Wen Wu, Bo Li, Chao Zhang, Chung-Cheng Chiu, Qiujia Li, Junwen Bai, Tara Sainath, Phil Woodland. Proceedings of the 62nd Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers). 2024.

The end-to-end ASR model is often desired in the streaming multilingual scenario since it is easier to deploy and can benefit from pre-trained speech models such as powerful foundation models. Meanwhile, the heterogeneous nature and imbalanced data abundance of different languages may cause performance degradation, leading to asynchronous peak performance for different languages during training, especially on tail ones. Sometimes even the data itself may become unavailable as a result of the enhanced privacy protection. Existing work tends to significantly increase the model size or learn language-specific decoders to accommodate each language separately. In this study, we explore simple yet effective Language-Dependent Adapter (LDA) finetuning under a cascaded Conformer transducer framework enhanced by teacher pseudolabeling for tail languages in the streaming multilingual ASR. The adapter only accounts for 0.4% of the full model per language. It is plugged into the frozen foundation model and is the only trainable module during the finetuning process with noisy student training. The final model merges the adapter parameters from different checkpoints for different languages. The model performance is validated on a challenging multilingual dictation dataset, which includes 39 tail languages across Latin, Greek, Arabic, etc. Our proposed method brings 12.2% word error rate reduction on average and up to 37.5% on a single locale. Furthermore, we show that our parameter-efficient LDA can match the quality of the full model finetuning, thus greatly alleviating the asynchronous peak performance issue.

Foundation models (FMs), that are trained on broad data at scale and are adaptable to a wide range of downstream tasks, have brought large interest in the research community. Benefiting from the diverse data sources such as different modalities, languages and application domains, foundation models have demonstrated strong generalization and knowledge transfer capabilities. In this paper, we present a pioneering study towards building an efficient solution for FM-based speech recognition systems. We adopt the recently developed self-supervised BEST-RQ for pretraining, and extend the joint training strategy JUST Hydra for finetuning using both source and unsuper-vised target domain data. The FM encoder adapter and decoder are then finetuned to the target domain with a small amount of super-vised in-domain data. On a large-scale YouTube and Voice Search task, our method is shown to be both data and model parameter efficient. It achieves the same quality with only 21.6M supervised in-domain data and 130.8M finetuned parameters, compared to the 731.1M model trained from scratch on additional 300M supervised in-domain data.

We propose Conditional Adapter (CoDA), a parameter-efficient transfer learning method that also improves inference efficiency. CoDA generalizes beyond standard adapter approaches to enable a new way of balancing speed and accuracy using conditional computation. Starting with an existing dense pretrained model, CoDA adds sparse activation together with a small number of new parameters and a light-weight training phase. Our experiments demonstrate that the CoDA approach provides an unexpectedly efficient way to transfer knowledge. Across a variety of language, vision, and speech tasks, CoDA achieves a 2x to 8x inference speed-up compared to the state-of-the-art Adapter approaches with moderate to no accuracy loss and the same parameter efficiency.

Climate change is posing new challenges to crop-related concerns, including food insecurity, supply stability, and economic planning. Accurately predicting crop yields is crucial for addressing these challenges. However, this prediction task is exceptionally complicated since crop yields depend on numerous factors such as weather, land surface, and soil quality, as well as their interactions. In recent years, machine learning models have been successfully applied in this domain. However, these models either restrict their tasks to a relatively small region, or only study over a single or few years, which makes them hard to generalize spatially and temporally. In this paper, we introduce a novel graph-based recurrent neural network for crop yield prediction, to incorporate both geographical and temporal knowledge in the model, and further boost predictive power. Our method is trained, validated, and tested on over 2000 counties from 41 states in the US mainland, covering years from 1981 to 2019. As far as we know, this is the first machine learning method that embeds geographical knowledge in crop yield prediction and predicts crop yields at the county level nationwide. We also laid a solid foundation by comparing our model on a nationwide scale with other well-known baseline methods, including linear models, tree-based models, and deep learning methods. Experiments show that our proposed method consistently outperforms the existing state-of-the-art methods on various metrics, validating the effectiveness of geospatial and temporal information.

