Decoder-only large language model (LLM)-based embedding models are beginning to outperform BERT or T5-based embedding models in general-purpose text embedding tasks, including dense vector-based retrieval. In this work, we introduce the NV-Embed model, incorporating architectural designs, training procedures, and curated datasets to significantly enhance the performance of LLM as a versatile embedding model, while maintaining its simplicity and reproducibility.For model architecture, we propose a latent attention layer to obtain pooled embeddings, which consistently improves retrieval and downstream task accuracy compared to mean pooling or using the last <EOS> token embedding from LLMs. To enhance representation learning, we remove the causal attention mask of LLMs during contrastive training. For training algorithm, we introduce a two-stage contrastive instruction-tuning method. It first applies contrastive training with instructions on retrieval datasets, utilizing in-batch negatives and curated hard negative examples. At stage-2, it blends various non-retrieval into instruction tuning, which not only enhances non-retrieval task accuracy but also improves retrieval performance. For training data, we utilize the hard-negative mining, synthetic data generation and existing public available datasets to boost the performance of embedding model. By combining these techniques, our NV-Embed- v1 model secured the No.1 position on the Massive Text Embedding Benchmark (MTEB) (as of May 24, 2024), across 56 embedding tasks. NV-Embed-v2 has reclaimed and maintained the top spot on MTEB since August 30, 2024, demonstrating the sustained effectiveness of the proposed methods over time. Also, it achieved the highest scores in the Long Doc section and the second-highest scores in the QA section of the AIR Benchmark, which covers a range of out-of-domain information retrieval topics beyond those in MTEB. We further provide the analysis of model compression techniques for generalist embedding models. We open-source the model at: https://huggingface.co/nvidia/NV-Embed-v2 .

State-of-the-art retrieval models typically address a straightforward search scenario, in which retrieval tasks are fixed (e.g., finding a passage to answer a specific question) and only a single modality is supported for both queries and retrieved results. This paper introduces techniques for advancing information retrieval with multimodal large language models (MLLMs), enabling a broader search scenario, termed universal multimodal retrieval, where multiple modalities and diverse retrieval tasks are accommodated. To this end, we first study fine-tuning an MLLM as a bi-encoder retriever on 10 datasets with 16 retrieval tasks. Our empirical results show that the fine-tuned MLLM retriever is capable of understanding challenging queries, composed of both text and image, but it underperforms compared to a smaller CLIP retriever in cross-modal retrieval tasks due to the modality bias exhibited by MLLMs. To address the issue, we propose modality-aware hard negative mining to mitigate the modality bias exhibited by MLLM retrievers. Second, we propose continuously fine-tuning the universal multimodal retriever to enhance its text retrieval capability while preserving multimodal retrieval capability. As a result, our model, MM-Embed, achieves state-of-the-art performance on the multimodal retrieval benchmark M-BEIR, which spans multiple domains and tasks, while also surpassing the state-of-the-art text retrieval model, NV-Embed-v1, on the MTEB retrieval benchmark. Finally, we explore prompting the off-the-shelf MLLMs as zero-shot rerankers to refine the ranking of the candidates from the multimodal retriever. We find that, through prompt-and-reranking, MLLMs can further improve multimodal retrieval when the user queries (e.g., text-image composed queries) are more complex and challenging to understand. These findings also pave the way for advancing universal multimodal retrieval in the future. We release the model weights at: https://huggingface.co/nvidia/MM-Embed.

Fugatto is a versatile audio synthesis and transformation model capable of following free-form text instructions with optional audio inputs. While large language models (LLMs) trained with text on a simple next-token prediction objective can learn to infer instructions directly from the data, models trained solely on audio data lack this capacity. This is because audio data does not inherently contain the instructions that were used to generate it. To overcome this challenge, we introduce a specialized dataset generation approach optimized for producing a wide range of audio generation and transformation tasks, ensuring the data reveals meaningful relationships between audio and language. Another challenge lies in achieving compositional abilities -- such as combining, interpolating between, or negating instructions -- using data alone. To address it, we propose ComposableART, an inference-time technique that extends classifier-free guidance to compositional guidance. It enables the seamless and flexible composition of instructions, leading to highly customizable audio outputs outside the training distribution. Our evaluations across a diverse set of tasks demonstrate that Fugatto performs competitively with specialized models, while ComposableART enhances its sonic palette and control over synthesis. Most notably, we highlight our framework's ability to execute emergent sounds and tasks -- sonic phenomena that transcend conventional audio generation -- unlocking new creative possibilities. \href{https://fugatto.github.io/}{Demo Website.}

