Deep learning with differential privacy (DP) has garnered significant attention over the past years, leading to the development of numerous methods aimed at enhancing model accuracy and training efficiency. This paper delves into the problem of training Transformer models with differential privacy. Our treatment is modular: the logic is to 'reduce' the problem of training DP Transformer to the more basic problem of training DP vanilla neural nets. The latter is better understood and amenable to many model-agnostic methods. Such 'reduction' is done by first identifying the hardness unique to DP Transformer training: the attention distraction phenomenon and a lack of compatibility with existing techniques for efficient gradient clipping. To deal with these two issues, we propose the Re-Attention Mechanism and Phantom Clipping, respectively. We believe that our work not only casts new light on training DP Transformers but also promotes a modular treatment to advance research in the field of differentially private deep learning.

A recommendation system is of vital importance in delivering personalization services, which often brings continuous dual improvement in user experience and organization revenue. However, the data of one single organization may not be enough to build an accurate recommendation model for inactive or new cold-start users. Moreover, due to the recent regulatory restrictions on user privacy and data security, as well as the commercial conflicts, the raw data in different organizations cannot be merged to alleviate the scarcity issue in training a model. In order to learn users’ preferences from such cross-silo data of different organizations and then provide recommendations to the cold-start users, we propose a novel federated learning framework, i.e., federated cross-organization recommendation ecosystem (FedCORE). Specifically, we first focus on the ecosystem problem of cross-organization federated recommendation, including cooperation patterns and privacy protection. For the former, we propose a privacy-aware collaborative training and inference algorithm. For the latter, we define four levels of privacy leakage and propose some methods for protecting the privacy. We then conduct extensive experiments on three real-world datasets and two seminal recommendation models to study the impact of cooperation in our proposed ecosystem and the effectiveness of privacy protection.

Traditional federated learning (FL) algorithms, such as FedAvg, fail to handle non-i.i.d data because they learn a global model by simply averaging biased local models that are trained on non-i.i.d local data, therefore failing to model the global data distribution. In this paper, we present a novel Bayesian FL algorithm that successfully handles such a non-i.i.d FL setting by enhancing the local training task with an auxiliary task that explicitly estimates the global data distribution. One key challenge in estimating the global data distribution is that the data are partitioned in FL, and therefore the ground-truth global data distribution is inaccessible. To address this challenge, we propose an expectation-propagation-inspired probabilistic neural network, dubbed federated neural propagation (FedNP), which efficiently estimates the global data distribution given non-i.i.d data partitions. Our algorithm is sampling-free and end-to-end differentiable, can be applied with any conventional FL frameworks and learns richer global data representation. Experiments on both image classification tasks with synthetic non-i.i.d image data partitions and real-world non-i.i.d speech recognition tasks demonstrate that our framework effectively alleviates the performance deterioration caused by non-i.i.d data.

In order to mine latent semantics from text data, word embedding and topic modeling are two major methodologies in industry. From a pragmatic perspective, each of these two lines of semantic models faces increasing challenges from real-life applications. However, modern text mining tasks typically require a panoramic view of the latent semantics. Hence, discovering heterogeneous semantics (e.g., heterogeneous types of latent topics) is critical for the performance of these tasks, and it is necessary to design a model that meets this demand. Furthermore, with the arrival of the big data era and the increasing awareness of data privacy, it is necessary to study the issues of mining heterogeneous semantics with high efficiency while avoiding compromising data privacy. In this work, we develop a novel method called Heterogeneous Latent Topic Discovery (HLTD) which seamlessly integrates topic modeling with word embedding to discover heterogeneous latent topics. By coupling parameter-server architecture with new private sampling algorithms, HLTD can be efficiently trained with effective protection of underlying data privacy. We evaluate HLTD through a wide range of qualitative and quantitative metrics in industry. Extensive experiments demonstrates the superiority of HLTD over the state-of-the-arts.

Automatic Speech Recognition (ASR) is playing a vital role in a wide range of real-world applications. However, Commercial ASR solutions are typically “one-size-fits-all” products and clients are inevitably faced with the risk of severe performance degradation in field test. Meanwhile, with new data regulations such as the European Union’s General Data Protection Regulation (GDPR) coming into force, ASR vendors, which traditionally utilize the speech training data in a centralized approach, are becoming increasingly helpless to solve this problem, since accessing clients’ speech data is prohibited. Here, we show that by seamlessly integrating three machine learning paradigms (i.e., T ransfer learning, F ederated learning, and E volutionary learning (TFE)), we can successfully build a win-win ecosystem for ASR clients and vendors and solve all the aforementioned problems plaguing them. Through large-scale quantitative experiments, we show that with TFE, the clients can enjoy far better ASR solutions than the “one-size-fits-all” counterpart, and the vendors can exploit the abundance of clients’ data to effectively refine their own ASR products.

