Pre-training is known to generate universal representations for downstream tasks in large-scale deep learning such as large language models. Existing literature, e.g., Kim et al. (2020), empirically observe that the downstream tasks can inherit the adversarial robustness of the pre-trained model. We provide theoretical justifications for this robustness inheritance phenomenon. Our theoretical results reveal that feature purification plays an important role in connecting the adversarial robustness of the pre-trained model and the downstream tasks in two-layer neural networks. Specifically, we show that (i) with adversarial training, each hidden node tends to pick only one (or a few) feature; (ii) without adversarial training, the hidden nodes can be vulnerable to attacks. This observation is valid for both supervised pre-training and contrastive learning. With purified nodes, it turns out that clean training is enough to achieve adversarial robustness in downstream tasks.

In this paper, we propose VidLA, an approach for video-language alignment at scale. There are two major limitations of previous video-language alignment approaches. First, they do not capture both short-range and long-range temporal dependencies and typically employ complex hierarchical deep network architectures that are hard to integrate with existing pretrained image-text foundation models. To effectively address this limitation, we instead keep the network architecture simple and use a set of data tokens that operate at different temporal resolutions in a hierarchical manner, accounting for the temporally hierarchical nature of videos. By employing a simple two-tower architecture, we are able to initialize our video-language model with pretrained image-text foundation models, thereby boosting the final performance. Second, existing video-language alignment works struggle due to the lack of semantically aligned large-scale training data. To overcome it, we leverage recent LLMs to curate the largest video-language dataset to date with better visual grounding. Furthermore, unlike existing video-text datasets which only contain short clips, our dataset is enriched with video clips of varying durations to aid our temporally hierarchical data to-kens in extracting better representations at varying temporal scales. Overall, empirical results show that our proposed approach surpasses state-of-the-art methods on Multiple retrieval benchmarks, especially on longer videos, and performs competitively on classification benchmarks.

Diffusion-based generative modeling has been achieving state-of-the-art results on various generation tasks. Most diffusion models, however, are limited to a single-generation modeling. Can we generalize diffusion models with the ability of multi-modal generative training for more generalizable modeling? In this paper, we propose a principled way to define a diffusion model by constructing a unified multi-modal diffusion model in a common {\em diffusion space}. We define the forward diffusion process to be driven by an information aggregation from multiple types of task-data, {\it e.g.}, images for a generation task and labels for a classification task. In the reverse process, we enforce information sharing by parameterizing a shared backbone denoising network with additional modality-specific decoder heads. Such a structure can simultaneously learn to generate different types of multi-modal data with a multi-task loss, which is derived from a new multi-modal variational lower bound that generalizes the standard diffusion model. We propose several multi-modal generation settings to verify our framework, including image transition, masked-image training, joint image-label and joint image-representation generative modeling. Extensive experimental results on ImageNet indicate the effectiveness of our framework for various multi-modal generative modeling, which we believe is an important research direction worthy of more future explorations.

Multi-task learning (MTL) considers learning a joint model for multiple tasks by optimizing a convex combination of all task losses. To solve the optimization problem, existing methods use an adaptive weight updating scheme, where task weights are dynamically adjusted based on their respective losses to prioritize difficult tasks. However, these algorithms face a great challenge whenever label noise is present, in which case excessive weights tend to be assigned to noisy tasks that have relatively large Bayes optimal errors, thereby overshadowing other tasks and causing performance to drop across the board. To overcome this limitation, we propose Multi-Task Learning with Excess Risks (ExcessMTL), an excess risk-based task balancing method that updates the task weights by their distances to convergence instead. Intuitively, ExcessMTL assigns higher weights to worse-trained tasks that are further from convergence. To estimate the excess risks, we develop an efficient and accurate method with Taylor approximation. Theoretically, we show that our proposed algorithm achieves convergence guarantees and Pareto stationarity. Empirically, we evaluate our algorithm on various MTL benchmarks and demonstrate its superior performance over existing methods in the presence of label noise. Our code is available at https://github.com/yifei-he/ExcessMTL.

Graph machine learning (GML) is effective in many business applications. However, making GML easy to use and applicable to industry applications with massive datasets remain challenging. We developed GraphStorm, which provides an end-to-end solution for scalable graph construction, graph model training and inference. GraphStorm has the following desirable properties: (a) Easy to use: it can perform graph construction and model training and inference with just a single command; (b) Expert-friendly: GraphStorm contains many advanced GML modeling techniques to handle complex graph data and improve model performance; (c) Scalable: every component in GraphStorm can operate on graphs with billions of nodes and can scale model training and inference to different hardware without changing any code. GraphStorm has been used and deployed for over a dozen billion-scale industry applications after its release in May 2023. It is open-sourced in Github: https://github.com/awslabs/graphstorm.

