Zhuang, K. Li, H. Cheng, X. Guo, F. Huang, R. Ji, and X. Sun, "Disco: Remedy213 self-supervised learning on lightweight models with distilled contrastive learning,"arXiv preprint214 arXiv:2104.09124,2021.215

For vision transformers, we train linear probes on representations from individual tokens or on the representation averaged over all tokens, at the output of different transformer layers (each layer meaning a full transformer block including self-attention and MLP). Moreover, ResNets differ from ViTs in that the number of channels changes throughout the model, with fewer channels in the earlier layers. Wetrain alinear probe on each individual token and plot the average accuracy over the test set, in percent. Here we plot the results for each token a subset of layers in 3models: ViT-B/32 trained with aclassification token (CLS) or global average pooling (GAP), as well as a ResNet50. There are two main observations tobemade.

Deep networks are heavily over-parameterized, yet their learned representations often admit low-rank structure. We introduce a framework for estimating a model's intrinsic dimensionality by treating learned representations as projections onto a low-rank subspace of the model's full capacity. Our approach: train a full-rank teacher, factorize its weights at multiple ranks, and train each factorized student via distillation to measure performance as a function of rank. We define effective rank as a region, not a point: the smallest contiguous set of ranks for which the student reaches 85-95% of teacher accuracy. To stabilize estimates, we fit accuracy vs. rank with a monotone PCHIP interpolant and identify crossings of the normalized curve. We also define the effective knee as the rank maximizing perpendicular distance between the smoothed accuracy curve and its endpoint secant; an intrinsic indicator of where marginal gains concentrate. On ViT-B/32 fine-tuned on CIFAR-100 (one seed, due to compute constraints), factorizing linear blocks and training with distillation yields an effective-rank region of approximately [16, 34] and an effective knee at r* ~ 31. At rank 32, the student attains 69.46% top-1 accuracy vs. 73.35% for the teacher (~94.7% of baseline) while achieving substantial parameter compression. We provide a framework to estimate effective-rank regions and knees across architectures and datasets, offering a practical tool for characterizing the intrinsic dimensionality of deep models.

Vision-language models like CLIP have demonstrated remarkable zero-shot capabilities in classification and retrieval. However, these models often struggle with compositional reasoning - the ability to understand the relationships between concepts. A recent benchmark, SugarCrepe++, reveals that previous works on improving compositionality have mainly improved lexical sensitivity but neglected semantic understanding. In addition, downstream retrieval performance often deteriorates, although one would expect that improving compositionality should enhance retrieval. In this work, we introduce CLIC (Compositionally-aware Learning in CLIP), a fine-tuning method based on a novel training technique combining multiple images and their associated captions. CLIC improves compositionality across architectures as well as differently pre-trained CLIP models, both in terms of lexical and semantic understanding, and achieves consistent gains in retrieval performance. This even applies to the recent CLIPS, which achieves SOTA retrieval performance. Nevertheless, the short fine-tuning with CLIC leads to an improvement in retrieval and to the best compositional CLIP model on SugarCrepe++. All our models and code are available at https://clic-compositional-clip.github.io

We raise a challenging research question in the following: Is there a benchmark that reflects the exact reliance on spurious features of CLIP?