We develop a thermodynamic theory for machine learning (ML) systems. Similar to physical thermodynamic systems which are characterized by energy and entropy, ML systems possess these characteristics as well. This comparison inspire us to integrate the concept of temperature into ML systems grounded in the fundamental principles of thermodynamics, and establish a basic thermodynamic framework for machine learning systems with non-Boltzmann distributions. We introduce the concept of states within a ML system, identify two typical types of state, and interpret model training and refresh as a process of state phase transition. We consider that the initial potential energy of a ML system is described by the model's loss functions, and the energy adheres to the principle of minimum potential energy. For a variety of energy forms and parameter initialization methods, we derive the temperature of systems during the phase transition both analytically and asymptotically, highlighting temperature as a vital indicator of system data distribution and ML training complexity. Moreover, we perceive deep neural networks as complex heat engines with both global temperature and local temperatures in each layer. The concept of work efficiency is introduced within neural networks, which mainly depends on the neural activation functions. We then classify neural networks based on their work efficiency, and describe neural networks as two types of heat engines.
相關內容
機器學習（Machine Learning）是一個研究計算學習方法的國際論壇。該雜志發表文章，報告廣泛的學習方法應用于各種學習問題的實質性結果。該雜志的特色論文描述研究的問題和方法，應用研究和研究方法的問題。有關學習問題或方法的論文通過實證研究、理論分析或與心理現象的比較提供了堅實的支持。應用論文展示了如何應用學習方法來解決重要的應用問題。研究方法論文改進了機器學習的研究方法。所有的論文都以其他研究人員可以驗證或復制的方式描述了支持證據。論文還詳細說明了學習的組成部分，并討論了關于知識表示和性能任務的假設。
官網地址：
Explainable AI methods facilitate the understanding of model behaviour, yet, small, imperceptible perturbations to inputs can vastly distort explanations. As these explanations are typically evaluated holistically, before model deployment, it is difficult to assess when a particular explanation is trustworthy. Some studies have tried to create confidence estimators for explanations, but none have investigated an existing link between uncertainty and explanation quality. We artificially simulate epistemic uncertainty in text input by introducing noise at inference time. In this large-scale empirical study, we insert different levels of noise perturbations and measure the effect on the output of pre-trained language models and different uncertainty metrics. Realistic perturbations have minimal effect on performance and explanations, yet masking has a drastic effect. We find that high uncertainty doesn't necessarily imply low explanation plausibility; the correlation between the two metrics can be moderately positive when noise is exposed during the training process. This suggests that noise-augmented models may be better at identifying salient tokens when uncertain. Furthermore, when predictive and epistemic uncertainty measures are over-confident, the robustness of a saliency map to perturbation can indicate model stability issues. Integrated Gradients shows the overall greatest robustness to perturbation, while still showing model-specific patterns in performance; however, this phenomenon is limited to smaller Transformer-based language models.
The Skolem problem is a long-standing open problem in linear dynamical systems: can a linear recurrence sequence (LRS) ever reach 0 from a given initial configuration? Similarly, the positivity problem asks whether the LRS stays positive from an initial configuration. Deciding Skolem (or positivity) has been open for half a century: the best known decidability results are for LRS with special properties (e.g., low order recurrences). But these problems are easier for "uninitialized" variants, where the initial configuration is not fixed but can vary arbitrarily: checking if there is an initial configuration from which the LRS stays positive can be decided in polynomial time (Tiwari in 2004, Braverman in 2006). In this paper, we consider problems that lie between the initialized and uninitialized variants. More precisely, we ask if 0 (resp. negative numbers) can be avoided from every initial configuration in a neighborhood of a given initial configuration. This can be considered as a robust variant of the Skolem (resp. positivity) problem. We show that these problems lie at the frontier of decidability: if the neighbourhood is given as part of the input, then robust Skolem and robust positivity are Diophantine hard, i.e., solving either would entail major breakthroughs in Diophantine approximations, as happens for (non-robust) positivity. However, if one asks whether such a neighbourhood exists, then the problems turn out to be decidable with PSPACE complexity. Our techniques also allow us to tackle robustness for ultimate positivity, which asks whether there is a bound on the number of steps after which the LRS remains positive. There are two variants depending on whether we ask for a "uniform" bound on this number of steps. For the non-uniform variant, when the neighbourhood is open, the problem turns out to be tractable, even when the neighbourhood is given as input.
