Search

Knowledge-based Visual Qustion-answering (K-VQA) necessitates the use of background knowledge beyond what is depicted in the image. Current zero-shot K-VQA methods usually translate an image to a single type of textual decision context and use a text-based model to answer the question based on it, which conflicts with the fact that K-VQA questions often require the combination of multiple question-answering strategies. In light of this, we propose Rationale-based Ensemble of Answer Context Tactics (REACT) to achieve a dynamic ensemble of multiple question-answering tactics, comprising Answer Candidate Generation (ACG) and Rationale-based Strategy Fusion (RSF). In ACG, we generate three distinctive decision contexts to provide different strategies for each question, resulting in the generation of three answer candidates. RSF generates automatic and mechanistic rationales from decision contexts for each candidate, allowing the model to select the correct answer from all candidates. We conduct comprehensive experiments on the OK-VQA and A-OKVQA datasets, and our method significantly outperforms state-of-the-art LLM-based baselines on all datasets.

Large Language Models (LLMs) have demonstrated remarkable capabilities across various domains and are moving towards more specialized areas. Recent advanced proprietary models such as GPT-4 and Gemini have achieved significant advancements in biomedicine, which have also raised privacy and security challenges. The construction of specialized generalists hinges largely on high-quality datasets, enhanced by techniques like supervised fine-tuning and reinforcement learning from human or AI feedback, and direct preference optimization. However, these leading technologies (e.g., preference learning) are still significantly limited in the open source community due to the scarcity of specialized data. In this paper, we present the UltraMedical collections, which consist of high-quality manual and synthetic datasets in the biomedicine domain, featuring preference annotations across multiple advanced LLMs. By utilizing these datasets, we fine-tune a suite of specialized medical models based on Llama-3 series, demonstrating breathtaking capabilities across various medical benchmarks. Moreover, we develop powerful reward models skilled in biomedical and general reward benchmark, enhancing further online preference learning within the biomedical LLM community., Comment: Datasets and models are available at https://github.com/TsinghuaC3I/UltraMedical

The task of multimodal relation extraction has attracted significant research attention, but progress is constrained by the scarcity of available training data. One natural thought is to extend existing datasets with cross-modal generative models. In this paper, we consider a novel problem setting, where only unimodal data, either text or image, are available during training. We aim to train a multimodal classifier from synthetic data that perform well on real multimodal test data. However, training with synthetic data suffers from two obstacles: lack of data diversity and label information loss. To alleviate the issues, we propose Mutual Information-aware Multimodal Iterated Relational dAta GEneration (MI2RAGE), which applies Chained Cross-modal Generation (CCG) to promote diversity in the generated data and exploits a teacher network to select valuable training samples with high mutual information with the ground-truth labels. Comparing our method to direct training on synthetic data, we observed a significant improvement of 24.06% F1 with synthetic text and 26.42% F1 with synthetic images. Notably, our best model trained on completely synthetic images outperforms prior state-of-the-art models trained on real multimodal data by a margin of 3.76% in F1. Our codebase will be made available upon acceptance.

Generative retrieval (Wang et al., 2022; Tay et al., 2022) is a new approach for end-to-end document retrieval that directly generates document identifiers given an input query. Techniques for designing effective, high-quality document IDs remain largely unexplored. We introduce ACID, in which each document's ID is composed of abstractive keyphrases generated by a large language model, rather than an integer ID sequence as done in past work. We compare our method with the current state-of-the-art technique for ID generation, which produces IDs through hierarchical clustering of document embeddings. We also examine simpler methods to generate natural-language document IDs, including the naive approach of using the first k words of each document as its ID or words with high BM25 scores in that document. We show that using ACID improves top-10 and top-20 accuracy by 15.6% and 14.4% (relative) respectively versus the state-of-the-art baseline on the MSMARCO 100k retrieval task, and 4.4% and 4.0% respectively on the Natural Questions 100k retrieval task. Our results demonstrate the effectiveness of human-readable, natural-language IDs in generative retrieval with LMs. The code for reproducing our results and the keyword-augmented datasets will be released on formal publication.

