论文检索

Deep Learning · Large Language Models

George-Octavian Bărbulescu, Peter Triantafillou

LLMs have been found to memorize training textual sequences and regurgitate verbatim said sequences during text generation time. This fact is known to be the cause of privacy and related (e.g., copyright) problems. Unlearning in LLMs then takes the form of devising new algorithms that will properly deal with these side-effects of memorized data, while not hurting the model's utility. We offer a fresh perspective towards this goal, namely, that each textual sequence to be forgotten should be treated differently when being unlearned based on its degree of memorization within the LLM. We contribute a new metric for measuring unlearning quality, an adversarial attack showing that SOTA algorithms lacking this perspective fail for privacy, and two new unlearning methods based on Gradient Ascent and Task Arithmetic, respectively. A comprehensive performance evaluation across an extensive suite of NLP tasks then mapped the solution space, identifying the best solutions under different scales in model capacities and forget set sizes and quantified the gains of the new approaches.

Optimization · Stochastic

Haitong LIU, Qiang Li, Hoi To Wai

This paper studies the performative prediction problem where a learner aims to minimize the expected loss with a decision-dependent data distribution. Such setting is motivated when outcomes can be affected by the prediction model, e.g., in strategic classification. We consider a state-dependent setting where the data distribution evolves according to an underlying controlled Markov chain. We focus on stochastic derivative free optimization (DFO) where the learner is given access to a loss function evaluation oracle with the above Markovian data. We propose a two-timescale DFO($\lambda$) algorithm that features (i) a sample accumulation mechanism that utilizes every observed sample to estimate the overall gradient of performative risk, and (ii) a two-timescale diminishing step size that balances the rates of DFO updates and bias reduction. Under a general non-convex optimization setting, we show that DFO($\lambda$) requires ${\cal O}( 1 /\epsilon^3)$ samples (up to a log factor) to attain a near-stationary solution with expected squared gradient norm less than $\epsilon > 0$. Numerical experiments verify our analysis.

Applications · Physics

Yi Xiao, LEI BAI, Wei Xue, Hao Chen, Kun Chen, kang chen, Tao Han, Wanli Ouyang

Data-driven weather forecasting models are advancing rapidly, yet they rely on initial states (i.e., analysis states) typically produced by traditional data assimilation algorithms. Four-dimensional variational assimilation (4DVar) is one of the most widely adopted data assimilation algorithms in numerical weather prediction centers; it is accurate but computationally expensive. In this paper, we aim to couple the AI forecasting model, FengWu, with 4DVar to build a self-contained data-driven global weather forecasting framework, FengWu-4DVar. To achieve this, we propose an *AI-embedded* 4DVar algorithm that includes three components: (1) a 4DVar objective function embedded with the FengWu forecasting model and its error representation to enhance efficiency and accuracy; (2) a spherical-harmonic-transform-based (SHT-based) approximation strategy for capturing the horizontal correlation of background error; and (3) an auto-differentiation (AD) scheme for determining the optimal analysis fields. Experimental results show that under the ERA5 simulated observational data with varying proportions and noise levels, FengWu-4DVar can generate accurate analysis fields; remarkably, it has achieved stable self-contained global weather forecasts for an entire year for the first time, demonstrating its potential for real-world applications. Additionally, our framework is approximately 100 times faster than the traditional 4DVar algorithm under similar experimental conditions, highlighting its significant computational efficiency.

Deep Learning · Theory

Yahong Yang, Juncai He

Constructing the architecture of a neural network is a challenging pursuit for the machine learning community, and the dilemma of whether to go deeper or wider remains a persistent question. This paper explores a comparison between deeper neural networks (DeNNs) with a flexible number of layers and wider neural networks (WeNNs) with limited hidden layers, focusing on their optimal generalization error in Sobolev losses. Analytical investigations reveal that the architecture of a neural network can be significantly influenced by various factors, including the number of sample points, parameters within the neural networks, and the regularity of the loss function. Specifically, a higher number of parameters tends to favor WeNNs, while an increased number of sample points and greater regularity in the loss function lean towards the adoption of DeNNs. We ultimately apply this theory to address partial differential equations using deep Ritz and physics-informed neural network (PINN) methods, guiding the design of neural networks.

