分享

Pose-estimation methods enable extracting human motion from common videos in the structured form of 3D skeleton sequences. Despite great application opportunities, effective content-based access to such spatio-temporal motion data is a challenging problem. In this paper, we focus on the recently introduced text-motion retrieval tasks, which aim to search for database motions that are the most relevant to a specified natural-language textual description (text-to-motion) and vice-versa (motion-to-text). Despite recent efforts to explore these promising avenues, a primary challenge remains the insufficient data available to train robust text-motion models effectively. To address this issue, we propose to investigate joint-dataset learning - where we train on multiple text-motion datasets simultaneously - together with the introduction of a Cross-Consistent Contrastive Loss function (CCCL), which regularizes the learned text-motion common space by imposing uni-modal constraints that augment the representation ability of the trained network. To learn a proper motion representation, we also introduce a transformer-based motion encoder, called MoT++, which employs spatio-temporal attention to process sequences of skeleton data. We demonstrate the benefits of the proposed approaches on the widely-used KIT Motion-Language and HumanML3D datasets. We perform detailed experimentation on joint-dataset learning and cross-dataset scenarios, showing the effectiveness of each introduced module in a carefully conducted ablation study and, in turn, pointing out the limitations of state-of-the-art methods.

In many real-world scenarios, recorded videos suffer from accidental focus blur, and while video deblurring methods exist, most specifically target motion blur. This paper introduces a framework optimised for the joint task of focal deblurring (refocusing) and video super-resolution (VSR). The proposed method employs novel map guided transformers, in addition to image propagation, to effectively leverage the continuous spatial variance of focal blur and restore the footage. We also introduce a flow re-focusing module to efficiently align relevant features between the blurry and sharp domains. Additionally, we propose a novel technique for generating synthetic focal blur data, broadening the model's learning capabilities to include a wider array of content. We have made a new benchmark dataset, DAVIS-Blur, available. This dataset, a modified extension of the popular DAVIS video segmentation set, provides realistic out-of-focus blur degradations as well as the corresponding blur maps. Comprehensive experiments on DAVIS-Blur demonstrate the superiority of our approach. We achieve state-of-the-art results with an average PSNR performance over 1.9dB greater than comparable existing video restoration methods. Our source code will be made available at https://github.com/crispianm/DaBiT

Diffusion models have emerged as powerful tools for solving inverse problems due to their exceptional ability to model complex prior distributions. However, existing methods predominantly assume known forward operators (i.e., non-blind), limiting their applicability in practical settings where acquiring such operators is costly. Additionally, many current approaches rely on pixel-space diffusion models, leaving the potential of more powerful latent diffusion models (LDMs) underexplored. In this paper, we introduce LatentDEM, an innovative technique that addresses more challenging blind inverse problems using latent diffusion priors. At the core of our method is solving blind inverse problems within an iterative Expectation-Maximization (EM) framework: (1) the E-step recovers clean images from corrupted observations using LDM priors and a known forward model, and (2) the M-step estimates the forward operator based on the recovered images. Additionally, we propose two novel optimization techniques tailored for LDM priors and EM frameworks, yielding more accurate and efficient blind inversion results. As a general framework, LatentDEM supports both linear and non-linear inverse problems. Beyond common 2D image restoration tasks, it enables new capabilities in non-linear 3D inverse rendering problems. We validate LatentDEM's performance on representative 2D blind deblurring and 3D sparse-view reconstruction tasks, demonstrating its superior efficacy over prior arts.

Diffusion models excel in solving imaging inverse problems due to their ability to model complex image priors. However, their reliance on large, clean datasets for training limits their practical use where clean data is scarce. In this paper, we propose EMDiffusion, an expectation-maximization (EM) approach to train diffusion models from corrupted observations. Our method alternates between reconstructing clean images from corrupted data using a known diffusion model (E-step) and refining diffusion model weights based on these reconstructions (M-step). This iterative process leads the learned diffusion model to gradually converge to the true clean data distribution. We validate our method through extensive experiments on diverse computational imaging tasks, including random inpainting, denoising, and deblurring, achieving new state-of-the-art performance.

