Haoxin Li, Yuan Liu, Hanwang Zhang, and Boyang Li. Mitigating and Evaluating Static Bias of Action Representations in the Background and the Foreground. International Conference on Computer Vision (ICCV) (Oral Presentation). 2023.

TL;DR: Video classifiers often rely on static features to make biased predictions. We show that static bias is caused by not only the background, but also the foreground, such as human actor's outfit or instrument. We propose a theory-inspired method to mitigate bias without having to locate the source.

Wenliang Dai, Junnan Li, Dongxu Li, Anthony Meng Huat Tiong, Junqi Zhao, Weisheng Wang, Boyang Li, Pascale Fung, and Steven Hoi. InstructBLIP: Towards General-purpose Vision-Language Models with Instruction Tuning. NeurIPS. 2023.

TL;DR: An instruction-tuned vision-language model that achieves state-of-the-art performance on several benchmarks.

Yidan Sun, Qin Chao, Yangfeng Ji, and Boyang Li. Synopses of Movie Narratives: a Video-Language Dataset for Story Understanding. ArXiv Preprint 2203.05711. 2022.

TL;DR: A large, plot-level, multimodal dataset for story understanding. The storytelling techniques of the paper create unique challenges for the current generation of multimodal networks.

Anthony Meng Huat Tiong, Junnan Li, Boyang Li, Silvio Savarese, and Steven C.H. Hoi. Plug-and-Play VQA: Zero-shot VQA by Conjoining Large Pretrained Models with Zero Training. Findings of the Conference on Empirical Methods in Natural Language Processing (Findings of EMNLP). 2022.

TL;DR: An unexpected modular approach to visual question answering. We translate relevant portions of an image to text and rely on text-based reasoning entirely. On VQAv2, the results are better than Deepmind's Flamingo by 8.5%. This is in contrast to conventional wisdom that end-to-end learning is necessary for good performance.

Xu Guo, Boyang Li, and Han Yu. Improving the Sample Efficiency of Prompt Tuning with Domain Adaptation. Findings of the Conference on Empirical Methods in Natural Language Processing (Findings of EMNLP). 2022.

TL;DR: Prompt tuning is great for large model deployment but requires a lot of training data. To our knowledge, we are the first in using unlabeled data in the target domain to improve prompt tuning.

Jun Chen, Han Guo, Kai Yi, Boyang Li, and Mohamed Elhoseiny. VisualGPT: Data-efficient Adaptation of Pretrained Language Models for Image Captioning. The IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR). 2022.

TL;DR: We developed a data-efficient method to adapt large-scale pretrained language models for image captioning and achieved SOTA results on X-ray image captioning.

Yinan Zhang, Boyang Li, Yong Liu, Hao Wang, Chunyan Miao. Initialization Matters: Regularizing Manifold-informed Initialization for Neural Recommendation Systems. ACM SIGKDD Conference on Knowledge Discovery and Data Mining (KDD). 2021.

TL;DR: If neural recommenders performed poorly (e.g., worse than well-tuned k-nearest-neighbors), it is probably because they did not use this data-dependent, manifold-informed initialization.

Xu Guo, Boyang Li, Han Yu, and Chunyan Miao. Latent-Optimized Adversarial Neural Transfer for Sarcasm Detection. The Annual Conference of the North American Chapter of the Association for Computational Linguistics (NAACL-HLT). 2021.

TL;DR: Sarcasm detection is an ideal problem for transfer learning. We identify the competition between losses in adversarial transfer learning and propose a modified optimization technique to solve the problem, which achieves the SOTA result on the iSarcasm dataset.

Chang Liu, Han Yu, Boyang Li, Zhiqi Shen, Zhanning Gao, Peiran Ren, Xuansong Xie, Lizhen Cui, and Chunyan Miao. Noise-resistant Deep Metric Learning with Ranking-based Instance Selection. The IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR). 2021.

TL;DR: We introduce a simple, efficient, and (we believe) the first technique for deep metric learning under noisy training data; the method outperforms 12 baseline methods under both synthetic and natural noise.

Yuanyuan Chen, Boyang Li, Han Yu, Pengcheng Wu, and Chunyan Miao. HyDRA: Hypergradient Data Relevance Analysis for Interpreting Deep Neural Networks. The AAAI Conference on Artificial Intelligence (AAAI). 2021.

TL;DR: We provide an approximate hypergradient method for estimating how training data contribute to individual network predictions and a theoretical bound on the approximation error.

Adam Noack, Isaac Ahern, Dejing Dou, and Boyang Li. An Empirical Study on the Relation between Network Interpretability and Adversarial Robustness Springer Nature Computer Science. 2020.

TL;DR: Does the interpretability of neural networks imply robustness against adversarial attack? We provide some positive empirical evidence.

Hannah Kim, Denys Katerenchuk, Daniel Billet, Jun Huan, Haesun Park, and Boyang Li. Understanding Actors and Evaluating Personae with Gaussian Embeddings. The AAAI Conference on Artificial Intelligence (AAAI). 2019.

TL;DR: We computationally model movie casting decisions and actors' versatility.

Pelin Dogan, Boyang Li, Leonid Sigal, Markus Gross. A Neural Multi-sequence Alignment TeCHnique (NeuMATCH). The Conference on Computer Vision and Pattern Recognition (CVPR). 2018.

TL;DR: We propose the first end-to-end optimizable network for aligning video and text sequences.

Ng Annalyn, Maarten Bos, Leonid Sigal, Boyang Li. Predicting Personality from Book Preferences with User-Generated Content Labels. IEEE Transaction on Affective Computing. 2018.

TL;DR: We can infer your personality from the books you read.