In this work, we introduce ChatQA 2, an Llama 3.0-based model with a 128K context window, designed to bridge the gap between open-source LLMs and leading proprietary models (e.g., GPT-4-Turbo-2024-04-09) in long context un- derstanding and retrieval-augmented generation (RAG) capabilities. These two capabilities are complementary to each other and essential for LLMs to process large volumes of information that cannot fit into a single prompt. We present a detailed continued training recipe to extend the context window of Llama3- 70B-base from 8K to 128K tokens, along with a three-stage instruction tun- ing process to enhance the model’s instruction-following, RAG performance, and long-context understanding capabilities. Our results demonstrate that the Llama3-ChatQA-2-70B model outperforms most existing state-of-the-art models, including GPT-4-Turbo-2024-04-09, Qwen2-72B-Instruct, and Llama3.1-70B- Instruct, on ultra-long tasks beyond 100K tokens, as well as on the RAG benchmark using only a 4K context window, showing the strong long context capability across varying sequence lengths. We further provide extensive comparisons between direct long-context and RAG solutions using the same state-of-the-art long-context LLMs. Interestingly, we find that the performance of strong long-context LLMs using RAG improves when retrieving a larger number of chunks. With a large set of top-k chunks, RAG consistently outperforms direct long-context solution using the same state-of-the-art long-context models (e.g., Llama3-ChatQA-2-70B and Qwen2-72B-Instruct) on both 32K and 128K benchmarks. We open-source the model weights, training data, and the evaluation setup for the for the community: https://chatqa2-project.github.io/

Understanding and reasoning over non-speech sounds and music are crucial for both humans and AI agents to interact effectively with their environments. In this paper, we introduce Audio Flamingo 2 (AF2), an Audio-Language Model (ALM) with advanced audio understanding and reasoning capabilities. AF2 leverages (i) a custom CLAP model, (ii) synthetic Audio QA data for fine-grained audio reasoning, and (iii) a multi-stage curriculum learning strategy. AF2 achieves state-of-the-art performance with only a 3B parameter small language model, surpassing large open-source and proprietary models across over 20 benchmarks. Next, for the first time, we extend audio understanding to long audio segments (30 secs to 5 mins) and propose LongAudio, a large and novel dataset for training ALMs on long audio captioning and question-answering tasks. Fine-tuning AF2 on LongAudio leads to exceptional performance on our proposed LongAudioBench, an expert-annotated benchmark for evaluating ALMs on long audio understanding capabilities. We conduct extensive ablation studies to confirm the efficacy of our approach.

Despite recent progress in large-scale reinforcement learning (RL) for reasoning, the training recipe for building high-performing reasoning models remains elusive. Key implementation details of frontier models, such as DeepSeek-R1, including data curation strategies and RL training recipe, are often omitted. Moreover, recent research indicates distillation remains more effective than RL for smaller models. In this work, we demonstrate that large-scale RL can significantly enhance the reasoning capabilities of strong, small- and mid-sized models, achieving results that surpass those of state-of-the-art distillation-based models. We systematically study the RL training process through extensive ablations and propose a simple yet effective approach: first training on math-only prompts, then on code-only prompts. Notably, we find that math-only RL not only significantly enhances the performance of strong distilled models on math benchmarks (e.g., +14.6\% / +17.2\% on AIME 2025 for the 7B / 14B models), but also code reasoning tasks (e.g., +6.8\% / +5.8\% on LiveCodeBench for the 7B / 14B models). In addition, extended code-only RL further improves code benchmark performance while causing minimal degradation in math results. We develop a robust data curation pipeline to collect challenging prompts with high-quality, verifiable answers and test cases to enable verification-based RL across both domains. The dataset will be released to support open research. Finally, we identify key experimental insights, including curriculum learning with progressively increasing response lengths and the stabilizing effect of on-policy parameter updates. We find that RL not only elicits the foundational reasoning capabilities acquired during pretraining and supervised fine-tuning (SFT), but also pushes the limits of the model’s reasoning ability, enabling it to solve problems that were previously unsolvable.

