A common strategy for Parameter-Efficient Fine-Tuning (PEFT) of pre-trained Vision Transformers (ViTs) involves adapting the model to downstream tasks by learning a low-rank adaptation matrix. This matrix is decomposed into a product of down-projection and up-projection matrices, with the bottleneck dimensionality being crucial for reducing the number of learnable parameters, as exemplified by prevalent methods like LoRA and Adapter. However, these low-rank strategies typically employ a fixed bottleneck dimensionality, which limits their flexibility in handling layer-wise variations. To address this limitation, we propose a novel PEFT approach inspired by Singular Value Decomposition (SVD) for representing the adaptation matrix. SVD decomposes a matrix into the product of a left unitary matrix, a diagonal matrix of scaling values, and a right unitary matrix.

Visual task adaptation has been demonstrated to be effective in adapting pre-trained Vision Transformers (ViTs) to general downstream visual tasks using specialized learnable layers or tokens. However, there is yet a large-scale benchmark to fully explore the effect of visual task adaptation on the realistic and important medical domain, particularly across diverse medical visual modalities, such as color images, X-ray, and CT. To close this gap, we present Med-VTAB, a large-scale Medical Visual Task Adaptation Benchmark consisting of 1.68 million medical images for diverse organs, modalities, and adaptation approaches. Based on Med-VTAB, we explore the scaling law of medical prompt tuning concerning tunable parameters and the generalizability of medical visual adaptation using non-medical/medical pre-train weights. Besides, we study the impact of patient ID out-of-distribution on medical visual adaptation, which is a real and challenging scenario. Furthermore, results from Med-VTAB indicate that a single pre-trained model falls short in medical task adaptation. Therefore, we introduce GMoE-Adapter, a novel method that combines medical and general pre-training weights through a gated mixture-of-experts adapter, achieving state-of-the-art results in medical visual task adaptation.

Pre-trained large transformer models have achieved remarkable performance in the fields of natural language processing and computer vision. Since the magnitude of available labeled electroencephalogram (EEG) data is much lower than that of text and image data, it is difficult for transformer models pre-trained from EEG to be developed as large as GPT-4 100T to fully unleash the potential of this architecture. In this paper, we show that transformers pre-trained from images as well as text can be directly fine-tuned for EEG-based prediction tasks. We design AdaCE, plug-and-play Adapters for Converting EEG data into image as well as text forms, to fine-tune pre-trained vision and language transformers. The proposed AdaCE module is highly effective for fine-tuning pre-trained transformers while achieving state-of-the-art performance on diverse EEG-based prediction tasks. For example, AdaCE on the pre-trained Swin-Transformer achieves 99.6%, an absolute improvement of 9.2%, on the EEG-decoding task of human activity recognition (UCI HAR). Furthermore, we empirically show that applying the proposed AdaCE to fine-tune larger pre-trained models can achieve better performance on EEG-based predicting tasks, indicating the potential of our adapters for even larger transformers. The plug-and-play AdaCE module can be applied to fine-tuning most of the popular pre-trained transformers on many other time-series data with multiple channels, not limited to EEG data and the models we use. Our code will be available at https://github.com/wangbxj1234/AdaCE.