparison. These results establish a clear operating boundary: immune-inspired persistence filtering earns its computational cost when anomalies are defined by temporal continuity, not by instantaneous magnitude.
Source: Research Square
Authors: Sherif Tawfik, Hung Le, Svetha Venkatesh
Institution: Deakin University
Language: English
Subfield: Biomedical Engineering
Field: Engineering
Domain: Physical Sciences
Open Access status: green
OpenAlex ID: W7204436054