The Internet of Medical Things (IoMT) has transformed healthcare by integrating medical devices, wearables, and electronic medical records (EMRs). These technologies enable real-time diagnostics, personalized monitoring, and proactive healthcare delivery. However, applying the IoMT to critical areas like lung cancer detection remains challenging due to multimodal data heterogeneity, real-time processing requirements, and resource constraints nature of the devices. To address these challenges, in this paper, we present a novel IoMT-based framework for lung cancer diagnostics that unifies imaging data, sensor readings, and EMRs into a cohesive pipeline. The proposed framework employs convolutional neural networks, vision transformers for imaging analysis, temporal convolutional networks for time-series sensor data, and an attention-based fusion mechanism for dynamic multimodal integration. These techniques are supported by preprocessing methods such as U-Net++ segmentation and temporal feature extraction to enhance data consistency and efficiency. Additionally, lightweight models are deployed on IoMT devices to ensure scalability and real-time inference, making the framework practical for resource-constrained environments. Extensive evaluation demonstrates that the framework achieves 98% accuracy, 98% F1-score, and 0.99 AUC (area under the curve) of the ROC (receiver operating characteristic (AUC-ROC). These results show the proposed framework outperforms state-of-the-art approaches and showcases its potential for large-scale clinical deployment.

Visual language models (VLMs) rapidly progressed with the recent success of large language models. There have been growing efforts on visual instruction tuning to extend the LLM with visual inputs, but lacks an in-depth study of the visual language pre-training process, where the model learns to perform joint modeling on both modalities. In this work, we examine the design options for VLM pre-training by augmenting LLM towards VLM through step-by-step controllable comparisons. We introduce three main findings: (1) freezing LLMs during pre-training can achieve decent zero-shot performance, but lack in-context learning capability, which requires unfreezing the LLM; (2) interleaved pre-training data is beneficial whereas image-text pairs alone are not optimal; (3) re-blending text-only instruction data to image-text data during instruction fine-tuning not only remedies the degradation of text-only tasks, but also boosts VLM task accuracy. With an enhanced pre-training recipe we build VILA, a Visual Language model family that consistently outperforms the state-of-the-art models, e.g., LLaVA-1.5, across main benchmarks without bells and whistles. Multi-modal pre-training also helps unveil appealing properties of VILA, including multi-image reasoning, enhanced in-context learning, and better world knowledge. VILA is also deployable on Jetson Orin for on-device VLM.

Extending the context window of large language models (LLMs) is getting popular recently, while the solution of augmenting LLMs with retrieval has existed for years. The natural questions are: i) Retrieval-augmentation versus long context window, which one is better for downstream tasks? ii) Can both methods be combined to get the best of both worlds? In this work, we answer these questions by studying both solutions using two state-of-the-art pretrained LLMs, i.e., a proprietary 43B GPT and Llama2-70B. Perhaps surprisingly, we find that LLM with 4K context window using simple retrieval-augmentation at generation can achieve comparable performance to finetuned LLM with 16K context window via positional interpolation on long context tasks, while taking much less computation. More importantly, we demonstrate that retrieval can significantly improve the performance of LLMs regardless of their extended context window sizes. Our best model, retrieval-augmented Llama2-70B with 32K context window, outperforms GPT-3.5-turbo-16k and Davinci003 in terms of average score on nine long context tasks including question answering, query-based summarization, and in-context few-shot learning tasks. It also outperforms its non-retrieval Llama2-70B-32k baseline by a margin, while being much faster at generation. Our study provides general insights on the choice of retrieval-augmentation versus long context extension of LLM for practitioners.

Augmenting large language models (LLMs) to understand audio – including non-speech sounds and non-verbal speech – is critically important for diverse real-world applications of LLMs. In this paper, we propose Audio Flamingo, a novel audio language model with 1) strong audio understanding abilities, 2) the ability to quickly adapt to unseen tasks via in-context learning and retrieval, and 3) strong multi-turn dialogue abilities. We introduce a series of training techniques, architecture design, and data strategies to enhance our model with these abilities. Extensive evaluations across various audio understanding tasks confirm the efficacy of our method, setting new state-of-the-art benchmarks. Our demo website is https://audioflamingo.github.io/ and the code is open-sourced at https://github.com/NVIDIA/audio-flamingo.