Deep Learning · Large Language Models

Apoorva Nitsure, Youssef Mroueh, Mattia Rigotti, Kristjan Greenewald, Brian Belgodere, Mikhail Yurochkin, Jiri Navratil, Igor Melnyk, Jarret Ross

We propose a distributional framework for benchmarking socio-technical risks of foundation models with quantified statistical significance. Our approach hinges on a new statistical relative testing based on first and second order stochastic dominance of real random variables. We show that the second order statistics in this test are linked to mean-risk models commonly used in econometrics and mathematical finance to balance risk and utility when choosing between alternatives. Using this framework, we formally develop a risk-aware approach for foundation model selection given guardrails quantified by specified metrics. Inspired by portfolio optimization and selection theory in mathematical finance, we define a metrics portfolio for each model as a means to aggregate a collection of metrics, and perform model selection based on the stochastic dominance of these portfolios. The statistical significance of our tests is backed theoretically by an asymptotic analysis via central limit theorems instantiated in practice via a bootstrap variance estimate. We use our framework to compare various large language models regarding risks related to drifting from instructions and outputting toxic content.

Reinforcement Learning · Deep RL

Stefan Sylvius Wagner Martinez, Stefan Harmeling

In this paper we adopt a representation-centric perspective on exploration in reinforcement learning, viewing exploration fundamentally as a density estimation problem. We investigate the effectiveness of clustering representations for exploration in 3-D environments, based on the observation that the importance of pixel changes between transitions is less pronounced in 3-D environments compared to 2-D environments, where pixel changes between transitions are typically distinct and significant. We propose a method that performs episodic and global clustering on random representations and on pre-trained DINO representations to count states, i.e, estimate pseudo-counts. Surprisingly, even random features can be clustered effectively to count states in 3-D environments, however when these become visually more complex, pre-trained DINO representations are more effective thanks to the pre-trained inductive biases in the representations. Overall, this presents a pathway for integrating pre-trained biases into exploration. We evaluate our approach on the VizDoom and Habitat environments, demonstrating that our method surpasses other well-known exploration methods in these settings.

Optimization

Ankit Vani, Frederick Tung, Gabriel Oliveira, Hossein Sharifi-Noghabi

Despite attaining high empirical generalization, the sharpness of models trained with sharpness-aware minimization (SAM) do not always correlate with generalization error. Instead of viewing SAM as minimizing sharpness to improve generalization, our paper considers a new perspective based on SAM's training dynamics. We propose that perturbations in SAM perform *perturbed forgetting*, where they discard undesirable model biases to exhibit learning signals that generalize better. We relate our notion of forgetting to the information bottleneck principle, use it to explain observations like the better generalization of smaller perturbation batches, and show that perturbed forgetting can exhibit a stronger correlation with generalization than flatness. While standard SAM targets model biases exposed by the steepest ascent directions, we propose a new perturbation that targets biases exposed through the model's outputs. Our output bias forgetting perturbations outperform standard SAM, GSAM, and ASAM on ImageNet, robustness benchmarks, and transfer to CIFAR-10,100, while sometimes converging to sharper regions. Our results suggest that the benefits of SAM can be explained by alternative mechanistic principles that do not require flatness of the loss surface.

Social Aspects · Trustworthy Machine Learning

Chaochao Chen, Jiaming Zhang, Yuyuan Li, Zhongxuan Han

The abundance of free internet data offers unprecedented opportunities for researchers and developers, but it also poses privacy risks. Utilizing data without explicit consent raises critical challenges in protecting personal information.Unlearnable examples have emerged as a feasible protection approach, which renders the data unlearnable, i.e., useless to third parties, by injecting imperceptible perturbations. However, these perturbations only exhibit unlearnable effects on either a particular dataset or label-consistent scenarios, thereby lacking broad applicability. To address both issues concurrently, we propose a universal perturbation generator that harnesses data with concept unlearnability, thereby broadening the scope of unlearnability beyond specific datasets or labels. Specifically, we leverage multi-modal pre-trained models to establish a connection between the data concepts in a shared embedding space. This connection enables the information transformation from image data to text concepts. Consequently, we can align the text embedding using concept-wise discriminant loss, and render the data unlearnable. Extensive experiments conducted on real-world datasets demonstrate the concept unlearnability, i.e., cross-dataset transferability and label-agnostic utility, of our proposed unlearnable examples, as well as their robustness against attacks.

