论文检索

Deep Learning · Large Language Models

Madhur Panwar, Kabir Ahuja, Navin Goyal

In-context learning (ICL) is one of the surprising and useful features of large language models and subject of intense research. Recently, stylized meta-learning-like ICL setups have been devised that train transformers on sequences of input-output pairs $(x, f(x))$. The function $f$ comes from a function class and generalization is checked by evaluating on sequences generated from unseen functions from the same class. One of the main discoveries in this line of research has been that for several function classes, such as linear regression, transformers successfully generalize to new functions in the class. However, the inductive biases of these models resulting in this behavior are not clearly understood. A model with unlimited training data and compute is a Bayesian predictor: it learns the pretraining distribution.In this paper we empirically examine how far this Bayesian perspective can help us understand ICL. To this end, we generalize the previous meta-ICL setup to hierarchical meta-ICL setup which involve unions of multiple task families. We instantiate this setup on a diverse range of linear and nonlinear function families and find that transformers can do ICL in this setting as well. Where Bayesian inference is tractable, we find evidence that high-capacity transformers mimic the Bayesian predictor. The Bayesian perspective provides insights into the inductive bias of ICL and how transformers perform a particular task when they are trained on multiple tasks. We also find that transformers can learn to generalize to new function classes that were not seen during pretraining. This involves deviation from the Bayesian predictor. We examine these deviations in more depth offering new insights and hypotheses.

Probabilistic Methods · Bayesian Models and Methods

Minyoung Kim, Timothy Hospedales

We propose a novel hierarchical Bayesian model for the few-shot meta learning problem. We consider episode-wise random variables to model episode-specific generative processes, where these local random variables are governed by a higher-level global random variable. The global variable captures information shared across episodes, while controlling how much the model needs to be adapted to new episodes in a principled Bayesian manner. Within our framework, prediction on a novel episode/task can be seen as a Bayesian inference problem. For tractable training, we need to be able to relate each local episode-specific solution to the global higher-level parameters. We propose a Normal-Inverse-Wishart model, for which establishing this local-global relationship becomes feasible due to the approximate closed-form solutions for the local posterior distributions. The resulting algorithm is more attractive than the MAML in that it does not maintain a costly computational graph for the sequence of gradient descent steps in an episode. Our approach is also different from existing Bayesian meta learning methods in that rather than modeling a single random variable for all episodes, it leverages a hierarchical structure that exploits the local-global relationships desirable for principled Bayesian learning with many related tasks.

Reinforcement Learning

Safa Messaoud, Billel Mokeddem, Zhenghai Xue, Linsey Pang, Bo An, Haipeng Chen, Sanjay Chawla

Learning expressive stochastic policies instead of deterministic ones has been proposed to achieve better stability, sample complexity and robustness. Notably, in Maximum Entropy reinforcement learning (MaxEnt RL), the policy is modeled as an expressive energy-based model (EBM) over the Q-values. However, this formulation requires the estimation of the entropy of such EBM distributions which is an open problem. To address this, previous MaxEnt RL methods either implicitly estimate the entropy, yielding high computational complexity and variance (SQL), or follow a variational inference approach that fits simplified distributions (e.g., Gaussian) for tractability (SAC). We propose Sein Soft Actor-Critic (S$^2$AC), a MaxEnt RL algorithm that learns expressive policies without compromising efficiency. S$^2$AC uses parameterized Stein Variational Gradient Descent (SVGD) as the underlying policy. At the core of S$^2$AC is a new solution to the above open challenge of entropy computation for EBMs. Our entropy formula is computationally efficient and only depends on first-order derivatives and vector products. Empirical results show that S$^2$AC yields more optimal solutions to the MaxEnt objective than SQL and SAC in the multi-goal environment, and outperforms SAC and SQL on the MuJoCo benchmark. Our code is available at: https://anonymous.4open.science/r/Stein-Soft-Actor-Critic/

Deep Learning · Graph Neural Networks

Milan Papez, Martin Rektoris, Vaclav Smidl, Tomáš Pevný

Daily internet communication relies heavily on tree-structured graphs, embodied by popular data formats such as XML and JSON. However, many recent generative (probabilistic) models utilize neural networks to learn a probability distribution over undirected cyclic graphs. This assumption of a generic graph structure brings various computational challenges, and, more importantly, the presence of non-linearities in neural networks does not permit tractable probabilistic inference. We address these problems by proposing sum-product-set networks, an extension of probabilistic circuits from unstructured tensor data to tree-structured graph data. To this end, we use random finite sets to reflect a variable number of nodes and edges in the graph and to allow for exact and efficient inference. We demonstrate that our tractable model performs comparably to various intractable models based on neural networks.