Pretraining auto-regressive large language models (LLMs) with retrieval demonstrates better perplexity and factual accuracy by leveraging external databases. However, the size of existing pretrained retrieval-augmented LLM is still limited (e.g., Retro has 7.5B parameters), which limits the effectiveness of instruction tuning and zero-shot generalization. In this work, we introduce Retro 48B, the largest LLM pretrained with retrieval. Specifically, we continue to pretrain a 43B GPT model on additional 100 billion tokens using the Retro augmentation method by retrieving from 1.2 trillion tokens. Notably, the obtained foundation model, Retro 48B, largely outperforms the counterpart GPT 43B trained on 1.2T tokens in terms of perplexity with only 2.58% additional GPU hours, demonstrating the significant scaling potential of the method. After instruction tuning on Retro, InstructRetro demonstrates significant improvement over the instruction-tuned GPT on a wide range of zero-shot tasks. Specifically, the average improvement of InstructRetro is 7% over its GPT counterpart across 8 short-form QA and reading comprehension tasks, 10% over GPT across 4 challenging long-form QA tasks, and 16% over GPT across 3 summarization tasks. Surprisingly, we find that one can ablate the encoder from InstructRetro architecture and directly use its decoder backbone, while achieving comparable results. Our results highlight the promising direction to obtain a better GPT decoder through continued pretraining with retrieval before instruction tuning. Our code and checkpoints are publicly available at: https://huggingface.co/nvidia/retro-48b-instruct-4k.

In this work, we introduce ChatQA, a suite of models that outperform GPT-4 on retrieval-augmented generation (RAG) and conversational question answering (QA). To enhance generation, we propose a two-stage instruction tuning method that significantly boosts the performance of RAG. For effective retrieval, we introduce a dense retriever optimized for conversational QA, which yields results comparable to the alternative state-of-the-art query rewriting models, while substantially reducing deployment costs. We also present the ChatRAG Bench, which encompasses ten datasets covering comprehensive evaluations on RAG, table-related QA, arithmetic calculations, and scenarios involving unanswerable questions. Our ChatQA-1.0-70B (score: 54.14), built on Llama2, a weaker foundation model than GPT-4, can slightly outperform GPT-4-0613 (score: 53.90) and GPT-4-Turbo-2024-04-09 (score: 54.03) on the ChatRAG Bench, without relying on any synthetic data from OpenAI GPT models. Notably, Llama3-ChatQA-1.5-70B model surpasses the accuracy of GPT-4-Turbo-2024-04-09 by a margin. These results demonstrate the exceptional quality of the proposed ChatQA recipe. To advance research in this field, we open-sourced the model weights, instruction tuning data, ChatRAG Bench, and retriever for the community.

Large language models (LLMs) typically utilize the top-k contexts from a retriever in retrieval-augmented generation (RAG). In this work, we propose a novel method called RankRAG, which instruction-tunes a single LLM for both context ranking and answer generation in RAG. In particular, the instruction-tuned LLMs work surprisingly well by adding a small fraction of ranking data into the training blend, and outperform existing expert ranking models, including the same LLM exclusively fine-tuned on a large amount of ranking data. For generation, we compare our model with many strong baselines, including ChatQA-1.5, an open-sourced model with the state-of-the-art performance on RAG benchmarks. Specifically, our Llama3-RankRAG-8B and Llama3-RankRAG-70B significantly outperform Llama3-ChatQA-1.5-8B and Llama3-ChatQA-1.5-70B, respectively, on nine general knowledge-intensive benchmarks for RAG. In addition, it also performs comparably to GPT-4 on five RAG benchmarks in the biomedical domain without instruction fine-tuning on biomedical data, demonstrating its superb capability for generalization to new domains.

Generative modeling with ordinary differential equations (ODEs) has achieved fantastic results on a variety of applications. Yet, few works have focused on controlling the generated content of a pre-trained ODE-based generative model. In this paper, we propose to optimize the output of ODE models according to a guidance function to achieve controllable generation. We point out that, the gradients can be efficiently back-propagated from the output to any intermediate time steps on the ODE trajectory, by decomposing the back-propagation and computing vectorJacobian products. To further accelerate the computation of the back-propagation, we propose to use a non-uniform discretization to approximate the ODE trajectory, where we measure how straight the trajectory is and gather the straight parts into one discretization step. This allows us to save ∼ 90% of the back-propagation time with ignorable error. Our framework, named FlowGrad, outperforms the state-of-the-art baselines on text-guided image manipulation. Moreover, FlowGrad enables us to find global semantic directions in frozen ODE-based generative models that can be used to manipulate new images without extra optimization.

