Text-to-speech(TTS) has undergone remarkable improvements in performance, particularly with the advent of Denoising Diffusion Probabilistic Models (DDPMs). However, the perceived quality of audio depends not solely on its content, pitch, rhythm, and energy, but also on the physical environment. In this work, we propose ViT-TTS, the first visual TTS model with scalable diffusion transformers. ViT-TTS complement the phoneme sequence with the visual information to generate high-perceived audio, opening up new avenues for practical applications of AR and VR to allow a more immersive and realistic audio experience. To mitigate the data scarcity in learning visual acoustic information, we 1) introduce a self-supervised learning framework to enhance both the visual-text encoder and denoiser decoder; 2) leverage the diffusion transformer scalable in terms of parameters and capacity to learn visual scene information. Experimental results demonstrate that ViT-TTS achieves new state-of-the-art results, outperforming cascaded systems and other baselines regardless of the visibility of the scene. With low-resource data (1h, 2h, 5h), ViT-TTS achieves comparative results with rich-resource baselines.~\footnote{Audio samples are available at \url{https://ViT-TTS.github.io/.}}
Voice conversion is an increasingly popular technology, and the growing number of real-time applications requires models with streaming conversion capabilities. Unlike typical (non-streaming) voice conversion, which can leverage the entire utterance as full context, streaming voice conversion faces significant challenges due to the missing future information, resulting in degraded intelligibility, speaker similarity, and sound quality. To address this challenge, we propose DualVC, a dual-mode neural voice conversion approach that supports both streaming and non-streaming modes using jointly trained separate network parameters. Furthermore, we propose intra-model knowledge distillation and hybrid predictive coding (HPC) to enhance the performance of streaming conversion. Additionally, we incorporate data augmentation to train a noise-robust autoregressive decoder, improving the model's performance on long-form speech conversion. Experimental results demonstrate that the proposed model outperforms the baseline models in the context of streaming voice conversion, while maintaining comparable performance to the non-streaming topline system that leverages the complete context, albeit with a latency of only 252.8 ms.
Pre-trained seq2seq models have achieved state-of-the-art results in the grammatical error correction task. However, these models still suffer from a prediction bias due to their unidirectional decoding. Thus, we propose a bidirectional Transformer reranker (BTR), that re-estimates the probability of each candidate sentence generated by the pre-trained seq2seq model. The BTR preserves the seq2seq-style Transformer architecture but utilizes a BERT-style self-attention mechanism in the decoder to compute the probability of each target token by using masked language modeling to capture bidirectional representations from the target context. For guiding the reranking, the BTR adopts negative sampling in the objective function to minimize the unlikelihood. During inference, the BTR gives final results after comparing the reranked top-1 results with the original ones by an acceptance threshold. Experimental results show that, in reranking candidates from a pre-trained seq2seq model, T5-base, the BTR on top of T5-base could yield 65.47 and 71.27 F0.5 scores on the CoNLL-14 and BEA test sets, respectively, and yield 59.52 GLEU score on the JFLEG corpus, with improvements of 0.36, 0.76 and 0.48 points compared with the original T5-base. Furthermore, when reranking candidates from T5-large, the BTR on top of T5-base improved the original T5-large by 0.26 points on the BEA test set.
Multi-aspect controllable text generation aims to generate fluent sentences that possess multiple desired attributes simultaneously. Traditional methods either combine many operators in the decoding stage, often with costly iteration or search in the discrete text space, or train separate controllers for each aspect, resulting in a degeneration of text quality due to the discrepancy between different aspects. To address these limitations, we introduce a novel approach for multi-aspect control, namely MacLaSa, that estimates compact latent space for multiple aspects and performs efficient sampling with a robust sampler based on ordinary differential equations (ODEs). To eliminate the domain gaps between different aspects, we utilize a Variational Autoencoder (VAE) network to map text sequences from varying data sources into close latent representations. The estimated latent space enables the formulation of joint energy-based models (EBMs) and the plugging in of arbitrary attribute discriminators to achieve multi-aspect control. Afterwards, we draw latent vector samples with an ODE-based sampler and feed sampled examples to the VAE decoder to produce target text sequences. Experimental results demonstrate that MacLaSa outperforms several strong baselines on attribute relevance and textual quality while maintaining a high inference speed.
