Log-likelihood evaluation enables important capabilities in generative models, including model comparison, certain fine-tuning objectives, and many downstream applications. Yet paradoxically, some of today's best generative models -- diffusion and flow-based models -- still require hundreds to thousands of neural function evaluations (NFEs) to compute a single likelihood. While recent distillation methods have successfully accelerated sampling to just a few steps, they achieve this at the cost of likelihood tractability: existing approaches either abandon likelihood computation entirely or still require expensive integration over full trajectories. We present fast flow joint distillation (F2D2), a framework that simultaneously reduces the number of NFEs required for both sampling and likelihood evaluation by two orders of magnitude. Our key insight is that in continuous normalizing flows, the coupled ODEs for sampling and likelihood are computed from a shared underlying velocity field, allowing us to jointly distill both the sampling trajectory and cumulative divergence using a single flow map. F2D2 is modular, compatible with existing flow-based few-step sampling models, and requires only an additional divergence prediction head. Experiments demonstrate F2D2's capability of achieving accurate log-likelihood with few-step evaluations while maintaining high sample quality, solving a long-standing computational bottleneck in flow-based generative models. As an application of our approach, we propose a lightweight self-guidance method that enables a 2-step MeanFlow to outperform a 1024 step flow matching model with only a single additional backward NFE.
Joint Alignment and Distillation for Video Generation via Sample-Guided Distribution Matching
Tsinghua University（清华大学）；Kuaishou Technology（快手科技）；Harbin Institute of Technology (ShenZhen)（哈尔滨工业大学（深圳））；Beijing University of Posts and Telecommunications（北京邮电大学）
This work gives an in-depth derivation of the trainable evidence lower bound obtained from the marginal joint log-Likelihood with the goal of training a Multi-Modal Variational Autoencoder (M$^2$VAE).
Learning disentangled and interpretable representations is an important step towards accomplishing comprehensive data representations on the manifold. In this paper, we propose a novel representation learning algorithm which combines the inference abilities of Variational Autoencoders (VAE) with the generalization capability of Generative Adversarial Networks (GAN). The proposed model, called InfoVAEGAN, consists of three networks~: Encoder, Generator and Discriminator. InfoVAEGAN aims to jointly learn discrete and continuous interpretable representations in an unsupervised manner by using two…
Hallo-Live: Real-Time Streaming Joint Audio-Video Avatar Generation with Asynchronous Dual-Stream and Human-Centric Preference Distillation
Shanghai Innovation Institute（上海创新研究院）；Fudan University（复旦大学）；University of Science and Technology of China（中国科学技术大学）；Nanjing University（南京大学）；Baidu（百度）
Based on the manifold hypothesis, real-world data often lie on a low-dimensional manifold, while normalizing flows as a likelihood-based generative model are incapable of finding this manifold due to their structural constraints. So, one interesting question arises: $\textit{"Can we find sub-manifold(s) of data in normalizing flows and estimate the density of the data on the sub-manifold(s)?"}$. In this paper, we introduce two approaches, namely per-pixel penalized log-likelihood and hierarchical training, to answer the mentioned question. We propose a single-step method for joint manifold lea…
Deep Neural Networks are being extensively used in communication systems and Automatic Modulation Classification (AMC) in particular. However, they are very susceptible to small adversarial perturbations that are carefully crafted to change the network decision. In this work, we build on knowledge distillation ideas and adversarial training in order to build more robust AMC systems. We first outline the importance of the quality of the training data in terms of accuracy and robustness of the model. We then propose to use the Maximum Likelihood function, which could solve the AMC problem in off…