Boxin Wang, Wei Ping, Peng Xu, Lawrence McAfee, Zihan Liu, Mohammad Shoeybi, Yi Dong, Oleksii Kuchaiev, Bo Li, Chaowei Xiao, Anima Anandkumar, Bryan Catanzaro. Proceedings of the 2023 Conference on Empirical Methods in Natural Language Processing. 2023.

In this work, we present CleanUNet 2, a speech denoising model that combines the advantages of waveform denoiser and spectrogram denoiser and achieves the best of both worlds. CleanUNet 2 uses a two-stage framework inspired by popular speech synthesis methods that consist of a waveform model and a spectrogram model. Specifically, CleanUNet 2 builds upon CleanUNet, the state-of-the-art waveform denoiser, and further boosts its performance by taking predicted spectrograms from a spectrogram denoiser as the input. We demonstrate that CleanUNet 2 outperforms previous methods in terms of various objective and subjective evaluations.

Peng Xu, Mostofa Patwary, Shrimai Prabhumoye, Virginia Adams, Ryan Prenger, Wei Ping, Nayeon Lee, Mohammad Shoeybi, Bryan Catanzaro. Proceedings of the 2022 Conference on Empirical Methods in Natural Language Processing. 2022.

Speech-to-text alignment is a critical component of neural text-to-speech (TTS) models. Autoregressive TTS models typically use an attention mechanism to learn these alignments on-line. However, these alignments tend to be brittle and often fail to generalize to long utterances and out-of-domain text, leading to missing or repeating words. Most non-autoregressive end-to-end TTS models rely on durations extracted from external sources. In this paper we leverage the alignment mechanism proposed in RAD-TTS and demonstrate its applicability to wide variety of neural TTS models. The alignment learning framework combines the forward-sum algorithm, Viterbi algorithm, and an efficient static prior. In our experiments, the framework improves all tested TTS architectures, both autoregressive (Flowtron, Tacotron 2) and non-autoregressive (FastPitch, FastSpeech 2, RAD-TTS). Specifically, it improves alignment convergence speed, simplifies the training pipeline by eliminating need for external aligners, enhances robustness to errors on long utterances and improves the perceived speech synthesis quality, as judged by human evaluators.

In this work, we present CleanUNet, a causal speech denoising model on the raw waveform. The proposed model is based on an encoder-decoder architecture combined with several self-attention blocks to refine its bottleneck representations, which is crucial to obtain good results. The model is optimized through a set of losses defined over both waveform and multi-resolution spectrograms. The proposed method outperforms the state-of-the-art models in terms of denoised speech quality from various objective and subjective evaluation metrics.1

Pretrained language models (LMs) are susceptible to generate text with nonfactual information. In this work, we measure and improve the factual accuracy of large-scale LMs for open-ended text generation. We design the FactualityPrompts test set and metrics to measure the factuality of LM generations. Based on that, we study the factual accuracy of LMs with parameter sizes ranging from 126M to 530B. Interestingly, we find that larger LMs are more factual than smaller ones, although a previous study suggests that larger LMs can be less truthful in terms of misconceptions. In addition, popular sampling algorithms (e.g., top-p) in open-ended text generation can harm the factuality due to the ``uniform randomness'' introduced at every sampling step. We propose the factual-nucleus sampling algorithm that dynamically adapts the randomness to improve the factuality of generation while maintaining quality. Furthermore, we analyze the inefficiencies of the standard training method in learning correct associations between entities from factual text corpus (e.g., Wikipedia). We propose a factuality-enhanced training method that uses TopicPrefix for better awareness of facts and sentence completion as the training objective, which can vastly reduce the factual errors.

Pre-trained language models (LMs) are shown to easily generate toxic language. In this work, we systematically explore domain-adaptive training to reduce the toxicity of language models. We conduct this study on three dimensions: training corpus, model size, and parameter efficiency. For the training corpus, we demonstrate that using self-generated datasets consistently outperforms the existing baselines across various model sizes on both automatic and human evaluations, even when it uses a 3 1 smaller training corpus. We then comprehensively study detoxifying LMs with parameter sizes ranging from 126M up to 530B (3× larger than GPT3), a scale that has never been studied before. We find that i) large LMs have similar toxicity levels as smaller ones given the same pre-training corpus, and ii) large LMs require more endeavor to unlearn the toxic content seen at pretraining. We also explore parameter-efficient training methods for detoxification. We demonstrate that adding and training adapter-only layers in LMs not only saves a lot of parameters but also achieves a better trade-off between toxicity and perplexity than whole model adaptation for large-scale models. Our code will be available at: https://github.com/NVIDIA/Megatron-LM/.