Transformer-based encoder-decoder models that generate outputs in a left-to-right fashion have become standard for sequence-to-sequence tasks. In this paper, we propose a framework for decoding that produces sequences from the "outside-in": at each step, the model chooses to generate a token on the left, on the right, or join the left and right sequences. We argue that this is more principled than prior bidirectional decoders. Our proposal supports a variety of model architectures and includes several training methods, such as a dynamic programming algorithm that marginalizes out the latent ordering variable. Our model sets state-of-the-art (SOTA) on the 2022 and 2023 shared tasks, beating the next best systems by over 4.7 and 2.7 points in average accuracy respectively. The model performs particularly well on long sequences, can implicitly learn the split point of words composed of stem and affix, and performs better relative to the baseline on datasets that have fewer unique lemmas (but more examples per lemma).
Finetuning pretrained language models (LMs) have enabled appealing performance on a diverse array of tasks. The intriguing task-agnostic property has driven a shifted focus from task-specific to task-agnostic distillation of LMs. While task-agnostic, compute-efficient, performance-preserved LMs can be yielded by task-agnostic distillation, previous studies mainly sit in distillation of either encoder-only LMs (e.g., BERT) or decoder-only ones (e.g., GPT) yet largely neglect that distillation of encoder-decoder LMs (e.g., T5) can posit very distinguished behaviors. Frustratingly, we discover that existing task-agnostic distillation methods can fail to handle the distillation of encoder-decoder LMs. To the demand, we explore a few paths and uncover a path named as MiniEnD that successfully tackles the distillation of encoder-decoder LMs in a task-agnostic fashion. We examine MiniEnD on language understanding and abstractive summarization. The results showcase that MiniEnD is generally effective and is competitive compared to other alternatives. We further scale MiniEnD up to distillation of 3B encoder-decoder language models with interpolated distillation. The results imply the opportunities and challenges in distilling large language models (e.g., LLaMA).
Recent studies have shown that sequence-to-sequence (seq2seq) models struggle with compositional generalization (CG), i.e., the ability to systematically generalize to unseen compositions of seen components. There is mounting evidence that one of the reasons hindering CG is the representation of the encoder uppermost layer is entangled, i.e., the syntactic and semantic representations of sequences are entangled. However, we consider that the previously identified representation entanglement problem is not comprehensive enough. Additionally, we hypothesize that the source keys and values representations passing into different decoder layers are also entangled. Starting from this intuition, we propose \textsc{CompoSition} (\textbf{Compo}se \textbf{S}yntactic and Semant\textbf{i}c Representa\textbf{tion}s), an extension to seq2seq models which learns to compose representations of different encoder layers dynamically for different tasks, since recent studies reveal that the bottom layers of the Transformer encoder contain more syntactic information and the top ones contain more semantic information. Specifically, we introduce a \textit{composed layer} between the encoder and decoder to compose different encoder layers' representations to generate specific keys and values passing into different decoder layers. \textsc{CompoSition} achieves competitive results on two comprehensive and realistic benchmarks, which empirically demonstrates the effectiveness of our proposal. Codes are available at~\url{https://github.com/thinkaboutzero/COMPOSITION}.
To translate speech for automatic dubbing, machine translation needs to be isochronous, i.e. translated speech needs to be aligned with the source in terms of speech durations. We introduce target factors in a transformer model to predict durations jointly with target language phoneme sequences. We also introduce auxiliary counters to help the decoder to keep track of the timing information while generating target phonemes. We show that our model improves translation quality and isochrony compared to previous work where the translation model is instead trained to predict interleaved sequences of phonemes and durations.
