Abstract:Deep Learning Recommendation Model(DLRM)s utilize the embedding layer to represent various categorical features. Traditional DLRMs adopt unified embedding size for all features, leading to suboptimal performance and redundant parameters. Thus, lots of Automatic Embedding size Search (AES) works focus on obtaining mixed embedding sizes with strong model performance. However, previous AES works can hardly address several challenges together: (1) The search results of embedding sizes are unstable; (2) Recommendation effect with AES results is unsatisfactory; (3) Memory cost of embeddings is uncontrollable. To address these challenges, we propose a novel one-shot AES framework called AdaS&S, in which a supernet encompassing various candidate embeddings is built and AES is performed as searching network architectures within it. Our framework contains two main stages: In the first stage, we decouple training parameters from searching embedding sizes, and propose the Adaptive Sampling method to yield a well-trained supernet, which further helps to produce stable AES results. In the second stage, to obtain embedding sizes that benefits the model effect, we design a reinforcement learning search process which utilizes the supernet trained previously. Meanwhile, to adapt searching to specific resource constraint, we introduce the resource competition penalty to balance the model effectiveness and memory cost of embeddings. We conduct extensive experiments on public datasets to show the superiority of AdaS&S. Our method could improve AUC by about 0.3% while saving about 20% of model parameters. Empirical analysis also shows that the stability of searching results in AdaS&S significantly exceeds other methods.
Abstract:Given a Hyperparameter Optimization(HPO) problem, how to design an algorithm to find optimal configurations efficiently? Bayesian Optimization(BO) and the multi-fidelity BO methods employ surrogate models to sample configurations based on history evaluations. More recent studies obtain better performance by integrating BO with HyperBand(HB), which accelerates evaluation by early stopping mechanism. However, these methods ignore the advantage of a suitable evaluation scheme over the default HyperBand, and the capability of BO is still constrained by skewed evaluation results. In this paper, we propose FlexHB, a new method pushing multi-fidelity BO to the limit as well as re-designing a framework for early stopping with Successive Halving(SH). Comprehensive study on FlexHB shows that (1) our fine-grained fidelity method considerably enhances the efficiency of searching optimal configurations, (2) our FlexBand framework (self-adaptive allocation of SH brackets, and global ranking of configurations in both current and past SH procedures) grants the algorithm with more flexibility and improves the anytime performance. Our method achieves superior efficiency and outperforms other methods on various HPO tasks. Empirical results demonstrate that FlexHB can achieve up to 6.9X and 11.1X speedups over the state-of-the-art MFES-HB and BOHB respectively.
Abstract:We present Distributed Equivalent Substitution (DES) training, a novel distributed training framework for recommender systems with large-scale dynamic sparse features. Our framework achieves faster convergence with less communication overhead and better computing resource utilization. DES strategy splits a weights-rich operator into sub-operators with co-located weights and aggregates partial results with much smaller communication cost to form a computationally equivalent substitution to the original operator. We show that for different types of models that recommender systems use, we can always find computational equivalent substitutions and splitting strategies for their weights-rich operators with theoretical communication load reduced ranging from 72.26% to 99.77%. We also present an implementation of DES that outperforms state-of-the-art recommender systems. Experiments show that our framework achieves up to 83% communication savings compared to other recommender systems, and can bring up to 4.5x improvement on throughput for deep models.
Abstract:Although attention-based Neural Machine Translation (NMT) has achieved remarkable progress in recent years, it still suffers from issues of repeating and dropping translations. To alleviate these issues, we propose a novel key-value memory-augmented attention model for NMT, called KVMEMATT. Specifically, we maintain a timely updated keymemory to keep track of attention history and a fixed value-memory to store the representation of source sentence throughout the whole translation process. Via nontrivial transformations and iterative interactions between the two memories, the decoder focuses on more appropriate source word(s) for predicting the next target word at each decoding step, therefore can improve the adequacy of translations. Experimental results on Chinese=>English and WMT17 German<=>English translation tasks demonstrate the superiority of the proposed model.
Abstract:Chinese poetry generation is a very challenging task in natural language processing. In this paper, we propose a novel two-stage poetry generating method which first plans the sub-topics of the poem according to the user's writing intent, and then generates each line of the poem sequentially, using a modified recurrent neural network encoder-decoder framework. The proposed planning-based method can ensure that the generated poem is coherent and semantically consistent with the user's intent. A comprehensive evaluation with human judgments demonstrates that our proposed approach outperforms the state-of-the-art poetry generating methods and the poem quality is somehow comparable to human poets.