Devendra Sachan, Mostofa Patwary, Mohammad Shoeybi, Neel Kant, Wei Ping, William L. Hamilton, Bryan Catanzaro. Proceedings of the 59th Annual Meeting of the Association for Computational Linguistics and the 11th International Joint Conference on Natural Language Processing (Volume 1: Long Papers). 2021.

Transformers have achieved success in both language and vision domains. However, it is prohibitively expensive to scale them to long sequences such as long documents or high-resolution images, because self-attention mechanism has quadratic time and memory complexities with respect to the input sequence length. In this paper, we propose Long-Short Transformer (Transformer-LS), an efficient self-attention mechanism for modeling long sequences with linear complexity for both language and vision tasks. It aggregates a novel long-range attention with dynamic projection to model distant correlations and a short-term attention to capture fine-grained local correlations. We propose a dual normalization strategy to account for the scale mismatch between the two attention mechanisms. Transformer-LS can be applied to both autoregressive and bidirectional models without additional complexity. Our method outperforms the state-of-the-art models on multiple tasks in language and vision domains, including the Long Range Arena benchmark, autoregressive language modeling, and ImageNet classification. For instance, Transformer-LS achieves 0.97 test BPC on enwik8 using half the number of parameters than previous method, while being faster and is able to handle 3x as long sequences compared to its full-attention version on the same hardware. On ImageNet, it can obtain the state-of-the-art results (e.g., a moderate size of 55.8M model solely trained on 224x224 ImageNet-1K can obtain Top-1 accuracy 84.1%), while being more scalable on high-resolution images. The source code and models are released at https://github.com/NVIDIA/transformer-ls.

In this work, we propose ParaNet, a non-autoregressive seq2seq model that converts text to spectrogram. It is fully convolutional and brings 46.7 times speed-up over the lightweight Deep Voice 3 at synthesis, while obtaining reasonably good speech quality. ParaNet also produces stable alignment between text and speech on the challenging test sentences by iteratively improving the attention in a layer-by-layer manner. Furthermore, we build the parallel text-to-speech system by applying various parallel neural vocoders, which can synthesize speech from text through a single feed-forward pass. We also explore a novel VAE-based approach to train the inverse autoregressive flow (IAF) based parallel vocoder from scratch, which avoids the need for distillation from a separately trained WaveNet as previous work.

In this work, we propose WaveFlow, a small-footprint generative flow for raw audio, which is directly trained with maximum likelihood. It handles the long-range structure of 1-D waveform with a dilated 2-D convolutional architecture, while modeling the local variations using expressive autoregressive functions. WaveFlow provides a unified view of likelihood-based models for 1-D data, including WaveNet and WaveGlow as special cases. It generates high-fidelity speech as WaveNet, while synthesizing several orders of magnitude faster as it only requires a few sequential steps to generate very long waveforms with hundreds of thousands of time-steps. Furthermore, it can significantly reduce the likelihood gap that has existed between autoregressive models and flow-based models for efficient synthesis. Finally, our small-footprint WaveFlow has only 5.91M parameters, which is 15{\texttimes} smaller than WaveGlow. It can generate 22.05 kHz high-fidelity audio 42.6{\texttimes} faster than real-time (at a rate of 939.3 kHz) on a V100 GPU without engineered inference kernels.

We propose a Topic Compositional Neural Language Model (TCNLM), a novel method designed to simultaneously capture both the global semantic meaning and the local word-ordering structure in a document. The TCNLM learns the global semantic coherence of a document via a neural topic model, and the probability of each learned latent topic is further used to build a Mixture-of-Experts (MoE) language model, where each expert (corresponding to one topic) is a recurrent neural network (RNN) that accounts for learning the local structure of a word sequence. In order to train the MoE model efficiently, a matrix factorization method is applied, by extending each weight matrix of the RNN to be an ensemble of topic-dependent weight matrices. The degree to which each member of the ensemble is used is tied to the document-dependent probability of the corresponding topics. Experimental results on several corpora show that the proposed approach outperforms both a pure RNN-based model and other topic-guided language models. Further, our model yields sensible topics, and also has the capacity to generate meaningful sentences conditioned on given topics.