Recently, various studies have been directed towards exploring dense passage retrieval techniques employing pre-trained language models, among which the masked auto-encoder (MAE) pre-training architecture has emerged as the most promising. The conventional MAE framework relies on leveraging the passage reconstruction of decoder to bolster the text representation ability of encoder, thereby enhancing the performance of resulting dense retrieval systems. Within the context of building the representation ability of the encoder through passage reconstruction of decoder, it is reasonable to postulate that a ``more demanding'' decoder will necessitate a corresponding increase in the encoder's ability. To this end, we propose a novel token importance aware masking strategy based on pointwise mutual information to intensify the challenge of the decoder. Importantly, our approach can be implemented in an unsupervised manner, without adding additional expenses to the pre-training phase. Our experiments verify that the proposed method is both effective and robust on large-scale supervised passage retrieval datasets and out-of-domain zero-shot retrieval benchmarks.
Multi-query attention (MQA), which only uses a single key-value head, drastically speeds up decoder inference. However, MQA can lead to quality degradation, and moreover it may not be desirable to train a separate model just for faster inference. We (1) propose a recipe for uptraining existing multi-head language model checkpoints into models with MQA using 5% of original pre-training compute, and (2) introduce grouped-query attention (GQA), a generalization of multi-query attention which uses an intermediate (more than one, less than number of query heads) number of key-value heads. We show that uptrained GQA achieves quality close to multi-head attention with comparable speed to MQA.
Pretraining is the preliminary and fundamental step in developing capable language models (LM). Despite this, pretraining data design is critically under-documented and often guided by empirically unsupported intuitions. To address this, we pretrain 28 1.5B parameter decoder-only models, training on data curated (1) at different times, (2) with varying toxicity and quality filters, and (3) with different domain compositions. First, we quantify the effect of pretraining data age. A temporal shift between evaluation data and pretraining data leads to performance degradation, which is not overcome by finetuning. Second, we explore the effect of quality and toxicity filters, showing a trade-off between performance on standard benchmarks and risk of toxic generations. Our findings indicate there does not exist a one-size-fits-all solution to filtering training data. We also find that the effects of different types of filtering are not predictable from text domain characteristics. Lastly, we empirically validate that the inclusion of heterogeneous data sources, like books and web, is broadly beneficial and warrants greater prioritization. These findings constitute the largest set of experiments to validate, quantify, and expose many undocumented intuitions about text pretraining, which we hope will help support more informed data-centric decisions in LM development.
With the exponential growth of video data, there is an urgent need for automated technology to analyze and comprehend video content. However, existing video understanding models are often task-specific and lack a comprehensive capability of handling diverse tasks. The success of large language models (LLMs) like GPT has demonstrated their impressive abilities in sequence causal reasoning. Building upon this insight, we propose a novel framework called VideoLLM that leverages the sequence reasoning capabilities of pre-trained LLMs from natural language processing (NLP) for video sequence understanding. VideoLLM incorporates a carefully designed Modality Encoder and Semantic Translator, which convert inputs from various modalities into a unified token sequence. This token sequence is then fed into a decoder-only LLM. Subsequently, with the aid of a simple task head, our VideoLLM yields an effective unified framework for different kinds of video understanding tasks. To evaluate the efficacy of VideoLLM, we conduct extensive experiments using multiple LLMs and fine-tuning methods. We evaluate our VideoLLM on eight tasks sourced from four different datasets. The experimental results demonstrate that the understanding and reasoning capabilities of LLMs can be effectively transferred to video understanding tasks. We release the code at https://github.com/cg1177/VideoLLM.