Hierarchies are common structures used to organize data, such as e-commerce hierarchies associated with product data. With these product hierarchies, we aim to learn hierarchy-aware product text embeddings to improve fine-tuning performance on a variety of downstream e-commerce tasks. Existing methods leverage hierarchies by either aligning the text embeddings to separate hierarchical embeddings or by aligning the hierarchical information implicitly within a unified text Transformer. Although these models optimize to predict hierarchy information, performing further fine-tuning on new tasks is non-trivial. To bridge this gap, we propose a pre-training architecture to implicitly encode the hierarchy within the product text and then directly leverage a sub-set of the pre-training model during fine-tuning. Pre-training is done through Semantic and Structural Transformers (SST) where the Semantic-Transformer first encodes the product text into a contextual embedding, which is then used by the Structural-Transformer to infer the product’s path in the hierarchy. Fine-tuning is done using only the initial Semantic-Transformer, now that hierarchy-aware text embeddings are learned. With this design, we eliminate the need of linking each fine-tuning dataset with corresponding hierarchies. This leads to fine-tuning performance improvements on critical e-commerce downstream tasks over the existing state-of-the-art hierarchy models, even when hierarchy data $is$ available during fine-tuning. Moreover, this improvement is consistent even after augmenting our baseline models to support fine-tuning. We conclude by discussing how such implicit structural encodings can be leveraged beyond the e-commerce domain.

Developing text mining approaches to mine aspects from customer reviews has been well-studied due to its importance in understanding customer needs and product attributes. In contrast, it remains unclear how to predict the future emerging aspects of a new product that currently has little review information. This task, which we named product aspect forecasting, is critical for recommending new products, but also challenging because of the missing reviews. Here, we propose ForeSeer, a novel textual mining and product embedding approach progressively trained on temporal product graphs for this novel product aspect forecasting task. ForeSeer transfers reviews from similar products on a large product graph and exploits these reviews to predict aspects that might emerge in future reviews. A key novelty of our method is to jointly provide review, product, and aspect embeddings that are both time-sensitive and less affected by extremely imbalanced aspect frequencies. We evaluated ForeSeer on a real-world product review system containing 11,536,382 reviews and 11,000 products over 3 years. We observe that ForeSeer substantially outperformed existing approaches with at least 49.1% AUPRC improvement under the real setting where aspect associations are not given. ForeSeer further improves future link prediction on the product graph and the review aspect association prediction. Collectively, Foreseer offers a novel framework for review forecasting by effectively integrating review text, product network, and temporal information, opening up new avenues for online shopping recommendation and e-commerce applications.

Contrastive loss has been increasingly used in learning representations from multiple modalities. In the limit, the nature of the contrastive loss encourages modalities to exactly match each other in the latent space. Yet it remains an open question how the modality alignment affects the downstream task performance. In this paper, based on an information-theoretic argument, we first prove that exact modality alignment is sub-optimal in general for down-stream prediction tasks. Hence we advocate that the key of better performance lies in meaningful latent modality structures instead of perfect modality alignment. To this end, we propose three general approaches to construct latent modality structures. Specifically, we design 1) a deep feature separation loss for intra-modality regularization; 2) a Brownian-bridge loss for inter-modality regularization; and 3) a geometric consistency loss for both intra- and intermodality regularization. Extensive experiments are conducted on two popular multi-modal representation learning frameworks: the CLIP-based two-tower model and the ALBEF-based fusion model. We test our model on a variety of tasks including zero/few-shot image classification, image-text retrieval, visual question answering, visual reasoning, and visual entailment. Our method achieves consistent improvements over existing methods, demonstrating the effectiveness and generalizability of our proposed approach on latent modality structure regularization.

Zhan Shi, Guoyin Wang, Ke Bai, Jiwei Li, Xiang Li, Qingjun Cui, Belinda Zeng, Trishul Chilimbi, Xiaodan Zhu. Proceedings of the 2023 Conference on Empirical Methods in Natural Language Processing. 2023.

Model pre-training on large text corpora has been demonstrated effective for various downstream applications in the NLP domain. In the graph mining domain, a similar analogy can be drawn for pre-training graph models on large graphs in the hope of benefiting downstream graph applications, which has also been explored by several recent studies. However, no existing study has ever investigated the pre-training of text plus graph models on large heterogeneous graphs with abundant textual information (a.k.a. large graph corpora) and then fine-tuning the model on different related downstream applications with different graph schemas. To address this problem, we propose a framework of graph-aware language model pre-training (GaLM) on a large graph corpus, which incorporates large language models and graph neural networks, and a variety of fine-tuning methods on downstream applications. We conduct extensive experiments on Amazon's real internal datasets and large public datasets. Comprehensive empirical results and in-depth analysis demonstrate the effectiveness of our proposed methods along with lessons learned.