The advancements in zero-shot text-to-speech (TTS) methods, based on large-scale models, have demonstrated high fidelity in reproducing speaker characteristics. However, these models are too large for practical daily use. We propose a lightweight zero-shot TTS method using a mixture of adapters (MoA). Our proposed method incorporates MoA modules into the decoder and the variance adapter of a non-autoregressive TTS model. These modules enhance the ability to adapt a wide variety of speakers in a zero-shot manner by selecting appropriate adapters associated with speaker characteristics on the basis of speaker embeddings. Our method achieves high-quality speech synthesis with minimal additional parameters. Through objective and subjective evaluations, we confirmed that our method achieves better performance than the baseline with less than 40\% of parameters at 1.9 times faster inference speed. Audio samples are available on our demo page (https://ntt-hilab-gensp.github.io/is2024lightweightTTS/).

The use of large language model (LLM)-based preference evaluations has become widespread for comparing model responses, but it has revealed a notable bias towards longer responses, questioning the reliability of such evaluations. This paper explores the length bias in LLM evaluations from a data-centric perspective, analyzing 14 commonly used preference datasets and 10 reward models. Our findings indicate that human preference labeling favors longer responses and this spurious correlation is learned by the reward model and subsequently propagated to the aligned model during training. We decompose the preference evaluation metric, i.e., win rate, from the perspective of human to identify the deeper factors and conclude that the win rate is affected by two axes of model response: desirability and information mass, where the former is length-independent and related to trustworthiness, and the latter is length-dependent and can be represented by conditional entropy. Controlled experiments demonstrate that response length impacts evaluations by influencing information mass. To ensure reliable evaluation metrics that assess content quality without being confounded by response length, we propose AdapAlpaca, a simple yet effective adjustment to win rate measurement. Specifically, by adjusting the lengths of reference answers to match the test model's answers within the same interval, we debias information mass relative to length, ensuring a fair model evaluation. Furthermore, we investigate length bias in DPO using AlpacaEval and AdapAlpaca. By testing Tulu2 and Tulu2-dpo at 7B, 13B, and 70B scales, we found that DPO leads to higher human preference, but this gain is amplified by response length, with AlpacaEval showing higher win rates gain than AdapAlpaca.

The demand for deploying large language models(LLMs) on mobile devices continues to increase, driven by escalating data security concerns and cloud costs. However, network bandwidth and memory limitations pose challenges for deploying billion-level models on mobile devices. In this study, we investigate the outputs of different layers across various scales of LLMs and found that the outputs of most layers exhibit significant similarity. Moreover, this similarity becomes more pronounced as the model size increases, indicating substantial redundancy in the depth direction of the LLMs. Based on this observation, we propose an efficient model volume compression strategy, termed FoldGPT, which combines block removal and block parameter sharing.This strategy consists of three parts: (1) Based on the learnable gating parameters, we determine the block importance ranking while modeling the coupling effect between blocks. Then we delete some redundant layers based on the given removal rate. (2) For the retained blocks, we apply a specially designed group parameter sharing strategy, where blocks within the same group share identical weights, significantly compressing the number of parameters and slightly reducing latency overhead. (3) After sharing these Blocks, we "cure" the mismatch caused by sparsity with a minor amount of fine-tuning and introduce a tail-layer distillation strategy to improve the performance. Experiments demonstrate that FoldGPT outperforms previous state-of-the-art(SOTA) methods in efficient model compression, demonstrating the feasibility of achieving model lightweighting through straightforward block removal and parameter sharing.

Personalized Dialogue Generation (PDG) aims to create coherent responses according to roles or personas. Traditional PDG relies on external role data, which can be scarce and raise privacy concerns. Approaches address these issues by extracting role information from dialogue history, which often fail to generically model roles in continuous space. To overcome these limitations, we introduce a novel framework \textbf{MO}dels \textbf{R}oles from \textbf{P}ersonalized Dialogue \textbf{H}istory by \textbf{E}xploring and \textbf{U}tilizing Latent \textbf{S}pace (MORPHEUS) through a three-stage training process. Specifically, we create a persona codebook to represent roles in latent space compactly, and this codebook is used to construct a posterior distribution of role information. This method enables the model to generalize across roles, allowing the generation of personalized dialogues even for unseen roles. Experiments on both Chinese and English datasets demonstrate that MORPHEUS enhances the extraction of role information, and improves response generation without external role data. Additionally, MORPHEUS can be considered an efficient fine-tuning for large language models.