Satellite image time series in the optical and infrared spectrum suffer from frequent data gaps due to cloud cover, cloud shadows, and temporary sensor outages. It has been a long-standing problem of remote sensing research how to best reconstruct the missing pixel values and obtain complete, cloud-free image sequences. We approach that problem from the perspective of representation learning and develop U-TILISE, an efficient neural model that is able to implicitly capture spatio-temporal patterns of the spectral intensities, and that can therefore be trained to map a cloud-masked input sequence to a cloud-free output sequence. The model consists of a convolutional spatial encoder that maps each individual frame of the input sequence to a latent encoding; an attention-based temporal encoder that captures dependencies between those per-frame encodings and lets them exchange information along the time dimension; and a convolutional spatial decoder that decodes the latent embeddings back into multi-spectral images. We experimentally evaluate the proposed model on EarthNet2021, a dataset of Sentinel-2 time series acquired all over Europe, and demonstrate its superior ability to reconstruct the missing pixels. Compared to a standard interpolation baseline, it increases the PSNR by 1.8 dB at previously seen locations and by 1.3 dB at unseen locations.
We present a new self-supervised paradigm on point cloud sequence understanding. Inspired by the discriminative and generative self-supervised methods, we design two tasks, namely point cloud sequence based Contrastive Prediction and Reconstruction (CPR), to collaboratively learn more comprehensive spatiotemporal representations. Specifically, dense point cloud segments are first input into an encoder to extract embeddings. All but the last ones are then aggregated by a context-aware autoregressor to make predictions for the last target segment. Towards the goal of modeling multi-granularity structures, local and global contrastive learning are performed between predictions and targets. To further improve the generalization of representations, the predictions are also utilized to reconstruct raw point cloud sequences by a decoder, where point cloud colorization is employed to discriminate against different frames. By combining classic contrast and reconstruction paradigms, it makes the learned representations with both global discrimination and local perception. We conduct experiments on four point cloud sequence benchmarks, and report the results on action recognition and gesture recognition under multiple experimental settings. The performances are comparable with supervised methods and show powerful transferability.
This paper studies the computational offloading of video action recognition in edge computing. To achieve effective semantic information extraction and compression, following semantic communication we propose a novel spatiotemporal attention-based autoencoder (STAE) architecture, including a frame attention module and a spatial attention module, to evaluate the importance of frames and pixels in each frame. Additionally, we use entropy encoding to remove statistical redundancy in the compressed data to further reduce communication overhead. At the receiver, we develop a lightweight decoder that leverages a 3D-2D CNN combined architecture to reconstruct missing information by simultaneously learning temporal and spatial information from the received data to improve accuracy. To fasten convergence, we use a step-by-step approach to train the resulting STAE-based vision transformer (ViT_STAE) models. Experimental results show that ViT_STAE can compress the video dataset HMDB51 by 104x with only 5% accuracy loss, outperforming the state-of-the-art baseline DeepISC. The proposed ViT_STAE achieves faster inference and higher accuracy than the DeepISC-based ViT model under time-varying wireless channel, which highlights the effectiveness of STAE in guaranteeing higher accuracy under time constraints.
We present FIT: a transformer-based architecture with efficient self-attention and adaptive computation. Unlike original transformers, which operate on a single sequence of data tokens, we divide the data tokens into groups, with each group being a shorter sequence of tokens. We employ two types of transformer layers: local layers operate on data tokens within each group, while global layers operate on a smaller set of introduced latent tokens. These layers, comprising the same set of self-attention and feed-forward layers as standard transformers, are interleaved, and cross-attention is used to facilitate information exchange between data and latent tokens within the same group. The attention complexity is $O(n^2)$ locally within each group of size $n$, but can reach $O(L^{{4}/{3}})$ globally for sequence length of $L$. The efficiency can be further enhanced by relying more on global layers that perform adaptive computation using a smaller set of latent tokens. FIT is a versatile architecture and can function as an encoder, diffusion decoder, or autoregressive decoder. We provide initial evidence demonstrating its effectiveness in high-resolution image understanding and generation tasks. Notably, FIT exhibits potential in performing end-to-end training on gigabit-scale data, such as 6400$\times$6400 images, or 160K tokens (after patch tokenization), within a memory capacity of 16GB, without requiring specific optimizations or model parallelism.