Large-scale language model pretraining is a very successful form of self-supervised learning in natural language processing, but it is increasingly expensive to perform as the models and pretraining corpora have become larger over time. We propose NarrowBERT, a modified transformer encoder that increases the throughput for masked language model pretraining by more than $2\times$. NarrowBERT sparsifies the transformer model such that the self-attention queries and feedforward layers only operate on the masked tokens of each sentence during pretraining, rather than all of the tokens as with the usual transformer encoder. We also show that NarrowBERT increases the throughput at inference time by as much as $3.5\times$ with minimal (or no) performance degradation on sentence encoding tasks like MNLI. Finally, we examine the performance of NarrowBERT on the IMDB and Amazon reviews classification and CoNLL NER tasks and show that it is also comparable to standard BERT performance., Comment: To appear in ACL 2023 (main conference)

In video action recognition, shortcut static features can interfere with the learning of motion features, resulting in poor out-of-distribution (OOD) generalization. The video background is clearly a source of static bias, but the video foreground, such as the clothing of the actor, can also provide static bias. In this paper, we empirically verify the existence of foreground static bias by creating test videos with conflicting signals from the static and moving portions of the video. To tackle this issue, we propose a simple yet effective technique, StillMix, to learn robust action representations. Specifically, StillMix identifies bias-inducing video frames using a 2D reference network and mixes them with videos for training, serving as effective bias suppression even when we cannot explicitly extract the source of bias within each video frame or enumerate types of bias. Finally, to precisely evaluate static bias, we synthesize two new benchmarks, SCUBA for static cues in the background, and SCUFO for static cues in the foreground. With extensive experiments, we demonstrate that StillMix mitigates both types of static bias and improves video representations for downstream applications. Code is available at https://github.com/lihaoxin05/StillMix.

Interaction modeling is important for video action analysis. Recently, several works design specific structures to model interactions in videos. However, their structures are manually designed and non-adaptive, which require structures design efforts and more importantly could not model interactions adaptively. In this paper, we automate the process of structures design to learn adaptive structures for interaction modeling. We propose to search the network structures with differentiable architecture search mechanism, which learns to construct adaptive structures for different videos to facilitate adaptive interaction modeling. To this end, we first design the search space with several basic graph operations that explicitly capture different relations in videos. We experimentally demonstrate that our architecture search framework learns to construct adaptive interaction modeling structures, which provides more understanding about the relations between the structures and some interaction characteristics, and also releases the requirement of structures design efforts. Additionally, we show that the designed basic graph operations in the search space are able to model different interactions in videos. The experiments on two interaction datasets show that our method achieves competitive performance with state-of-the-arts.

Most of current Convolution Neural Network (CNN) based methods for optical flow estimation focus on learning optical flow on synthetic datasets with groundtruth, which is not practical. In this paper, we propose an unsupervised optical flow estimation framework named PCLNet. It uses pyramid Convolution LSTM (ConvLSTM) with the constraint of adjacent frame reconstruction, which allows flexibly estimating multi-frame optical flows from any video clip. Besides, by decoupling motion feature learning and optical flow representation, our method avoids complex short-cut connections used in existing frameworks while improving accuracy of optical flow estimation. Moreover, different from those methods using specialized CNN architectures for capturing motion, our framework directly learns optical flow from the features of generic CNNs and thus can be easily embedded in any CNN based frameworks for other tasks. Extensive experiments have verified that our method not only estimates optical flow effectively and accurately, but also obtains comparable performance on action recognition., Comment: IEEE International Conference on Multimedia and Expo(ICME). 2019

Egocentric interaction recognition aims to recognize the camera wearer's interactions with the interactor who faces the camera wearer in egocentric videos. In such a human-human interaction analysis problem, it is crucial to explore the relations between the camera wearer and the interactor. However, most existing works directly model the interactions as a whole and lack modeling the relations between the two interacting persons. To exploit the strong relations for egocentric interaction recognition, we introduce a dual relation modeling framework which learns to model the relations between the camera wearer and the interactor based on the individual action representations of the two persons. Specifically, we develop a novel interactive LSTM module, the key component of our framework, to explicitly model the relations between the two interacting persons based on their individual action representations, which are collaboratively learned with an interactor attention module and a global-local motion module. Experimental results on three egocentric interaction datasets show the effectiveness of our method and advantage over state-of-the-arts.