The growth rate of the GPU memory capacity has not been able to keep up with that of the size of large language models (LLMs), hindering the model training process. In particular, activations -- the intermediate tensors produced during forward propagation and reused in backward propagation -- dominate the GPU memory use. To address this challenge, we propose TBA to efficiently offload activations to high-capacity NVMe SSDs. This approach reduces GPU memory usage without impacting performance by adaptively overlapping data transfers with computation. TBA is compatible with popular deep learning frameworks like PyTorch, Megatron, and DeepSpeed, and it employs techniques such as tensor deduplication, forwarding, and adaptive offloading to further enhance efficiency. We conduct extensive experiments on GPT, BERT, and T5. Results demonstrate that TBA effectively reduces 47% of the activation peak memory usage. At the same time, TBA perfectly overlaps the I/O with the computation and incurs negligible performance overhead. We introduce the recompute-offload-keep (ROK) curve to compare the TBA offloading with other two tensor placement strategies, keeping activations in memory and layerwise full recomputation. We find that TBA achieves better memory savings than layerwise full recomputation while retaining the performance of keeping the activations in memory.

Biometric recognition systems are security systems based on intrinsic properties of their users, usually encoded in high dimension representations called embeddings, which potential theft would represent a greater threat than a temporary password or a replaceable key. To study the threat of embedding theft, we perform spoofing attacks on two behavioral biometric systems (an automatic speaker verification system and a handwritten digit analysis system) using a set of alignment techniques. Biometric recognition systems based on embeddings work in two phases: enrollment - where embeddings are collected and stored - then authentication - when new embeddings are compared to the stored ones -.The threat of stolen enrollment embeddings has been explored by the template reconstruction attack literature: reconstructing the original data to spoof an authentication system is doable with black-box access to their encoder. In this document, we explore the options available to perform template reconstruction attacks without any access to the encoder. To perform those attacks, we suppose general rules over the distribution of embeddings across encoders and use supervised and unsupervised algorithms to align an unlabeled set of embeddings with a set from a known encoder. The use of an alignment algorithm from the unsupervised translation literature gives promising results on spoofing two behavioral biometric systems.

This paper introduces an area-based source separation method designed for virtual meeting scenarios. The aim is to preserve speech signals from an unspecified number of sources within a defined spatial area in front of a linear microphone array, while suppressing all other sounds. Therefore, we employ an efficient neural network architecture adapted for multi-channel input to encompass the predefined target area. To evaluate the approach, training data and specific test scenarios including multiple target and interfering speakers, as well as background noise are simulated. All models are rated according to DNSMOS and scale-invariant signal-to-distortion ratio. Our experiments show that the proposed method separates speech from multiple speakers within the target area well, besides being of very low complexity, intended for real-time processing. In addition, a power reduction heatmap is used to demonstrate the networks' ability to identify sources located within the target area. We put our approach in context with a well-established baseline for speaker-speaker separation and discuss its strengths and challenges.

While previous chapters focused on recommendation systems (RSs) based on standardized, non-verbal user feedback such as purchases, views, and clicks -- the advent of LLMs has unlocked the use of natural language (NL) interactions for recommendation. This chapter discusses how LLMs' abilities for general NL reasoning present novel opportunities to build highly personalized RSs -- which can effectively connect nuanced and diverse user preferences to items, potentially via interactive dialogues. To begin this discussion, we first present a taxonomy of the key data sources for language-driven recommendation, covering item descriptions, user-system interactions, and user profiles. We then proceed to fundamental techniques for LLM recommendation, reviewing the use of encoder-only and autoregressive LLM recommendation in both tuned and untuned settings. Afterwards, we move to multi-module recommendation architectures in which LLMs interact with components such as retrievers and RSs in multi-stage pipelines. This brings us to architectures for conversational recommender systems (CRSs), in which LLMs facilitate multi-turn dialogues where each turn presents an opportunity not only to make recommendations, but also to engage with the user in interactive preference elicitation, critiquing, and question-answering.