We propose a large margin criterion for training neural language models.Conventionally, neural language models are trained by minimizing perplexity (PPL) on grammatical sentences.However, we demonstrate that PPL may not be the best metric to optimize in some tasks, and further propose a large margin formulation.The proposed method aims to enlarge the margin between the "good" and "bad" sentences in a task-specific sense.It is trained end-to-end and can be widely applied to tasks that involve re-scoring of generated text.Compared with minimum-PPL training, our method gains up to 1.1 WER reduction for speech recognition and 1.0 BLEU increase for machine translation.

Voice cloning is a highly desired feature for personalized speech interfaces. We introduce a neural voice cloning system that learns to synthesize a person's voice from only a few audio samples. We study two approaches: speaker adaptation and speaker encoding. Speaker adaptation is based on fine-tuning a multi-speaker generative model. Speaker encoding is based on training a separate model to directly infer a new speaker embedding, which will be applied to a multi-speaker generative model. In terms of naturalness of the speech and similarity to the original speaker, both approaches can achieve good performance, even with a few cloning audios. While speaker adaptation can achieve slightly better naturalness and similarity, cloning time and required memory for the speaker encoding approach are significantly less, making it more favorable for low-resource deployment.

Restricted Boltzmann machines (RBMs) and conditional RBMs (CRBMs) are popular models for a wide range of applications. In previous work, learning on such models has been dominated by contrastive divergence (CD) and its variants. Belief propagation (BP) algorithms are believed to be slow for structured prediction on conditional RBMs (e.g., Mnih et al. [2011]), and not as good as CD when applied in learning (e.g., Larochelle et al. [2012]). In this work, we present a matrix-based implementation of belief propagation algorithms on CRBMs, which is easily scalable to tens of thousands of visible and hidden units. We demonstrate that, in both maximum likelihood and max-margin learning, training conditional RBMs with BP as the inference routine can provide significantly better results than current state-of-the-art CD methods on structured prediction problems. We also include practical guidelines on training CRBMs with BP, and some insights on the interaction of learning and inference algorithms for CRBMs.

We introduce a technique for augmenting neural text-to-speech (TTS) with low-dimensional trainable speaker embeddings to generate different voices from a single model. As a starting point, we show improvements over the two state-of-the-art approaches for single-speaker neural TTS: Deep Voice 1 and Tacotron. We introduce Deep Voice 2, which is based on a similar pipeline with Deep Voice 1, but constructed with higher performance building blocks and demonstrates a significant audio quality improvement over Deep Voice 1. We improve Tacotron by introducing a post-processing neural vocoder, and demonstrate a significant audio quality improvement. We then demonstrate our technique for multi-speaker speech synthesis for both Deep Voice 2 and Tacotron on two multi-speaker TTS datasets. We show that a single neural TTS system can learn hundreds of unique voices from less than half an hour of data per speaker, while achieving high audio quality synthesis and preserving the speaker identities almost perfectly.

In this work, we propose an infinite restricted Boltzmann machine (RBM), whose maximum likelihood estimation (MLE) corresponds to a constrained convex optimization. We consider the Frank-Wolfe algorithm to solve the program, which provides a sparse solution that can be interpreted as inserting a hidden unit at each iteration, so that the optimization process takes the form of a sequence of finite models of increasing complexity. As a side benefit, this can be used to easily and efficiently identify an appropriate number of hidden units during the optimization. The resulting model can also be used as an initialization for typical state-of-the-art RBM training algorithms such as contrastive divergence, leading to models with consistently higher test likelihood than random initialization.

Marginal MAP inference involves making MAP predictions in systems defined with latent variables or missing information. It is significantly more difficult than pure marginalization and MAP tasks, for which a large class of efficient and convergent variational algorithms, such as dual decomposition, exist. In this work, we generalize dual decomposition to a generic powered-sum inference task, which includes marginal MAP, along with pure marginalization and MAP, as special cases. Our method is based on a block coordinate descent algorithm on a new convex decomposition bound, that is guaranteed to converge monotonically, and can be parallelized efficiently. We demonstrate our approach on various inference queries over real-world problems from the UAI approximate inference challenge, showing that our framework is faster and more reliable than previous methods.

In this work, we propose the marginal structured SVM (MSSVM) for structured prediction with hidden variables. MSSVM properly accounts for the uncertainty of hidden variables, and can significantly outperform the previously proposed latent structured SVM (LSSVM; Yu & Joachims (2009)) and other state-of-art methods, especially when that uncertainty is large. Our method also results in a smoother objective function, making gradient-based optimization of MSSVMs converge significantly faster than for LSSVMs. We also show that our method consistently outperforms hidden conditional random fields (HCRFs; Quattoni et al. (2007)) on both simulated and real-world datasets. Furthermore, we propose a unified framework that includes both our and several other existing methods as special cases, and provides insights into the comparison of different models in practice.