Abstract Dense embedding-based semantic matching is widely used in e-commerce product search to address the shortcomings of lexical matching such as sensitivity to spelling variants. The recent advances in BERT-like language model encoders, have however, not found their way to realtime search due to the strict inference latency requirement imposed on e-commerce websites. While bi-encoder BERT architectures enable fast approximate nearest neighbor search, training them effectively on query-product data remains a challenge due to training instabilities and the persistent generalization gap with cross-encoders. In this work, we propose a four-stage training procedure to leverage large BERT-like models for product search while preserving low inference latency. We introduce query-product interaction pre-finetuning to effectively pretrain BERT bi-encoders for matching and improve generalization. Through offline experiments on an e-commerce product dataset, we show that a distilled small BERT-based model (75M params) trained using our approach improves the search relevance metric by up to 23% over a baseline DSSM-based model with similar inference latency. The small model only suffers a 3% drop in relevance metric compared to the 20x larger teacher. We also show using online A/B tests at scale, that our approach improves over the production model in exact and substitute products retrieved.

Aligning signals from different modalities is an important step in vision-language representation learning as it affects the performance of later stages such as cross-modality fusion. Since image and text typically reside in different regions of the feature space, directly aligning them at instance level is challenging especially when features are still evolving during training. In this paper, we propose to align at a higher and more stable level using cluster representation. Specifically, we treat image and text as two “views” of the same entity, and encode them into a joint vision-language coding space spanned by a dictionary of cluster centers (codebook). We contrast positive and negative samples via their cluster assignments while simultaneously optimizing the cluster centers. To further smooth out the learning process, we adopt a teacher-student distillation paradigm, where the momentum teacher of one view guides the student learning of the other. We evaluated our approach on common vision language benchmarks and obtain new SoTA on zero-shot cross modality retrieval while being competitive on various other transfer tasks.

Vision-language representation learning largely benefits from image-text alignment through contrastive losses (e.g., InfoNCE loss). The success of this alignment strategy is attributed to its capability in maximizing the mutual information (MI) between an image and its matched text. However, simply performing cross-modal alignment (CMA) ignores data potential within each modality, which may result in degraded representations. For instance, although CMA-based models are able to map image-text pairs close together in the embedding space, they fail to ensure that similar inputs from the same modality stay close by. This problem can get even worse when the pre-training data is noisy. In this paper, we propose triple contrastive learning (TCL) for vision-language pre-training by leveraging both cross-modal and intra-modal self-supervision. Besides CMA, TCL introduces an intra-modal contrastive objective to provide complementary benefits in representation learning. To take advantage of localized and structural information from image and text input, TCL further maximizes the average MI between local regions of image/text and their global summary. To the best of our knowledge, ours is the first work that takes into account local structure information for multi-modality representation learning. Experimental evaluations show that our approach is competitive and achieves the new state of the art on various common downstream vision-language tasks such as image-text retrieval and visual question answering.

Contrastive learning (CL) has been the de facto technique for self-supervised representation learning (SSL), with impressive empirical success such as multi-modal representation learning. However, traditional CL loss only considers negative samples from a minibatch, which could cause biased gradients due to the non-decomposibility of the loss. For the first time, we consider optimizing a more generalized contrastive loss, where each data sample is associated with an infinite number of negative samples. We show that directly using minibatch stochastic optimization could lead to gradient bias. To remedy this, we propose an efficient Bayesian data augmentation technique to augment the contrastive loss into a decomposable one, where standard stochastic optimization can be directly applied without gradient bias. Specifically, our augmented loss defines a joint distribution over the model parameters and the augmented parameters, which can be conveniently optimized by a proposed stochastic expectation-maximization algorithm. Our framework is more general and is related to several popular SSL algorithms. We verify our framework on both small scale models and several large foundation models, including SSL of ImageNet and SSL for vision-language representation learning. Experiment results indicate the existence of gradient bias in all cases, and demonstrate the effectiveness of the proposed method on improving previous state of the arts. Remarkably, our method can outperform the strong MoCo-v3 under the same hyper-parameter setting with only around half of the minibatch size; and also obtains strong results in the recent public benchmark ELEVATER for few-shot image classification.

Spoken language understanding (SLU) is the task of inferring the semantics of spoken utterances. Traditionally, this has been achieved with a cascading combination of Automatic Speech Recognition (ASR) and Natural Language Understanding (NLU) modules that are optimized separately, which can lead to a suboptimal overall performance. More recently, End-to-End SLU (E2E SLU) was proposed to perform SLU directly from speech through a joint optimization of the modules, addressing some of the traditional SLU shortcomings. A key challenge of this approach is how to best integrate the feature learning of the ASR and NLU sub-tasks to maximize their performance. While it is known that in general, ASR models focus on low-level features, and NLU models need higher-level contextual information, ASR models can nonetheless also leverage top-down syntactic and semantic information to improve their recognition. Based on this insight, we propose Top-Down SLU (TD-SLU), a new transformer-based E2E SLU model that uses top-down attention and an attention gate to fuse high-level NLU features with low-level ASR features, which leads to a better optimization of both tasks. We have validated our model using the public FluentSpeech set, and a large custom dataset. Results show TD-SLU is able to outperform selected baselines both in terms of ASR and NLU quality metrics, and suggest that the added syntactic and semantic high-level information can improve the model’s performance.