Probabilistic topic modeling has been applied in a variety of industrial applications. Training a high-quality model usually requires a massive amount of data to provide comprehensive co-occurrence information for the model to learn. However, industrial data such as medical or financial records are often proprietary or sensitive, which precludes uploading to data centers. Hence, training topic models in industrial scenarios using conventional approaches faces a dilemma: A party (i.e., a company or institute) has to either tolerate data scarcity or sacrifice data privacy. In this article, we propose a framework named Industrial Federated Topic Modeling (iFTM), in which multiple parties collaboratively train a high-quality topic model by simultaneously alleviating data scarcity and maintaining immunity to privacy adversaries. iFTM is inspired by federated learning, supports two representative topic models (i.e., Latent Dirichlet Allocation and SentenceLDA) in industrial applications, and consists of novel techniques such as private Metropolis-Hastings, topic-wise normalization, and heterogeneous model integration. We conduct quantitative evaluations to verify the effectiveness of iFTM and deploy iFTM in two real-life applications to demonstrate its utility. Experimental results verify iFTM’s superiority over conventional topic modeling.

Due to the rising awareness of privacy protection and the voluminous scale of speech data, it is becoming infeasible for Automatic Speech Recognition (ASR) system developers to train the acoustic model with complete data as before. In this paper, we propose a novel Divide-and-Merge paradigm to solve salient problems plaguing the ASR field. In the Divide phase, multiple acoustic models are trained based upon different subsets of the complete speech data, while in the Merge phase two novel algorithms are utilized to generate a high-quality acoustic model based upon those trained on data subsets. We first propose the Genetic Merge Algorithm (GMA), which is a highly specialized algorithm for optimizing acoustic models but suffers from low efficiency. We further propose the SGD-Based Optimizational Merge Algorithm (SOMA), which effectively alleviates the efficiency bottleneck of GMA and maintains superior performance. Extensive experiments on public data show that the proposed methods can significantly outperform the state-of-the-art.

Topic modeling has been widely applied in a variety of industrial applications. Training a high-quality model usually requires massive amount of in-domain data, in order to provide comprehensive co-occurrence information for the model to learn. However, industrial data such as medical or financial records are often proprietary or sensitive, which precludes uploading to data centers. Hence training topic models in industrial scenarios using conventional approaches faces a dilemma: a party (i.e., a company or institute) has to either tolerate data scarcity or sacrifice data privacy. In this paper, we propose a novel framework named Federated Topic Modeling (FTM), in which multiple parties collaboratively train a high-quality topic model by simultaneously alleviating data scarcity and maintaining immune to privacy adversaries. FTM is inspired by federated learning and consists of novel techniques such as private Metropolis Hastings, topic-wise normalization and heterogeneous model integration. We conduct a series of quantitative evaluations to verify the effectiveness of FTM and deploy FTM in an Automatic Speech Recognition (ASR) system to demonstrate its utility in real-life applications. Experimental results verify FTM's superiority over conventional topic modeling.

Dialogue speech widely exists in scenarios such as chitchat, meeting and customer service. General-purpose speech recognition systems usually neglect the topic information in the context of dialogue speech, which has great potential for improving the performance of speech recognition. In this paper, we propose a transfer learning mechanism to conduct topic-aware recognition for dialogue speech. We first propose a new probabilistic topic model named Dialogue Speech Topic Model (DSTM) that is specialized for modeling the context of dialogue speech. We further propose a novel transfer learning mechanism for DSTM to significantly reduce its training cost while preserving its effectiveness for accurate topic inference. The experiment results demonstrate that proposed techniques in language model adaptation effectively improve the performance of the state-of-the-art Automatic Speech Recognition (ASR) system. Copyright © 2019 ISCA

With the recent advances of the sequence-to-sequence framework, generation approaches for the short text conversation (STC) become attractive. The traditional sequence-to-sequence approaches for the STC often suffer from poor diversity and general reply without substantiality. It is also hard to control the topic or semantics of the selected reply from multiple generated candidates. In this paper, a novel external-memory-driven sequence-to-sequence learning approach is proposed to address these problems. A tensor of the external memory is constructed to represent interpretable topics or semantics. During generation, a controllable memory trigger is extracted given the input sequence, and a reply is then generated using the memory trigger as well as the sequence-to-sequence model. Experiments show that the proposed approach can generate much richer diversity than the traditional sequence-to-sequence training with attention. Meanwhile, it achieves better quality score in human evaluation. It is also observed that by manually manipulating the memory trigger, it is possible to interpretably guide the topics or semantics of the reply.