Wenbo Li, Xin Yu, Kun Zhou, Yibing Song, Zhe Lin

Generative adversarial networks (GANs) have made great success in image inpainting yet still have difficulties tackling large missing regions. In contrast, iterative probabilistic algorithms, such as autoregressive and denoising diffusion models, have to be deployed with massive computing resources for decent effect. To achieve high-quality results with low computational cost, we present a novel pixel spread model (PSM) that iteratively employs decoupled probabilistic modeling, combining the optimization efficiency of GANs with the prediction tractability of probabilistic models. As a result, our model selectively spreads informative pixels throughout the image in a few iterations, largely enhancing the completion quality and efficiency. On multiple benchmarks, we achieve new state-of-the-art performance. Our code and models will be publicly available.

Deep Learning · Large Language Models

Chaoqi Wang, Yibo Jiang, Chenghao Yang, Han Liu, Yuxin Chen

The increasing capabilities of large language models (LLMs) raise opportunities for artificial general intelligence but concurrently amplify safety concerns, such as potential misuse of AI systems, necessitating effective AI alignment. Reinforcement Learning from Human Feedback (RLHF) has emerged as a promising pathway towards AI alignment but brings forth challenges due to its complexity and dependence on a separate reward model. Direct Preference Optimization (DPO) has been proposed as an alternative; and it remains equivalent to RLHF under the reverse KL regularization constraint. This paper presents $f$-DPO, a generalized approach to DPO by incorporating diverse divergence constraints. We show that under certain $f$-divergences, including Jensen-Shannon divergence, forward KL divergences and $\alpha$-divergences, the complex relationship between the reward and optimal policy can also be simplified by addressing the Karush–Kuhn–Tucker conditions. This eliminates the need for estimating the normalizing constant in the Bradley-Terry model and enables a tractable mapping between the reward function and the optimal policy. Our approach optimizes LLMs to align with human preferences in a more efficient and supervised manner under a broad set of divergence constraints. Empirically, adopting these divergences ensures a balance between alignment performance and generation diversity. Importantly, our $f$-DPO outperforms PPO-based methods in divergence efficiency, and divergence constraints directly influence expected calibration error (ECE).

Deep Learning · Generative Models and Autoencoders

Yeongmin Kim, Byeonghu Na, Minsang Park, JoonHo Jang, Dongjun Kim, Wanmo Kang, Il-chul Moon

With significant advancements in diffusion models, addressing the potential risks of dataset bias becomes increasingly important. Since generated outputs directly suffer from dataset bias, mitigating latent bias becomes a key factor in improving sample quality and proportion. This paper proposes time-dependent importance reweighting to mitigate the bias for the diffusion models. We demonstrate that the time-dependent density ratio becomes more precise than previous approaches, thereby minimizing error propagation in generative learning. While directly applying it to score-matching is intractable, we discover that using the time-dependent density ratio both for reweighting and score correction can lead to a tractable form of the objective function to regenerate the unbiased data density. Furthermore, we theoretically establish a connection with traditional score-matching, and we demonstrate its convergence to an unbiased distribution. The experimental evidence supports the usefulness of the proposed method, which outperforms baselines including time-independent importance reweighting on CIFAR-10, CIFAR-100, FFHQ, and CelebA with various bias settings. Our code is available at https://github.com/alsdudrla10/TIW-DSM.

Applications · Language, Speech and Dialog

Haozhe Ji, Pei Ke, Hongning Wang, Minlie Huang

Despite the remarkable advances in language modeling, current mainstream decoding methods still struggle to generate texts that align with human texts across different aspects. In particular, sampling-based methods produce less-repetitive texts which are often disjunctive in discourse, while search-based methods maintain topic coherence at the cost of increased repetition. Overall, these methods fall short in achieving holistic alignment across a broad range of aspects. In this work, we frame decoding from a language model as an optimization problem with the goal of strictly matching the expected performance with human texts measured by multiple metrics of desired aspects simultaneously. The resulting decoding distribution enjoys an analytical solution that scales the input language model distribution via a sequence-level energy function defined by these metrics. And most importantly, we prove that this induced distribution is guaranteed to improve the perplexity on human texts, which suggests a better approximation to the underlying distribution of human texts. To facilitate tractable sampling from this globally normalized distribution, we adopt the Sampling-Importance-Resampling technique. Experiments on various domains and model scales demonstrate the superiority of our method in metrics alignment with human texts and human evaluation over strong baselines.