Self-supervised training has shown promising gains in pretraining models and facilitating the downstream finetuning for speech recognition, like multilingual ASR. Most existing methods adopt a 2-stage scheme where the self-supervised loss is optimized in the first pretraining stage, and the standard supervised finetuning resumes in the second stage. In this paper, we propose an end-to-end (E2E) Joint Unsupervised and Supervised Training (JUST) method to combine the supervised RNN-T loss and the self-supervised contrastive and masked language modeling (MLM) losses. We validate its performance on the public dataset Multilingual LibriSpeech (MLS), which includes 8 languages and is extremely imbalanced. On MLS, we explore (1) JUST trained from scratch, and (2) JUST finetuned from a pretrained checkpoint. Experiments show that JUST can consistently outperform other existing state-of-the-art methods, and beat the monolingual baseline by a significant margin, demonstrating JUST’s capability of handling low-resource languages in multilingual ASR. Our average WER of all languages outperforms average monolingual baseline by 33.3%, and the state-of-the-art 2-stage XLSR by 32%. On low-resource languages like Polish, our WER is less than half of the monolingual baseline and even beats the supervised transfer learning method which uses external supervision.

Multi-label classification (MLC) is a prediction task where each sample can have more than one label. We propose a novel contrastive learning boosted multi-label prediction model based on a Gaussian mixture variational autoencoder (C-GMVAE), which learns a multimodal prior space and employs a contrastive loss. Many existing methods introduce extra complex neural modules like graph neural networks to capture the label correlations, in addition to the prediction modules. We find that by using contrastive learning in the supervised setting, we can exploit label information effectively in a data-driven manner, and learn meaningful feature and label embeddings which capture the label correlations and enhance the predictive power. Our method also adopts the idea of learning and aligning latent spaces for both features and labels. In contrast to previous works based on a unimodal prior, C-GMVAE imposes a Gaussian mixture structure on the latent space, to alleviate the posterior collapse and over-regularization issues. C-GMVAE outperforms existing methods on multiple public datasets and can often match other models’ full performance with only 50% of the training data. Furthermore, we show that the learnt embeddings provide insights into the interpretation of label-label interactions.

Understanding how environmental characteristics affect biodiversity patterns, from individual species to communities of species, is critical for mitigating effects of global change. A central goal for conservation planning and monitoring is the ability to accurately predict the occurrence of species communities and how these communities change over space and time. This in turn leads to a challenging and long-standing problem in the field of computer science - how to perform accurate multi-label classification with hundreds of labels? The key challenge of this problem is its exponential-sized output space with regards to the number of labels to be predicted. Therefore, it is essential to facilitate the learning process by exploiting correlations (or dependency) among labels. Previous methods mostly focus on modelling the correlation on label pairs; however, complex relations between real-world objects often go beyond second order. In this paper, we propose a novel framework for multi-label classification, High-order Tie-in Variational Autoencoder (HOT-VAE), which performs adaptive high-order label correlation learning. We experimentally verify that our model outperforms the existing state-of-the-art approaches on a bird distribution dataset on both conventional F1 scores and a variety of ecological metrics. To show our method is general, we also perform empirical analysis on seven other public real-world datasets in several application domains, and Hot-VAE exhibits superior performance to previous methods.

Building ASR models across many languages is a challenging multi-task learning problem due to large variations and heavily unbalanced data. Existing work has shown positive transfer from high resource to low resource languages. However, degradations on high resource languages are commonly observed due to interference from the heterogeneous multilingual data and reduction in per-language capacity. We conduct a capacity study on a 15-language task, with the amount of data per language varying from 7.6K to 53.5K hours. We adopt GShard [1] to efficiently scale up to 10B parameters. Empirically, we find that (1) scaling the number of model parameters is an effective way to solve the capacity bottleneck - our 500M-param model already outperforms monolingual baselines and scaling it to 1B and 10B brought further quality gains; (2) larger models are not only more data efficient, but also more efficient in terms of training cost as measured in TPU days - the 1B-param model reaches the same accuracy at 34% of training time as the 500M-param model; (3) given a fixed capacity budget, adding depth works better than width and large encoders do better than large decoders; (4) with continuous training, they can be adapted to new languages and domains.