Mixed linear regression is a well-studied problem in parametric statistics and machine learning. Given a set of samples, tuples of covariates and labels, the task of mixed linear regression is to find a small list of linear relationships that best fit the samples. Usually it is assumed that the label is generated stochastically by randomly selecting one of two or more linear functions, applying this chosen function to the covariates, and potentially introducing noise to the result. In that situation, the objective is to estimate the ground-truth linear functions up to some parameter error. The popular expectation maximization (EM) and alternating minimization (AM) algorithms have been previously analyzed for this. In this paper, we consider the more general problem of agnostic learning of mixed linear regression from samples, without such generative models. In particular, we show that the AM and EM algorithms, under standard conditions of separability and good initialization, lead to agnostic learning in mixed linear regression by converging to the population loss minimizers, for suitably defined loss functions. In some sense, this shows the strength of AM and EM algorithms that converges to ``optimal solutions'' even in the absence of realizable generative models.
We consider the problem of joint learning of multiple linear dynamical systems. This has received significant attention recently under different types of assumptions on the model parameters. The setting we consider involves a collection of $m$ linear systems each of which resides on a node of a given undirected graph $G = ([m], \mathcal{E})$. We assume that the system matrices are marginally stable, and satisfy a smoothness constraint w.r.t $G$ -- akin to the quadratic variation of a signal on a graph. Given access to the states of the nodes over $T$ time points, we then propose two estimators for joint estimation of the system matrices, along with non-asymptotic error bounds on the mean-squared error (MSE). In particular, we show conditions under which the MSE converges to zero as $m$ increases, typically polynomially fast w.r.t $m$. The results hold under mild (i.e., $T \sim \log m$), or sometimes, even no assumption on $T$ (i.e. $T \geq 2$).
Large Language Models (LLMs) have highlighted the necessity of effective unlearning mechanisms to comply with data regulations and ethical AI practices. LLM unlearning aims at removing undesired data influences and associated model capabilities without compromising utility out of the scope of unlearning. While interest in studying LLM unlearning is growing,the impact of the optimizer choice for LLM unlearning remains under-explored. In this work, we shed light on the significance of optimizer selection in LLM unlearning for the first time, establishing a clear connection between {second-order optimization} and influence unlearning (a classical approach using influence functions to update the model for data influence removal). This insight propels us to develop a second-order unlearning framework, termed SOUL, built upon the second-order clipped stochastic optimization (Sophia)-based LLM training method. SOUL extends the static, one-shot model update using influence unlearning to a dynamic, iterative unlearning process. Our extensive experiments show that SOUL consistently outperforms conventional first-order methods across various unlearning tasks, models, and metrics, suggesting the promise of second-order optimization in providing a scalable and easily implementable solution for LLM unlearning.
Perceptual hashing algorithms (PHAs) are utilized extensively for identifying illegal online content. Given their crucial role in sensitive applications, understanding their security strengths and weaknesses is critical. This paper compares three major PHAs deployed widely in practice: PhotoDNA, PDQ, and NeuralHash, and assesses their robustness against three typical attacks: normal image editing attacks, malicious adversarial attacks, and hash inversion attacks. Contrary to prevailing studies, this paper reveals that these PHAs exhibit resilience to black-box adversarial attacks when realistic constraints regarding the distortion and query budget are applied, attributed to the unique property of random hash variations. Moreover, this paper illustrates that original images can be reconstructed from the hash bits, raising significant privacy concerns. By comprehensively exposing their security vulnerabilities, this paper contributes to the ongoing efforts aimed at enhancing the security of PHAs for effective deployment.
