UME: Upcycling Mixture-of-Experts for Scalable and Efficient Automatic Speech Recognition

Recent advancements in scaling up models have significantly improved performance in Automatic Speech Recognition (ASR) tasks. However, training large ASR models from scratch remains costly. To address this issue, we introduce UME, a novel method that efficiently Upcycles pretrained dense ASR checkpoints into larger Mixture-of-Experts (MoE) architectures. Initially, feed-forward networks are converted into MoE layers. By reusing the pretrained weights, we establish a robust foundation for the expanded model, significantly reducing optimization time. Then, layer freezing and expert balancing strategies are employed to continue training the model, further enhancing performance. Experiments on a mixture of 170k-hour Mandarin and English datasets show that UME: 1) surpasses the pretrained baseline by a margin of 11.9% relative error rate reduction while maintaining comparable latency; 2) reduces training time by up to 86.7% and achieves superior accuracy compared to training models of the same size from scratch.

Legommenders: A Comprehensive Content-Based Recommendation Library with LLM Support

We present Legommenders, a unique library designed for content-based recommendation that enables the joint training of content encoders alongside behavior and interaction modules, thereby facilitating the seamless integration of content understanding directly into the recommendation pipeline. Legommenders allows researchers to effortlessly create and analyze over 1,000 distinct models across 15 diverse datasets. Further, it supports the incorporation of contemporary large language models, both as feature encoder and data generator, offering a robust platform for developing state-of-the-art recommendation models and enabling more personalized and effective content delivery.

STORE: Streamlining Semantic Tokenization and Generative Recommendation with A Single LLM

Traditional recommendation models often rely on unique item identifiers (IDs) to distinguish between items, which can hinder their ability to effectively leverage item content information and generalize to long-tail or cold-start items. Recently, semantic tokenization has been proposed as a promising solution that aims to tokenize each item's semantic representation into a sequence of discrete tokens. In this way, it preserves the item's semantics within these tokens and ensures that semantically similar items are represented by similar tokens. These semantic tokens have become fundamental in training generative recommendation models. However, existing generative recommendation methods typically involve multiple sub-models for embedding, quantization, and recommendation, leading to an overly complex system. In this paper, we propose to streamline the semantic tokenization and generative recommendation process with a unified framework, dubbed STORE, which leverages a single large language model (LLM) for both tasks. Specifically, we formulate semantic tokenization as a text-to-token task and generative recommendation as a token-to-token task, supplemented by a token-to-text reconstruction task and a text-to-token auxiliary task. All these tasks are framed in a generative manner and trained using a single LLM backbone. Extensive experiments have been conducted to validate the effectiveness of our STORE framework across various recommendation tasks and datasets. We will release the source code and configurations for reproducible research.

Do self-supervised speech and language models extract similar representations as human brain?

Speech and language models trained through self-supervised learning (SSL) demonstrate strong alignment with brain activity during speech and language perception. However, given their distinct training modalities, it remains unclear whether they correlate with the same neural aspects. We directly address this question by evaluating the brain prediction performance of two representative SSL models, Wav2Vec2.0 and GPT-2, designed for speech and language tasks. Our findings reveal that both models accurately predict speech responses in the auditory cortex, with a significant correlation between their brain predictions. Notably, shared speech contextual information between Wav2Vec2.0 and GPT-2 accounts for the majority of explained variance in brain activity, surpassing static semantic and lower-level acoustic-phonetic information. These results underscore the convergence of speech contextual representations in SSL models and their alignment with the neural network underlying speech perception, offering valuable insights into both SSL models and the neural basis of speech and language processing.

Neural2Speech: A Transfer Learning Framework for Neural-Driven Speech Reconstruction

Reconstructing natural speech from neural activity is vital for enabling direct communication via brain-computer interfaces. Previous efforts have explored the conversion of neural recordings into speech using complex deep neural network (DNN) models trained on extensive neural recording data, which is resource-intensive under regular clinical constraints. However, achieving satisfactory performance in reconstructing speech from limited-scale neural recordings has been challenging, mainly due to the complexity of speech representations and the neural data constraints. To overcome these challenges, we propose a novel transfer learning framework for neural-driven speech reconstruction, called Neural2Speech, which consists of two distinct training phases. First, a speech autoencoder is pre-trained on readily available speech corpora to decode speech waveforms from the encoded speech representations. Second, a lightweight adaptor is trained on the small-scale neural recordings to align the neural activity and the speech representation for decoding. Remarkably, our proposed Neural2Speech demonstrates the feasibility of neural-driven speech reconstruction even with only 20 minutes of intracranial data, which significantly outperforms existing baseline methods in terms of speech fidelity and intelligibility.