JointDistill: Adaptive Multi-Task Distillation for Joint Depth Estimation and Scene Segmentation
Northwestern Polytechnical University（北华理工大学）；Hong Kong University of Science and Technology (Guangzhou)（香港理工大学（广州））；National University of Singapore（新加坡国立大学）；Harbin Engineering University（哈尔滨工程大学）
GDSD: Reinforcement Learning as Guided Denoiser Self-Distillation for Diffusion Language Models
UCL Dept. of Statistical Science（伦敦大学学院统计科学系）；UCL Centre for AI（伦敦大学学院人工智能中心）；Alibaba Group（阿里巴巴集团）；Dept. of EEE（电子工程系）；Imperial College London（伦敦帝国理工学院）；UNIST（全南大学）；University of Basel（巴塞尔大学）
CompassOPD: Cross-Family On-Policy Distillation via Within-Family Likelihood Shifts
Institute of Information Engineering, Chinese Academy of Sciences（中国科学院信息工程研究所）；School of Cyber Security, University of Chinese Academy of Sciences（中国科学院大学网络空间安全学院）
LinkedIn（领英）；University of California, Davis（加州大学戴维斯分校）；Arizona State University（亚利桑那州立大学）；Clemson University（克莱姆森大学）；Pennsylvania State University（宾夕法尼亚州立大学）；University of Wisconsin–Madison（威斯康星大学麦迪逊分校）
Google LLC（谷歌有限责任公司）；University of Washington（华盛顿大学）
We trained a keyword spotting model using federated learning on real user devices and observed significant improvements when the model was deployed for inference on phones. To compensate for data domains that are missing from on-device training caches, we employed joint federated-centralized training. And to learn in the absence of curated labels on-device, we formulated a confidence filtering strategy based on user-feedback signals for federated distillation. These techniques created models that significantly improved quality metrics in offline evaluations and user-experience metrics in live …
University of Illinois at Urbana-Champaign（伊利诺伊大学厄巴纳-香槟分校）；Adobe Research（奥多比研究院）
Keyphrase extraction is the task of extracting a small set of phrases that best describe a document. Most existing benchmark datasets for the task typically have limited numbers of annotated documents, making it challenging to train increasingly complex neural networks. In contrast, digital libraries store millions of scientific articles online, covering a wide range of topics. While a significant portion of these articles contain keyphrases provided by their authors, most other articles lack such kind of annotations. Therefore, to effectively utilize these large amounts of unlabeled articles,…
Shanghai Jiao Tong University（上海交通大学）；Ping An Asset Management Co., Ltd（平安资产管理有限责任公司）；State Grid Fujian Electric Power Company（国网福建省电力有限公司）
The extra trust brought by the model interpretation has made it an indispensable part of machine learning systems. But to explain a distilled model's prediction, one may either work with the student model itself, or turn to its teacher model. This leads to a more fundamental question: if a distilled model should give a similar prediction for a similar reason as its teacher model on the same input? This question becomes even more crucial when the two models have dramatically different structure, taking GBDT2NN for example. This paper conducts an empirical study on the new approach to explaining…
University of Maryland, Baltimore County（马里兰大学巴尔的摩郡分校）；LANL（洛斯阿拉莫斯国家实验室）
Non-negative matrix factorization (NMF) with missing-value completion is a well-known effective Collaborative Filtering (CF) method used to provide personalized user recommendations. However, traditional CF relies on the privacy-invasive collection of users' explicit and implicit feedback to build a central recommender model. One-shot federated learning has recently emerged as a method to mitigate the privacy problem while addressing the traditional communication bottleneck of federated learning. In this paper, we present the first unsupervised one-shot federated CF implementation, named FedSP…
École de technologie supérieure（高等技术学院）；Genetec Inc.（Genetec公司）
Currently, the divergence in distributions of design and operational data, and large computational complexity are limiting factors in the adoption of CNNs in real-world applications. For instance, person re-identification systems typically rely on a distributed set of cameras, where each camera has different capture conditions. This can translate to a considerable shift between source (e.g. lab setting) and target (e.g. operational camera) domains. Given the cost of annotating image data captured for fine-tuning in each target domain, unsupervised domain adaptation (UDA) has become a popular a…
