Few-shot CLIP adaptation is often treated as a parameterization problem. Full fine-tuning is assumed to overfit, so parameter-efficient fine-tuning (PEFT) methods are preferred. In this paper, we show that this conclusion conflates parameterization with optimization. Across 11 datasets, 3 CLIP backbones, and 5 shot levels, dense full fine-tuning with plain SGD outperforms representative PEFT methods under the reported evaluation protocol, whereas Adam/AdamW collapses. We trace this collapse to unreliable support-set preconditioning. The few-shot second moment systematically underestimates the corresponding full-data statistic, already at initialization. Adam's inverse-square-root normalization converts this error into an oversized adaptive update. Matching Adam's first-step update budget to SGD resolves most of the collapse, whereas numerator-side alignment does not explain the failure. Shampoo and SOAP also fail, so richer preconditioning alone does not resolve the problem. Theory formalizes how few-shot estimation error is amplified into unsafe updates. SGD succeeds because it avoids unreliable preconditioners and stays close to pretrained CLIP. Interpolation probes show low-loss connections along tested directions, and curvature probes show favorable directional curvature. These results identify optimizer-induced preconditioning, not dense full fine-tuning itself, as the main source of failure in few-shot CLIP adaptation.
Transductive few-shot learning has triggered an abundant literature focusing on vision-only models, but is still at a nascent stage within the recent context of foundational vision-language models (VLMs). Only a few recent methods addressed the problem, pointing to the potential of tranduction in VLMs and to the need for VLM-tailored methods. Building on this momentum, we leverage information-theoretic concepts and recent progress in parameter-efficient fine-tuning (PEFT), developing a highly competitive transductive few-shot CLIP method. Specifically, we introduce a novel Language-aware Information MaximizatiOn (LIMO) loss integrating three complementary terms: (i) the mutual information between the vision inputs and the textual class descriptions; (ii) a Kullback-Leibler (KL) divergence penalizing deviation of the network's probabilistic outputs from the text-driven zero-shot predictions; and (iii) a standard cross-entropy loss based on the labeled shots. Furthermore, we challenge the commonly followed fine-tuning practices in the context of transductive few-shot learning, and explore PEFT strategies, completely overlooked in this context. Surprisingly, we observe substantial boosts in performances, which points to the potential of adapting a subset of the model's parameters in the transductive few-shot setting. We report comprehensive evaluations, which show that LIMO outperforms the very recent transductive few-shot CLIP methods by a large margin and yields significant gains over the best-performing inductive methods. Our code is publicly available at:https://github.com/ghassenbaklouti/LIMO
Nairouz Mrabah, Ismail Ben Ayed, Mahdi S. Hosseini, Eric Granger
原文摘要
The limited supervision available for the few-shot adaptation of vision-language models makes gradient estimation particularly noisy. We find that the gradient magnitudes are severely miscalibrated, while the signs of a small top-magnitude subset remain informative. This motivates SignSparseSGD and DynSignSparseSGD, two state-free optimizers that use gradient magnitudes only to select where to update and fixed-magnitude signs to decide how to update. The static variant fixes the support after the first minibatch. The dynamic variant refreshes it at every step. Across 11 datasets, three CLIP backbones, and five shot levels, the proposed methods are competitive with or outperform parameter-efficient and optimization-based baselines on average, while exhibiting markedly lower sensitivity to long training schedules. They introduce no task-specific modules, maintain no first- or second-moment optimizer states, and use one shared configuration across datasets, backbones, and shot levels. Our code is included in the supplementary materials and will be released upon publication.
Sahar Dastani, Fereshteh Shakeri, Samuel Barbeau, Kunal Mahatha, Ismail Ben Ayed, Herve Lombaert, Christian Desrosiers
原文摘要
Self-attention has become the standard for modern sequence modeling, enabling Transformers to model long-range dependencies across a wide range of modalities. Despite these advances, its quadratic cost with respect to sequence length remains a significant computational bottleneck. Recent work addresses this limitation by introducing linear variants of self-attention, namely with kernel approximations of the softmax operator. However, their first-order aggregation restricts expressivity and struggles to capture higher-order structure. We propose Linear Stochastic Attention (LiSA), a linear attention plug-in module that enables higher-order token interactions via a stochastic diffusion process based on truncated random walks. LiSA incorporates adaptive locality, favoring semantically similar neighbors to preserve fine-grained structures, while maintaining stable and efficient training. We conduct extensive experiments on classification, semantic segmentation, and object detection tasks and show improved results over state-of-the-art methods.