No old images, still learning new domains: Lamarck inherits each expert from the last and leads CDDB-Hard by 5.2 points
Synopsis
In replay-free domain-incremental learning, test images carry no domain label, so a router picks one per-domain expert for each image. Lamarck keeps a training-free Mahalanobis router and spends its effort on the experts: an adapter after each deep ViT block corrects every image token, and each expert, adapters included, is initialized from its predecessor. On CDDB-Hard the adapters add 4.36 ACC points over prompt-only experts, and the full learner beats the best of twelve frozen-backbone baselines by 5.18 points while posting the highest mean accuracy across three benchmarks.
Interpretation
Instead of training the router, the per-domain learner improves the experts themselves: Lamarck adds a residual adapter after each deep block of the frozen ViT that applies a learned correction to every image token. Earlier per-domain learners put subsequent work into training the router, and the nine compared prompt learners place prompts at the input or in blocks 1 to 5 and train nothing in the deep blocks. Under the same Mahalanobis router, the adapters raise Lamarck 4.36 ACC points over its own prompt-only experts on CDDB-Hard, 2.14 on DomainNet and 0.33 on OfficeHome; on CDDB-Hard, deep-block prompts with as many parameters recover only 3.17 of those 4.36 points.
Each new expert, prompts, adapters and readout together, is initialized from its predecessor, so an image misrouted to a later expert loses less accuracy. Independently trained experts share no starting point, so a misrouted image absorbs the full accuracy gap of a later expert on an earlier domain. On CDDB-Hard, inheritance lifts later-expert accuracy on earlier domains to 82.12%, against 70.06% with reset adapters and 59.45% for independent experts (chance is 50%); the later-expert cost falls from 27.81 with reset adapters to 15.75.
Under one frozen ViT-B/16 protocol, Lamarck reaches the highest final accuracy on CDDB-Hard and the highest mean ACC over the three benchmarks. Among the twelve replay-free baselines, Hash was previously best on CDDB-Hard and HiDe-Prompt best in mean ACC. Over five seeds, Lamarck scores 92.94 on CDDB-Hard, 5.18 points above Hash, and 83.62 in mean ACC against 79.67 for HiDe-Prompt; it is second on DomainNet and OfficeHome at 69.46 and 88.47.
Perspective
The results rest on a frozen ViT-B/16, the DomainNet, OfficeHome and CDDB-Hard benchmarks, fixed domain orders and five seeds, with 15 epochs per domain at batch size 128; the second backbone, DINOv2 ViT-B/14, was tested only on OfficeHome and CDDB-Hard, and the reversed domain order only on CDDB-Hard. Inference runs the ViT twice per image, once to route and once to classify, at 71.2 GFLOPs per image, comparable to other two-pass learners.
On DomainNet, SOYO's experts are more accurate than Lamarck's on every domain and SOYO's ACC is higher; under the Mahalanobis router, inheritance changes ACC against reset adapters by −0.41 on DomainNet, +0.15 on OfficeHome and +0.21 on CDDB-Hard, and its gain becomes clear only under weaker routers or injected routing errors. Without adapters, inherited experts trail independent ones in ACC on CDDB-Hard. Storage grows by one expert per domain, totaling 9.84M parameters.