A hierarchical Bayesian approach that permits simultaneous inference for the regression coefficient matrix and the error precision (inverse covariance) matrix in the multivariate linear model is proposed. Assuming a natural ordering of the elements of the response, the precision matrix is reparameterized so it can be estimated with univariate-response linear regression techniques. A novel generalized bridge regression prior that accommodates both sparse and dense settings and is competitive with alternative methods for univariate-response regression is proposed and used in this framework. Two component-wise Markov chain Monte Carlo algorithms are developed for sampling, including a data augmentation algorithm based on a scale mixture of normals representation. Numerical examples demonstrate that the proposed method is competitive with comparable joint mean-covariance models, particularly in estimation of the precision matrix. The method is also used to estimate the 253 by 253 precision matrices of two classes of spectra extracted from images taken by the Hubble Space Telescope. Some interesting structural patterns in the estimates are discussed.
Safe reinforcement learning (RL) is crucial for deploying RL agents in real-world applications, as it aims to maximize long-term rewards while satisfying safety constraints. However, safe RL often suffers from sample inefficiency, requiring extensive interactions with the environment to learn a safe policy. We propose Efficient Safe Policy Optimization (ESPO), a novel approach that enhances the efficiency of safe RL through sample manipulation. ESPO employs an optimization framework with three modes: maximizing rewards, minimizing costs, and balancing the trade-off between the two. By dynamically adjusting the sampling process based on the observed conflict between reward and safety gradients, ESPO theoretically guarantees convergence, optimization stability, and improved sample complexity bounds. Experiments on the Safety-MuJoCo and Omnisafe benchmarks demonstrate that ESPO significantly outperforms existing primal-based and primal-dual-based baselines in terms of reward maximization and constraint satisfaction. Moreover, ESPO achieves substantial gains in sample efficiency, requiring 25--29% fewer samples than baselines, and reduces training time by 21--38%.
AI is undergoing a paradigm shift with the rise of models (e.g., BERT, DALL-E, GPT-3) that are trained on broad data at scale and are adaptable to a wide range of downstream tasks. We call these models foundation models to underscore their critically central yet incomplete character. This report provides a thorough account of the opportunities and risks of foundation models, ranging from their capabilities (e.g., language, vision, robotics, reasoning, human interaction) and technical principles(e.g., model architectures, training procedures, data, systems, security, evaluation, theory) to their applications (e.g., law, healthcare, education) and societal impact (e.g., inequity, misuse, economic and environmental impact, legal and ethical considerations). Though foundation models are based on standard deep learning and transfer learning, their scale results in new emergent capabilities,and their effectiveness across so many tasks incentivizes homogenization. Homogenization provides powerful leverage but demands caution, as the defects of the foundation model are inherited by all the adapted models downstream. Despite the impending widespread deployment of foundation models, we currently lack a clear understanding of how they work, when they fail, and what they are even capable of due to their emergent properties. To tackle these questions, we believe much of the critical research on foundation models will require deep interdisciplinary collaboration commensurate with their fundamentally sociotechnical nature.
We propose a novel approach to multimodal sentiment analysis using deep neural networks combining visual analysis and natural language processing. Our goal is different than the standard sentiment analysis goal of predicting whether a sentence expresses positive or negative sentiment; instead, we aim to infer the latent emotional state of the user. Thus, we focus on predicting the emotion word tags attached by users to their Tumblr posts, treating these as "self-reported emotions." We demonstrate that our multimodal model combining both text and image features outperforms separate models based solely on either images or text. Our model's results are interpretable, automatically yielding sensible word lists associated with emotions. We explore the structure of emotions implied by our model and compare it to what has been posited in the psychology literature, and validate our model on a set of images that have been used in psychology studies. Finally, our work also provides a useful tool for the growing academic study of images - both photographs and memes - on social networks.
小貼士
登錄享
注冊地址： 北京市海淀區羊坊店路18號2幢3層301-191