Self-supervised disentangled representation learning is a critical task in sequence modeling. The learnt representations contribute to better model interpretability as well as the data generation, and improve the sample efficiency for downstream tasks. We propose a novel sequence representation learning method, named Contrastively Disentangled Sequential Variational Autoencoder (C-DSVAE), to extract and separate the static (time-invariant) and dynamic (time-variant) factors in the latent space. Different from previous sequential variational autoencoder methods, we use a novel evidence lower bound which maximizes the mutual information between the input and the latent factors, while penalizes the mutual information between the static and dynamic factors. We leverage contrastive estimations of the mutual information terms in training, together with simple yet effective augmentation techniques, to introduce additional inductive biases. Our experiments show that C-DSVAE significantly outperforms the previous state-of-the-art methods on multiple metrics.

Multi-label classification is the challenging task of predicting the presence and absence of multiple targets, involving representation learning and label correlation modeling. We propose a novel framework for multi-label classification, Multivariate Probit Variational AutoEncoder (MPVAE), that effectively learns latent embedding spaces as well as label correlations. MPVAE learns and aligns two probabilistic embedding spaces for labels and features respectively. The decoder of MPVAE takes in the samples from the embedding spaces and models the joint distribution of output targets under a Multivariate Probit model by learning a shared covariance matrix. We show that MPVAE outperforms the existing state-of-the-art methods on important computational sustainability applications as well as on other application domains, using public real-world datasets. MPVAE is further shown to remain robust under noisy settings. Lastly, we demonstrate the interpretability of the learned covariance by a case study on a bird observation dataset.

A key problem in computational sustainability is to understand the distribution of species across landscapes over time. This question gives rise to challenging large-scale prediction problems since (i) hundreds of species have to be simultaneously modeled and (ii) the survey data are usually inflated with zeros due to the absence of species for a large number of sites. The problem of tackling both issues simultaneously, which we refer to as the zero-inflated multi-target regression problem, has not been addressed by previous methods in statistics and machine learning. In this paper, we propose a novel deep model for the zero-inflated multi-target regression problem. To this end, we first model the joint distribution of multiple response variables as a multivariate probit model and then couple the positive outcomes with a multivariate log-normal distribution. By penalizing the difference between the two distributions’ covariance matrices, a link between both distributions is established. The whole model is cast as an end-to-end learning framework and we provide an efficient learning algorithm for our model that can be fully implemented on GPUs. We show that our model outperforms the existing state-of-the-art baselines on two challenging real-world species distribution datasets concerning bird and fish populations.

Many real-world tasks involve identifying signals from data satisfying background or prior knowledge. In domains like materials discovery, due to the flaws and biases in raw experimental data, the identification of X-ray diffraction (XRD) signals often requires significant (manual) expert work to find refined signals that are similar to the ideal theoretical ones. Automatically refining the raw XRD signals utilizing simulated theoretical data is thus desirable. We propose imitation refinement, a novel approach to refine imperfect input signals, guided by a pre-trained classifier incorporating prior knowledge from simulated theoretical data, such that the refined signals imitate the ideal ones. The classifier is trained on the ideal simulated data to classify signals and learns an embedding space where each class is represented by a prototype. The refiner learns to refine the imperfect signals with small modifications, such that their embeddings are closer to the corresponding prototypes. We show that the refiner can be trained in both supervised and unsupervised fashions. We further illustrate the effectiveness of the proposed approach both qualitatively and quantitatively in an X-ray diffraction signal refinement task in materials discovery.

Low precision operations can provide scalability, memory savings, portability, and energy efficiency. This paper proposes SWALP, an approach to low precision training that averages low-precision SGD iterates with a modified learning rate schedule. SWALP is easy to implement and can match the performance of full-precision SGD even with all numbers quantized down to 8 bits, including the gradient accumulators. Additionally, we show that SWALP converges arbitrarily close to the optimal solution for quadratic objectives, and to a noise ball asymptotically smaller than low precision SGD in strongly convex settings.