Leveraging Label Information for Multimodal Emotion Recognition

Multimodal emotion recognition (MER) aims to detect the emotional status of a given expression by combining the speech and text information. Intuitively, label information should be capable of helping the model locate the salient tokens/frames relevant to the specific emotion, which finally facilitates the MER task. Inspired by this, we propose a novel approach for MER by leveraging label information. Specifically, we first obtain the representative label embeddings for both text and speech modalities, then learn the label-enhanced text/speech representations for each utterance via label-token and label-frame interactions. Finally, we devise a novel label-guided attentive fusion module to fuse the label-aware text and speech representations for emotion classification. Extensive experiments were conducted on the public IEMOCAP dataset, and experimental results demonstrate that our proposed approach outperforms existing baselines and achieves new state-of-the-art performance.

Co-evolving Vector Quantization for ID-based Recommendation

Category information plays a crucial role in enhancing the quality and personalization of recommendations. Nevertheless, the availability of item category information is not consistently present, particularly in the context of ID-based recommendations. In this work, we propose an alternative approach to automatically learn and generate entity (i.e., user and item) categorical information at different levels of granularity, specifically for ID-based recommendation. Specifically, we devise a co-evolving vector quantization framework, namely COVE, which enables the simultaneous learning and refinement of code representation and entity embedding in an end-to-end manner, starting from the randomly initialized states. With its high adaptability, COVE can be easily integrated into existing recommendation models. We validate the effectiveness of COVE on various recommendation tasks including list completion, collaborative filtering, and click-through rate prediction, across different recommendation models. We will publish the code and data for other researchers to reproduce our work.

Neighborhood-based Hard Negative Mining for Sequential Recommendation

Negative sampling plays a crucial role in training successful sequential recommendation models. Instead of merely employing random negative sample selection, numerous strategies have been proposed to mine informative negative samples to enhance training and performance. However, few of these approaches utilize structural information. In this work, we observe that as training progresses, the distributions of node-pair similarities in different groups with varying degrees of neighborhood overlap change significantly, suggesting that item pairs in distinct groups may possess different negative relationships. Motivated by this observation, we propose a Graph-based Negative sampling approach based on Neighborhood Overlap (GNNO) to exploit structural information hidden in user behaviors for negative mining. GNNO first constructs a global weighted item transition graph using training sequences. Subsequently, it mines hard negative samples based on the degree of overlap with the target item on the graph. Furthermore, GNNO employs curriculum learning to control the hardness of negative samples, progressing from easy to difficult. Extensive experiments on three Amazon benchmarks demonstrate GNNO's effectiveness in consistently enhancing the performance of various state-of-the-art models and surpassing existing negative sampling strategies. The code will be released at \url{https://github.com/floatSDSDS/GNNO}.

Multi-modal Pre-training for Medical Vision-language Understanding and Generation: An Empirical Study with A New Benchmark

With the availability of large-scale, comprehensive, and general-purpose vision-language (VL) datasets such as MSCOCO, vision-language pre-training (VLP) has become an active area of research and proven to be effective for various VL tasks such as visual-question answering. However, studies on VLP in the medical domain have so far been scanty. To provide a comprehensive perspective on VLP for medical VL tasks, we conduct a thorough experimental analysis to study key factors that may affect the performance of VLP with a unified vision-language Transformer. To allow making sound and quick pre-training decisions, we propose RadioGraphy Captions (RGC), a high-quality, multi-modality radiographic dataset containing 18,434 image-caption pairs collected from an open-access online database MedPix. RGC can be used as a pre-training dataset or a new benchmark for medical report generation and medical image-text retrieval. By utilizing RGC and other available datasets for pre-training, we develop several key insights that can guide future medical VLP research and new strong baselines for various medical VL tasks.

OTF: Optimal Transport based Fusion of Supervised and Self-Supervised Learning Models for Automatic Speech Recognition

Self-Supervised Learning (SSL) Automatic Speech Recognition (ASR) models have shown great promise over Supervised Learning (SL) ones in low-resource settings. However, the advantages of SSL are gradually weakened when the amount of labeled data increases in many industrial applications. To further improve the ASR performance when abundant labels are available, we first explore the potential of combining SL and SSL ASR models via analyzing their complementarity in recognition accuracy and optimization property. Then, we propose a novel Optimal Transport based Fusion (OTF) method for SL and SSL models without incurring extra computation cost in inference. Specifically, optimal transport is adopted to softly align the layer-wise weights to unify the two different networks into a single one. Experimental results on the public 1k-hour English LibriSpeech dataset and our in-house 2.6k-hour Chinese dataset show that OTF largely outperforms the individual models with lower error rates.