An arXiv preprint has tested a replay-free artificial immune network that learns visual classes one at a time. Both probes reported initial-class retention of 0.978 at step 9 in a sklearn-digits stream, after class 0 had been followed by classes 1 through 9. The study’s primary measure, balanced accuracy over the classes encountered, was 1.000 at step 0 and 0.939 with a logistic-regression probe and 0.902 with a one-nearest-neighbour probe at step 9. Initial retention measures preservation of the first class as later classes arrive.
A memory built as classes arrive
The model is a Deep artificial immune network, or Deep AIN. It uses structured, gradient-free immune affinity and forms depth by passing each layer’s response maps to the next: the first layer receives images, while later layers learn over earlier immune responses. Online updates evaluate affinity, stimulate responses, expand them clonally, mutate them and apply suppression. Class labels are used only by external probes that assess the representation, not in the immune updates. When the binding-profile coordinate system changes, earlier samples are re-embedded through the current repertoire.
The experiments used four grayscale benchmarks: sklearn digits, MNIST, Fashion-MNIST and KMNIST. Each stream started with class 0 and introduced classes 1 through 9 one at a time, using 80 initial samples, 60 samples at each new-class event and 35 held-out test samples per class. Results were generally averaged over three random seeds. Balanced accuracy over encountered classes was the primary measure; initial retention tracked the first class, and current-class accuracy tracked the newest class.
Structured affinity showed higher scores in one diagnostic
In a static sklearn-digits diagnostic, flat RBF—short for radial-basis function—reported balanced accuracy of 0.848. Shifted-template RBF reported 0.917, while convolution-style ZNCC, a normalized cross-correlation measure, reported 0.931. A separate MNIST feature-map row reported accuracy of 0.731. No uncertainty interval or inferential comparison was reported for this diagnostic.
Scores changed across datasets
With the same adaptive layer-wise affinity-scale rule, final balanced accuracy varied across the four streams and between the two main probes. For logistic regression and one-nearest-neighbour, respectively, the scores were 0.943 and 0.952 on sklearn digits; 0.862 and 0.857 on MNIST; 0.814 and 0.801 on Fashion-MNIST; and 0.711 and 0.853 on KMNIST. The replay-MLP comparison scores were 0.946, 0.858, 0.780 and 0.757, in the same dataset order. The reported comparison was similar on digits and MNIST, higher for the AIN on Fashion-MNIST, and dependent on probe choice on KMNIST.
Different readouts gave different scores
On KMNIST, the same learned binding-profile space produced different scores with different external readouts. Balanced accuracy was 0.612 with nearest-centroid, 0.711 with logistic regression, 0.818 with PCA plus RBF-SVM, and 0.853 with one-nearest-neighbour. PCA is principal-component analysis, while RBF-SVM is a support-vector-machine readout using a radial-basis-function setting. These probes were external readouts used to assess the representation, not components of the immune update process.
Scores varied with layer scale
Another comparison varied the affinity scale used by each layer; σ denotes that scale. In a two-layer model, logistic-probe balanced accuracy was 0.953 with fixed σ2 = 1.0, 0.978 with fixed σ2 = 2.0, and 0.968 with an adaptive nearest-neighbour scale multiplied by 6. Listed three-layer settings scored 0.924 with fixed σ2 = 2.0 and σ3 = 2.0, and 0.927 with adaptive scales multiplied by 6 and 6. The reported results did not show that adding depth automatically raised the score.
A narrow test, for now
The supplied evidence is limited to four grayscale benchmarks—sklearn digits, MNIST, Fashion-MNIST and KMNIST. The paper presents the work as a proof of concept rather than a general solution to visual continual learning, and the experiments do not establish performance on natural RGB images or more complex visual streams.
Results were generally averaged over three random seeds, but no uncertainty intervals or inferential tests were reported. The front matter identifies the work as arXiv version v1 dated 20 Aug 2026.
Paper data and sources
Original title: Structured Affinity for Unsupervised Visual Class-Incremental Memory in Deep Artificial Immune Networks
Authors: Siphesihle Sithungu
Journal/Repository: arXiv
Status: Preprint, not yet peer-reviewed
First online: 2026-08-20
DOI: Not available
Original paper · Full text