In real-time video communication, retransmitting lost packets over high-latency networks is not viable due to strict latency requirements. To counter packet losses without retransmission, two primary strategies are employed -- encoder-based forward error correction (FEC) and decoder-based error concealment. The former encodes data with redundancy before transmission, yet determining the optimal redundancy level in advance proves challenging. The latter reconstructs video from partially received frames, but dividing a frame into independently coded partitions inherently compromises compression efficiency, and the lost information cannot be effectively recovered by the decoder without adapting the encoder. We present a loss-resilient real-time video system called GRACE, which preserves the user's quality of experience (QoE) across a wide range of packet losses through a new neural video codec. Central to GRACE's enhanced loss resilience is its joint training of the neural encoder and decoder under a spectrum of simulated packet losses. In lossless scenarios, GRACE achieves video quality on par with conventional codecs (e.g., H.265). As the loss rate escalates, GRACE exhibits a more graceful, less pronounced decline in quality, consistently outperforming other loss-resilient schemes. Through extensive evaluation on various videos and real network traces, we demonstrate that GRACE reduces undecodable frames by 95% and stall duration by 90% compared with FEC, while markedly boosting video quality over error concealment methods. In a user study with 240 crowdsourced participants and 960 subjective ratings, GRACE registers a 38% higher mean opinion score (MOS) than other baselines.
The convergence of text, visual, and audio data is a key step towards human-like artificial intelligence, however the current Vision-Language-Speech landscape is dominated by encoder-only models which lack generative abilities. We propose closing this gap with i-Code V2, the first model capable of generating natural language from any combination of Vision, Language, and Speech data. i-Code V2 is an integrative system that leverages state-of-the-art single-modality encoders, combining their outputs with a new modality-fusing encoder in order to flexibly project combinations of modalities into a shared representational space. Next, language tokens are generated from these representations via an autoregressive decoder. The whole framework is pretrained end-to-end on a large collection of dual- and single-modality datasets using a novel text completion objective that can be generalized across arbitrary combinations of modalities. i-Code V2 matches or outperforms state-of-the-art single- and dual-modality baselines on 7 multimodal tasks, demonstrating the power of generative multimodal pretraining across a diversity of tasks and signals.
As deep learning gains popularity in modelling dynamical systems, we expose an underappreciated misunderstanding relevant to modelling dynamics on networks. Strongly influenced by graph neural networks, latent vertex embeddings are naturally adopted in many neural dynamical network models. However, we show that embeddings tend to induce a model that fits observations well but simultaneously has incorrect dynamical behaviours. Recognising that previous studies narrowly focus on short-term predictions during the transient phase of a flow, we propose three tests for correct long-term behaviour, and illustrate how an embedding-based dynamical model fails these tests, and analyse the causes, particularly through the lens of topological conjugacy. In doing so, we show that the difficulties can be avoided by not using embedding. We propose a simple embedding-free alternative based on parametrising two additive vector-field components. Through extensive experiments, we verify that the proposed model can reliably recover a broad class of dynamics on different network topologies from time series data.
We consider the well-known and important tasks of clone detection and information retrieval for source code. The most standard setup is to search clones inside the same language code snippets. But it is also useful to find code snippets with identical behaviour in different programming languages. Nevertheless multi- and cross-lingual clone detection has been little studied in literature. We present a novel training procedure, cross-consistency training (CCT) leveraging cross-lingual similarity, that we apply to train language models on source code in various programming languages. We show that this training is effective both for encoder- and decoder-based models. The trained encoder-based CCT-LM model achieves a new state of the art on POJ-104 (monolingual C++ clone detection benchmark) with 96.73\% MAP and AdvTest (monolingual Python code search benchmark) with 47.18\% MRR. The decoder-based CCT-LM model shows comparable performance in these tasks. In addition, we formulate the multi- and cross-lingual clone detection problem and present XCD, a new benchmark dataset produced from CodeForces submissions.