Multi-hop Deep Joint Source-Channel Coding with Deep Hash Distillation for Semantically Aligned Image Recovery
Department of Electrical and Electronic Engineering, Imperial College London, UK（电气与电子工程系，帝国理工学院伦敦分校，英国）；Lehrstuhl für Nachrichtentechnik, Technische Universität Dortmund, Germany（信息论系，德意志理工大学多特蒙德分校，德国）
I describe a trick for training flow models using a prescribed rule as a surrogate for maximum likelihood. The utility of this trick is limited for non-conditional models, but an extension of the approach, applied to maximum likelihood of the joint probability distribution of data and conditioning information, can be used to train sophisticated \textit{conditional} flow models. Unlike previous approaches, this method is quite simple: it does not require explicit knowledge of the distribution of conditions, auxiliary networks or other specific architecture, or additional loss terms beyond maxim…
University of Cambridge（剑桥大学）；Avignon University（阿维尼翁大学）；Samsung AI（三星人工智能）
Knowledge distillation has been widely used to compress existing deep learning models while preserving the performance on a wide range of applications. In the specific context of Automatic Speech Recognition (ASR), distillation from ensembles of acoustic models has recently shown promising results in increasing recognition performance. In this paper, we propose an extension of multi-teacher distillation methods to joint CTC-attention end-to-end ASR systems. We also introduce three novel distillation strategies. The core intuition behind them is to integrate the error rate metric to the teacher…
Reinforcement learning (RL) has achieved tremendous success as a general framework for learning how to make decisions. However, this success relies on the interactive hand-tuning of a reward function by RL experts. On the other hand, inverse reinforcement learning (IRL) seeks to learn a reward function from readily-obtained human demonstrations. Yet, IRL suffers from two major limitations: 1) reward ambiguity - there are an infinite number of possible reward functions that could explain an expert's demonstration and 2) heterogeneity - human experts adopt varying strategies and preferences, whi…
Sequential Beats Joint: On the Interplay between On-Policy Distillation and RLVR
New York University（纽约大学）；University of Chicago（芝加哥大学）；University of Waterloo（滑铁卢大学）；University of Alberta（阿尔伯塔大学）；NYU Shanghai（纽约大学上海分校）；Alberta Machine Intelligence Institute (Amii)（阿尔伯塔机器智能研究所）
Although with progress in introducing auxiliary amortized inference models, learning discrete latent variable models is still challenging. In this paper, we show that the annoying difficulty of obtaining reliable stochastic gradients for the inference model and the drawback of indirectly optimizing the target log-likelihood can be gracefully addressed in a new method based on stochastic approximation (SA) theory of the Robbins-Monro type. Specifically, we propose to directly maximize the target log-likelihood and simultaneously minimize the inclusive divergence between the posterior and the in…
Imperial College London（伦敦帝国学院）；New York University（纽约大学）；University of Liège（列日大学）；Federico Santa María Technical University（费德里科·圣玛丽亚理工大学）
We extend recent work (Brehmer, et. al., 2018) that use neural networks as surrogate models for likelihood-free inference. As in the previous work, we exploit the fact that the joint likelihood ratio and joint score, conditioned on both observed and latent variables, can often be extracted from an implicit generative model or simulator to augment the training data for these surrogate models. We show how this augmented training data can be used to provide a new cross-entropy estimator, which provides improved sample efficiency compared to previous loss functions exploiting this augmented traini…
Center for Nuclear Theory, Department of Physics and Astronomy, Stony Brook University（核理论中心，物理与天文学系，石溪大学）；Massachusetts Institute of Technology（麻省理工学院）；Université Paris-Saclay, CNRS/IN2P3, IJCLab（巴黎-萨克雷大学，CNRS/IN2P3，IJCLab）；Institute for Particle Physics Phenomenology（粒子物理学现象研究所）；Department of Physics, Durham University（物理系，杜ham大学）