Feature selection is an effective tool to deal with the "curse of dimensionality". To cope with the non-separable problem, feature selection in the kernel space has been investigated. However, previous study cannot adequately estimate the intrinsic dimensionality of the kernel space. Thus, it is difficult to accurately preserve the sketch of the kernel space using the learned basis, and the feature selection performance is affected. Moreover, the computing load of the algorithm reaches at least cubic with the number of training data. In this paper, we propose a fast framework to conduct feature selection in the kernel space. By designing a fast kernel subspace learning method, we automatically learn the intrinsic dimensionality and construct an orthogonal basis set of kernel space. The learned basis can accurately preserve the sketch of kernel space. Then backed by the constructed basis, we directly select features in kernel space. The whole proposed framework has a quadratic complexity with the number of training data, which is faster than existing kernel methods for feature selection. We evaluate our work under several typical datasets and find it not only preserves the sketch of the kernel space more accurately but also achieves better classification performance compared with many state-of-the-art methods.

Multi-instance learning, like other machine learning and data mining tasks, requires distance metrics. Although metric learning methods have been studied for many years, metric learners for multi-instance learning remain almost untouched. In this paper, we propose a framework called Multi-Instance MEtric Learning (MIMEL) to learn an appropriate distance under the multi-instance setting. The distance metric between two bags is defined using the Mahalanobis distance function. The problem is formulated by minimizing the KL divergence between two multivariate Gaussians under the constraints of maximizing the between-class bag distance and minimizing the within-class bag distance. To exploit the mechanism of how instances determine bag labels in multi-instance learning, we design a nonparametric density-estimation-based weighting scheme to assign higher "weights" to the instances that are more likely to be positive in positive bags. The weighting scheme itself has a small workload, which adds little extra computing costs to the proposed framework. Moreover, to further boost the classification accuracy, a kernel version of MIMEL is presented. We evaluate MIMEL, using not only several typical multi-instance tasks, but also two activity recognition datasets. The experimental results demonstrate that MIMEL achieves better classification accuracy than many state-of-the-art distance based algorithms or kernel methods for multi-instance learning.

In this paper, we present a novel near-duplicate video retrieval system serving one million web videos. To achieve both the effectiveness and efficiency, a visual word based approach is proposed, which quantizes each video frame into a word and represents the whole video as a bag of words. The system can respond to a query in 41ms with 78.4% MAP on average.

Kernel method is a powerful tool in multi-instance learning. However, many typical kernel methods for multi-instance learning ignore the correspondence information of instances between two bags or co-occurrence information, and result in poor performance. Additionally, most current multi-instance kernels unreasonably assign all instances in each bag an equal weight, which neglects the significance of some “key” instances in multi-instance learning. Last but not least, almost all the multi-instance kernels encounter a heavy computation load, which may fail in large datasets. To cope with these shortcomings, we propose a FAst kernel for Multi-instancE leaRning named as FAMER. FAMER constructs a Locally Sensitive Hashing (LSH) based similarity measure for multi-instance framework, and represents each bag as a histogram by embedding instances within the bag into an auxiliary space, which captures the correspondence information between two bags. By designing a bin-dependent weighting scheme, we not only impose different weights on instances according to their discriminative powers, but also exploit co-occurrence relations according to the joint statistics of instances. Without directly computing in a pairwise manner, the time complexity of FAMER is much smaller compared to other typical multi-instance kernels. The experiments demonstrate the effectiveness and efficiency of the proposed method.

Multi-instance learning, as other machine learning tasks, also suffers from the curse of dimensionality. Although dimensionality reduction methods have been investigated for many years, multi-instance dimensionality reduction methods remain untouched. On the other hand, most algorithms in multi- instance framework treat instances in each bag as independently and identically distributed samples, which fails to utilize the structure information conveyed by instances in a bag. In this paper, we propose a multi-instance dimensionality reduction method, which treats instances in each bag as non-i.i.d. samples. We regard every bag as a whole entity and define a bag margin objective function. By maximizing the margin of positive and negative bags, we learn a subspace to obtain more salient representation of original data. Experiments demonstrate the effectiveness of the proposed method.