Structured Affinity for Unsupervised Visual Class-Incremental Memory in Deep Artificial Immune Networks
The paper’s core claim is that visual immune memory works better when it keeps spatial response maps instead of collapsing images into scalar affinities.
Siphesihle Sithungu tests Deep Artificial Immune Networks as replay-free, class-incremental learners using structured templates, ZNCC filters, and feature-map binding profiles. The reported experiments cover sklearn digits, MNIST, Fashion-MNIST, and KMNIST, with scalar binding-profile variants trailing feature-map versions. On sklearn digits, calibrated two-layer feature-map Deep AIN reaches 0.978 balanced accuracy, while Fashion-MNIST reaches 0.814 and KMNIST 0.853 under the same calibration rule. The probes are described as external validation tools, not part of the AIN itself. ArXiv · AI/CL/LG's note
Siphesihle Sithungu tests Deep Artificial Immune Networks as replay-free, class-incremental learners using structured templates, ZNCC filters, and feature-map binding profiles. The reported experiments cover sklearn digits, MNIST, Fashion-MNIST, and KMNIST, with scalar binding-profile variants trailing feature-map versions. On sklearn digits, calibrated two-layer feature-map Deep AIN reaches 0.978 balanced accuracy, while Fashion-MNIST reaches 0.814 and KMNIST 0.853 under the same calibration rule. The probes are described as external validation tools, not part of the AIN itself. ArXiv · AI/CL/LG's note
score 4