Ruisong Zhou

Chengyue Wu, Teng Wang, Yixiao Ge et al. · tencent-ai

Foundation models have achieved great advances in multi-task learning with a unified interface of unimodal and multimodal tasks. However, the potential of such multi-task learners has not been exploited during transfer learning. In this work, we present a universal parameter-efficient transfer learning method, termed Predict-Interpolate Tuning ($π$-Tuning), for vision, language, and vision-language tasks. It aggregates the parameters of lightweight task-specific experts learned from similar tasks to aid the target downstream task. The task similarities are predicted in a unified modality-independent space, yielding a scalable graph to demonstrate task relationships. $π$-Tuning has several appealing benefits. First, it flexibly explores both intra- and inter-modal transferability between similar tasks to improve the accuracy and robustness of transfer learning, especially in data-scarce scenarios. Second, it offers a systematical solution for transfer learning with multi-task prediction-and-then-interpolation, compatible with diverse types of parameter-efficient experts, such as prompt and adapter. Third, an extensive study of task-level mutual benefits on 14 unimodal and 6 multimodal datasets shows that $π$-Tuning surpasses fine-tuning and other parameter-efficient transfer learning methods both in full-shot and low-shot regimes. The task graph also enables an in-depth interpretable analysis of task transferability across modalities. The code will be available at https://github.com/TencentARC/pi-Tuning.

Zhiyuan Ren, Yaohai Zhou, Yizhe Chen et al.

In this paper, we propose an efficient human pose estimation network -- SFM (slender fusion model) by fusing multi-level features and adding lightweight attention blocks -- HSA (High-Level Spatial Attention). Many existing methods on efficient network have already taken feature fusion into consideration, which largely boosts the performance. However, its performance is far inferior to large network such as ResNet and HRNet due to its limited fusion operation in the network. Specifically, we expand the number of fusion operation by building bridges between two pyramid frameworks without adding layers. Meanwhile, to capture long-range dependency, we propose a lightweight attention block -- HSA, which computes second-order attention map. In summary, SFM maximizes the number of feature fusion in a limited number of layers. HSA learns high precise spatial information by computing the attention of spatial attention map. With the help of SFM and HSA, our network is able to generate multi-level feature and extract precise global spatial information with little computing resource. Thus, our method achieve comparable or even better accuracy with less parameters and computational cost. Our SFM achieve 89.0 in PCKh@0.5, 42.0 in PCKh@0.1 on MPII validation set and 71.7 in AP, 90.7 in AP@0.5 on COCO validation with only 1.7G FLOPs and 1.5M parameters. The source code will be public soon.