Yong Cao

Haoyu He, Katrin Renz, Yong Cao et al.

Diffusion language models, as a promising alternative to traditional autoregressive (AR) models, enable faster generation and richer conditioning on bidirectional context. However, they suffer from a key discrepancy between training and inference: during inference, MDLMs progressively reveal the structure of the generated sequence by producing fewer and fewer masked tokens, whereas this structure is ignored in training as tokens are masked at random. Although this discrepancy between training and inference can lead to suboptimal performance, it has been largely overlooked by previous works, leaving closing this gap between the two stages an open problem. To address this, we frame the problem of learning effective denoising trajectories as a sequential decision-making problem and use the resulting framework to apply reinforcement learning. We propose a novel Masked Diffusion Policy Optimization (MDPO) to exploit the Markov property diffusion possesses and explicitly train the model under the same progressive refining schedule used at inference. MDPO matches the performance of the previous state-of-the-art (SOTA) method with 60x fewer gradient updates, while achieving average improvements of 9.6% on MATH500 and 54.2% on Countdown over SOTA when trained within the same number of weight updates. Additionally, we improve the remasking strategy of MDLMs as a plug-in inference replacement to overcome the limitation that the model cannot refine tokens flexibly. This training-free method, termed Running Confidence Remasking (RCR), consistently enhances performance and provides further improvements when used with MDPO. Our findings establish great potential for investigating the discrepancy between pre-training and inference of MDLMs. Code: https://github.com/autonomousvision/mdpo. Project Page: https://cli212.github.io/MDPO/.

Zhenshen Qu, Jingda Du, Yong Cao et al.

Recently, CNN object detectors have achieved high accuracy on remote sensing images but require huge labor and time costs on annotation. In this paper, we propose a new uncertainty-based active learning which can select images with more information for annotation and detector can still reach high performance with a fraction of the training images. Our method not only analyzes objects' classification uncertainty to find least confident objects but also considers their regression uncertainty to declare outliers. Besides, we bring out two extra weights to overcome two difficulties in remote sensing datasets, class-imbalance and difference in images' objects amount. We experiment our active learning algorithm on DOTA dataset with CenterNet as object detector. We achieve same-level performance as full supervision with only half images. We even override full supervision with 55% images and augmented weights on least confident images.

Pengwei Wang, Liangchen Wei, Yong Cao et al.

End-to-end Spoken Language Understanding (SLU) is proposed to infer the semantic meaning directly from audio features without intermediate text representation. Although the acoustic model component of an end-to-end SLU system can be pre-trained with Automatic Speech Recognition (ASR) targets, the SLU component can only learn semantic features from limited task-specific training data. In this paper, for the first time we propose to do large-scale unsupervised pre-training for the SLU component of an end-to-end SLU system, so that the SLU component may preserve semantic features from massive unlabeled audio data. As the output of the acoustic model component, i.e. phoneme posterior sequences, has much different characteristic from text sequences, we propose a novel pre-training model called BERT-PLM, which stands for Bidirectional Encoder Representations from Transformers through Permutation Language Modeling. BERT-PLM trains the SLU component on unlabeled data through a regression objective equivalent to the partial permutation language modeling objective, while leverages full bi-directional context information with BERT networks. The experiment results show that our approach out-perform the state-of-the-art end-to-end systems with over 12.5% error reduction.

Yihong Chen, Bei Chen, Xuguang Duan et al.

Almost all the knowledge empowered applications rely upon accurate knowledge, which has to be either collected manually with high cost, or extracted automatically with unignorable errors. In this paper, we study 20 Questions, an online interactive game where each question-response pair corresponds to a fact of the target entity, to acquire highly accurate knowledge effectively with nearly zero labor cost. Knowledge acquisition via 20 Questions predominantly presents two challenges to the intelligent agent playing games with human players. The first one is to seek enough information and identify the target entity with as few questions as possible, while the second one is to leverage the remaining questioning opportunities to acquire valuable knowledge effectively, both of which count on good questioning strategies. To address these challenges, we propose the Learning-to-Ask (LA) framework, within which the agent learns smart questioning strategies for information seeking and knowledge acquisition by means of deep reinforcement learning and generalized matrix factorization respectively. In addition, a Bayesian approach to represent knowledge is adopted to ensure robustness to noisy user responses. Simulating experiments on real data show that LA is able to equip the agent with effective questioning strategies, which result in high winning rates and rapid knowledge acquisition. Moreover, the questioning strategies for information seeking and knowledge acquisition boost the performance of each other, allowing the agent to start with a relatively small knowledge set and quickly improve its knowledge base in the absence of constant human supervision.