Ilia Igashov, Arne Schneuing, Marwin Segler, Michael Bronstein, Bruno Correia

Retrosynthesis planning is a fundamental challenge in chemistry which aims at designing multi-step reaction pathways from commercially available starting materials to a target molecule. Each step in multi-step retrosynthesis planning requires accurate prediction of possible precursor molecules given the target molecule and confidence estimates to guide heuristic search algorithms. We model single-step retrosynthesis as a distribution learning problem in a discrete state space. First, we introduce the Markov Bridge Model, a generative framework aimed to approximate the dependency between two intractable discrete distributions accessible via a finite sample of coupled data points. Our framework is based on the concept of a Markov bridge, a Markov process pinned at its endpoints. Unlike diffusion-based methods, our Markov Bridge Model does not need a tractable noise distribution as a sampling proxy and directly operates on the input product molecules as samples from the intractable prior distribution. We then address the retrosynthesis planning problem with our novel framework and introduce RetroBridge, a template-free retrosynthesis modeling approach that achieves state-of-the-art results on standard evaluation benchmarks.

Probabilistic Methods · Gaussian Processes

Yijue Dai, Wenzhong Yan, Feng Yin

Integrating information while recognizing dependence from multiple data sources and enhancing the predictive performance of the multi-output regression are challenging tasks. Multioutput Gaussian Process (MOGP) methods offer outstanding solutions with tractable predictions and uncertainty quantification. However, their practical applications are hindered by high computational complexity and storage demand. Additionally, there exist model mismatches in existing MOGP models when dealing with non-Gaussian data. To improve the model representation ability in terms of flexibility, optimality, and scalability, this paper introduces a novel multi-output regression framework, termed Graphical MOGP (GMOGP), which is empowered by: (i) Generating flexible Gaussian process priors consolidated from dentified parents, (ii) providing dependent processes with attention-based graphical representations, and (iii) achieving Pareto optimal solutions of kernel hyperparameters via a distributed learning framework. Numerical results confirm that the proposed GMOGP significantly outperforms state-of-the-art MOGP alternatives in predictive performance, as well as in time and memory efficiency, across various synthetic and real datasets.

Probabilistic Methods · Bayesian Models and Methods

Yongsheng Mei, Mahdi Imani, Tian Lan

Bayesian optimization (BO) has established itself as a leading strategy for efficiently optimizing expensive-to-evaluate functions. Existing BO methods mostly rely on Gaussian process (GP) surrogate models and are not applicable to (doubly-stochastic) Gaussian Cox processes, where the observation process is modulated by a latent intensity function modeled as a GP. In this paper, we propose a novel maximum *a posteriori* inference of Gaussian Cox processes. It leverages the Laplace approximation and change of kernel technique to transform the problem into a new reproducing kernel Hilbert space, where it becomes more tractable computationally. It enables us to obtain both a functional posterior of the latent intensity function and the covariance of the posterior, thus extending existing works that often focus on specific link functions or estimating the posterior mean. Using the result, we propose a BO framework based on the Gaussian Cox process model and further develop a Nyström approximation for efficient computation. Extensive evaluations on various synthetic and real-world datasets demonstrate significant improvement over state-of-the-art inference solutions for Gaussian Cox processes, as well as effective BO with a wide range of acquisition functions designed through the underlying Gaussian Cox process model.

Theory · Domain Adaptation and Transfer Learning

Simone Magistri, Tomaso Trinci, Albin Soutif--Cormerais, Joost van de Weijer, Andrew Bagdanov

Exemplar-Free Class Incremental Learning (EFCIL) aims to learn from a sequence of tasks without having access to previous task data. In this paper, we consider the challenging Cold Start scenario in which insufficient data is available in the first task to learn a high-quality backbone. This is especially challenging for EFCIL since it requires high plasticity, which results in feature drift which is difficult to compensate for in the exemplar-free setting. To address this problem, we propose a simple and effective approach that consolidates feature representations by regularizing drift in directions highly relevant to previous tasks and employs prototypes to reduce task-recency bias. Our method, called Elastic Feature Consolidation (EFC), exploits a tractable second-order approximation of feature drift based on an Empirical Feature Matrix (EFM). The EFM induces a pseudo-metric in feature space which we use to regularize feature drift in important directions and to update Gaussian prototypes used in a novel asymmetric cross entropy loss which effectively balances prototype rehearsal with data from new tasks. Experimental results on CIFAR-100, Tiny-ImageNet, ImageNet-Subset and ImageNet-1K demonstrate that Elastic Feature Consolidation is better able to learn new tasks by maintaining model plasticity and significantly outperform the state-of-the-art.

