Binghong Chen*, Tianzhe Wang*, Chengtao Li, Hanjun Dai, Le Song
This paper develops a novel algorithm for optimizing molecular properties via an Expectation Maximization (EM) like explainable evolutionary process. We show that our evolution-by-explanation algorithm is 79% better than the best baseline in terms of a generic metric combining aspects such as success rate, novelty, and diversity. Human expert evaluation on optimized molecules shows that 60% of top molecules obtained from our methods are deemed successful.
Zetian Jiang*, Tianzhe Wang*, Junchi Yan
We address the problem of multiple graph matching (MGM) in terms of both offline batch mode and online setting. We explore the concept of cycle-consistency over pairwise matchings and formulate the problem as finding optimal composition path on the supergraph, whose nodes refer to graphs and edge weights denote score function regarding consistency and affinity.
Tianzhe Wang, Kuan Wang, Han Cai, Ji Lin, Zhijian Liu, Song Han
We present a joint design methodology for efficient deep learning deployment. Unlike previous methods that separately optimize the neural network architecture, pruning policy, and quantization policy, we optimize them in a joint manner.
Han Cai, Chuang Gan, Tianzhe Wang, Zhekai Zhang, Song Han
We address the challenging problem of efficient deep learning model deployment, where the goal is to design neural network architectures that can fit different hardware platform constraints. Our key idea is to decouple model training from architecture search to save the cost. To this end, we propose to train a once-for-all network (OFA) that supports diverse architectural settings (depth, width, kernel size, and resolution).
Han Cai, Tianzhe Wang, Zhanghao Wu, Kuan Wang, Ji Lin, Song Han
We employ Proxyless Neural Architecture Search (ProxylessNAS) to auto design compact and specialized neural network architectures for the target hardware platform. ProxylessNAS makes latency differentiable, so we can optimize not only accuracy but also latency by gradient descent. Such direct optimization saves the search cost by 200 compared to conventional neural architecture search methods. Our work is followed by quantization-aware fine-tuning to further boost efficiency.
Tianzhe Wang*, Han Cai*, Ji Lin*, Yujun Lin*, Zhijian Liu*, Kuan Wang*, Ligeng Zhu*, Song Han
We propose design automation techniques for architecting efficient neural networks given a target hardware platform. We investigate automatically designing specialized and fast models, auto channel pruning, and auto mixed-precision quantization. We demonstrate such learning-based, automated design achieves superior performance and efficiency than rule-based human design. Moreover, we shorten the design cycle by 200 than previous work, so that we can afford to design specialized neural network models for different hardware platforms.
Tianzhe Wang*, Zetian Jiang*, Junchi Yan
Jointly matching of multiple graphs is challenging and recently has been an active topic in machine learning and computer vision. State-of-the-art methods have been devised, however, to our best knowledge there is no effective mechanism that can explicitly deal with the matching of a mixture of graphs belonging to multiple clusters, e.g., a collection of bikes and bottles. Seeing its practical importance, we propose a novel approach for multiple graph matching and clustering.
Shuai Wang, Yexin Yang, Tianzhe Wang, Yanmin Qian, Kai Yu
In this paper, knowledge distillation, which is also known as teacher-student learning, is used to reduce the performance gap between the large and small models. Knowledge distillation at two levels are investigated, the label level and the embedding level. Experiments are carried out on the VoxCeleb1 dataset, and results show that the performance of the small shallow model can be boosted significantly by the proposed methods, and the gap between the large deep and small shallow models can be substantially reduced.