Evaluation of Two Adaptation Designs for Small-Data Retinal Classification

Abstract A 2 × 2 experiment was conducted on 1,489 retinal fundus images with a frozen ResNet-50 to test two recent designs, the DualKernel Adapter (DKA) and Manifold-Constrained Hyper-Connections (mHC). Every trained model beat a linear probe by 0.013 to 0.031 macro AUC. mHC improved on a width-64 residual neck by +0.0075 macro AUC, and in an exploratory comparison planned in advance, a plain residual neck widened to the same parameter count achieved about the same improvement on average, so a simpler design delivered about the same benefit. DKA scored 0.0104 macro AUC below a standard adapter of nearly the same size. The standard adapter also made better use of additional training data.

This is a preprint: it has not been peer reviewed yet, and its findings may change.

Published in Research Square (preprint). The abstract is shown as published and remains the work of its authors; read the full article at the source.

doi.org/10.21203/rs.3.rs-11200257/v1