Applications · Neuroscience, Cognitive Science

Michael Kleinman, Alessandro Achille, Stefano Soatto

Critical learning periods are periods early in development where temporary sensory deficits can have a permanent effect on behavior and learned representations. Despite the radical differences between biological and artificial networks, critical learning periods have been empirically observed in both systems. This suggests that critical periods may be fundamental to learning and not an accident of biology.Yet, why exactly critical periods emerge in deep networks is still an open question, and in particular it is unclear whether the critical periods observed in both systems depend on particular architectural or optimization details. To isolate the key underlying factors, we focus on deep linear network models, and show that, surprisingly, such networks also display much of the behavior seen in biology and artificial networks, while being amenable to analytical treatment. We show that critical periods depend on the depth of the model and structure of the data distribution. We also show analytically and in simulations that the learning of features is tied to competition between sources. Finally, we extend our analysis to multi-task learning to show that pre-training on certain tasks can damage the transfer performance on new tasks, and show how this depends on the relationship between tasks and the duration of the pre-training stage. To the best of our knowledge, our work provides the first analytically tractable model that sheds light into why critical learning periods emerge in biological and artificial networks.

Deep Learning · Robustness

Dyah Adila, Changho Shin, Linrong Cai, Frederic Sala

Zero-shot inference is a powerful paradigm that enables the use of large pretrained models for downstream classification tasks without further training. However, these models are vulnerable to inherited biases that can impact their performance. The traditional solution is fine-tuning, but this undermines the key advantage of pretrained models, which is their ability to be used out-of-the-box. We propose RoboShot, a method that improves the robustness of pretrained model embeddings in a fully zero-shot fashion. First, we use language models (LMs) to obtain useful insights from task descriptions. These insights are embedded and used to remove harmful and boost useful components in embeddings---without any supervision. Theoretically, we provide a simple and tractable model for biases in zero-shot embeddings and give a result characterizing under what conditions our approach can boost performance. Empirically, we evaluate RoboShot on nine image and NLP classification tasks and show an average improvement of 15.98% over several zero-shot baselines. Additionally, we demonstrate that RoboShot is compatible with a variety of pretrained and language models and propose a way to further boost performance with a zero-shot adaptation variant.

Michał J Tyszkiewicz, Pascal Fua, Eduard Trulls

Diffusion models generating images conditionally on text, such as Dall-E 2 and Stable Diffusion, have recently made a splash far beyond the computer vision community. Here, we tackle the related problem of generating point clouds, both unconditionally, and conditionally with images. For the latter, we introduce a novel geometrically-motivated conditioning scheme based on projecting sparse image features into the point cloud and attaching them to each individual point, at every step in the denoising process. This approach improves geometric consistency and yields greater fidelity than current methods relying on unstructured, global latent codes. Additionally, we show how to apply recent continuous-time diffusion schemes. Our method performs on par or above the state of art on conditional and unconditional experiments on synthetic data, while being faster, lighter, and delivering tractable likelihoods. We show it can also scale to diverse indoors scenes.

Simian Luo, Xuelin Qian, Yanwei Fu, Yinda Zhang, Ying Tai, Zhenyu Zhang, Chengjie Wang, Xiangyang Xue

Auto-Regressive (AR) models have achieved impressive results in 2D image generation by modeling joint distributions in the grid space. While this approach has been extended to the 3D domain for powerful shape generation, it still has two limitations: expensive computations on volumetric grids and ambiguous auto-regressive order along grid dimensions. To overcome these limitations, we propose the Improved Auto-regressive Model (ImAM) for 3D shape generation, which applies discrete representation learning based on a latent vector instead of volumetric grids. Our approach not only reduces computational costs but also preserves essential geometric details by learning the joint distribution in a more tractable order. Moreover, thanks to the simplicity of our model architecture, we can naturally extend it from unconditional to conditional generation by concatenating various conditioning inputs, such as point clouds, categories, images, and texts. Extensive experiments demonstrate that ImAM can synthesize diverse and faithful shapes of multiple categories, achieving state-of-the-art performance.

Muyao Niu, Zhihang Zhong, Yinqiang Zheng

Low-light video enhancement in the visible (VIS) range is important yet technically challenging, and it is likely to become more tractable by introducing near-infrared (NIR) information for assistance, which in turn arouses a new challenge on how to obtain appropriate multispectral data for model training. In this paper, we defend the feasibility and superiority of NIR-assisted low-light video enhancement results by using unpaired 24-hour data for the first time, which significantly eases data collection and improves generalization performance on in-the-wild data. By accounting for different physical characteristics between unpaired daytime and nighttime videos, we first propose to turn daytime NIR & VIS into "nighttime mode". Specifically, we design a heuristic yet physics-inspired relighting algorithm to produce realistic pseudo nighttime NIR, and use a resampling strategy followed by a noiseGAN for nighttime VIS conversion. We further devise a temporal-aware network for video enhancement that extracts and fuses bi-directional temporal streams and is trained using real daytime videos and pseudo nighttime videos. We capture multi-spectral data using a co-axial camera and contribute Fulltime Multi-Spectral Video Dataset (FMSVD), the first dataset including aligned 24-hour NIR & VIS videos. Compared to alternative methods, we achieve significantly improved video quality as well as generalization ability on in-the-wild data in terms of both evaluation metrics and visual judgment. Codes and Data Available: https://github.com/MyNiuuu/NVEU.

Zenan Huang, Haobo Wang, Junbo Zhao, Nenggan Zheng

Existing machine learning (ML) models are often fragile in open environments because the data distribution frequently shifts. To address this problem, domain generalization (DG) aims to explore underlying invariant patterns for stable prediction across domains. In this work, we first characterize that this failure of conventional ML models in DG is attributed to an inadequate identification of causal structures. We further propose a novel and theoretically grounded invariant Directed Acyclic Graph (dubbed iDAG) searching framework that attains an invariant graphical relation as the proxy to the causality structure from the intrinsic data-generating process. To enable tractable computation, iDAG solves a constrained optimization objective built on a set of representative class-conditional prototypes. Additionally, we integrate a hierarchical contrastive learning module, which poses a strong effect of clustering, for enhanced prototypes as well as stabler prediction. Extensive experiments on the synthetic and real-world benchmarks demonstrate that iDAG outperforms the state-of-the-art approaches, verifying the superiority of causal structure identification for DG. The code of iDAG is available at https://github.com/lccurious/iDAG.

Hao Shen 0006, Zhong-Qiu Zhao, Yulun Zhang 0001, Zhao Zhang 0001

Multi-stage architectures have exhibited efficacy in image dehazing, which usually decomposes a challenging task into multiple more tractable sub-tasks and progressively estimates latent hazy-free images. Despite the remarkable progress, existing methods still suffer from the following shortcomings: (1) limited exploration of frequency domain information; (2) insufficient information interaction; (3) severe feature redundancy. To remedy these issues, we propose a novel Mutual Information-driven Triple interaction Network (MITNet) based on spatial-frequency dual domain information and two-stage architecture. To be specific, the first stage, named amplitude-guided haze removal, aims to recover the amplitude spectrum of the hazy images for haze removal. And the second stage, named phase-guided structure refined, devotes to learning the transformation and refinement of the phase spectrum. To facilitate the information exchange between two stages, an Adaptive Triple Interaction Module (ATIM) is developed to simultaneously aggregate cross-domain, cross-scale, and cross-stage features, where the fused features are further used to generate content-adaptive dynamic filters so that applying them to enhance global context representation. In addition, we impose the mutual information minimization constraint on paired scale encoder and decoder features from both stages. Such an operation can effectively reduce information redundancy and enhance cross-stage feature complementarity. Extensive experiments on multiple public datasets exhibit that our MITNet performs superior performance with lower model complexity. The code and models are available at https://github.com/it-hao/MITNet.

Joshua Ainslie, Tao Lei, Michiel de Jong, Santiago Ontañón, Siddhartha Brahma, Yury Zemlyanskiy, David Uthus, Mandy Guo, James Lee-Thorp, Yi Tay 等

Many natural language processing tasks benefit from long inputs, but processing long documents with Transformers is expensive – not only due to quadratic attention complexity but also from applying feedforward and projection layers to every token. However, not all tokens are equally important, especially for longer documents. We propose CoLT5, a long-input Transformer model that builds on this intuition by employing conditional computation, devoting more resources to important tokens in both feedforward and attention layers. We show that CoLT5 achieves stronger performance than LongT5 with much faster training and inference, achieving SOTA on the long-input SCROLLS benchmark. Moreover, CoLT5 can effectively and tractably make use of extremely long inputs, showing strong gains up to 64k input length.