Social Aspects · Trustworthy Machine Learning

Dong Chen, Hongyuan Qu, Guangwu Xu

Privacy attacks and poisoning attacks are two of the thorniest problems in federation learning (FL). Homomorphic encryption (HE), which allows certain mathematical operations to be done in the ciphertext state, provides a way to solve these two problems simultaneously. However, existing Paillier-based and CKKS-based privacy-preserving byzantine-robust FL (PBFL) solutions not only suffer from low efficiency but also expose the final model to the server. Additionally, these methods are limited to one robust aggregation algorithm (AGR) and are therefore vulnerable to AGR-tailored poisoning attacks. In this paper, we present AegisFL, an efficient PBLF system that provides the flexibility to change the AGR. We first observe that the core of the existing advanced AGRs is to calculate the inner products, $L_2$ norms and mean values for vectors. Based on this observation, we tailor a packing scheme for PBFL, which fits perfectly with RLWE-based fully homomorphic encryption. Under this packing scheme, the server only needs to perform one ciphertext multiplication to construct any required AGR, while the global model only belongs to honest clients. Finally, we conduct extensive experiments on different datasets and adversary settings, which also confirm the effectiveness and efficiency of our scheme.

Optimization · Large Scale, Parallel and Distributed

Egor Shulgin, Peter Richtarik

Modern advancements in large-scale machine learning would be impossible without the paradigm of data-parallel distributed computing. Since distributed computing with large-scale models imparts excessive pressure on communication channels, significant recent research has been directed toward co-designing communication compression strategies and training algorithms with the goal of reducing communication costs. While pure data parallelism allows better data scaling, it suffers from poor model scaling properties. Indeed, compute nodes are severely limited by memory constraints, preventing further increases in model size. For this reason, the latest achievements in training giant neural network models also rely on some form of model parallelism. In this work, we take a closer theoretical look at Independent Subnetwork Training (IST), which is a recently proposed and highly effective technique for solving the aforementioned problems. We identify fundamental differences between IST and alternative approaches, such as distributed methods with compressed communication, and provide a precise analysis of its optimization performance on a quadratic model.

Applications · Chemistry and Drug Discovery

Michael Sun, Minghao Guo, Weize Yuan, Veronika Thost, Crystal Owens, Aristotle Grosz, Sharvaa Selvan, Katelyn Zhou, Hassan Mohiuddin, Benjamin Pedretti 等

Recent research in molecular discovery has primarily been devoted to small, drug-like molecules, leaving many similarly important applications in material design without adequate technology. These applications often rely on more complex molecular structures with fewer examples that are carefully designed using known substructures. We propose a data-efficient and interpretable model for representing and reasoning over such molecules in terms of graph grammars that explicitly describe the hierarchical design space featuring motifs to be the design basis. We present a novel representation in the form of random walks over the design space, which facilitates both molecule generation and property prediction. We demonstrate clear advantages over existing methods in terms of performance, efficiency, and synthesizability of predicted molecules, and we provide detailed insights into the method's chemical interpretability.

Optimization · Stochastic

Simon Vary, Pierre Ablin, Bin Gao, P.-A. Absil

Optimization over the set of matrices $X$ that satisfy $X^\top B X = I_p$, referred to as the generalized Stiefel manifold, appears in many applications involving sampled covariance matrices such as the canonical correlation analysis (CCA), independent component analysis (ICA), and the generalized eigenvalue problem (GEVP). Solving these problems is typically done by iterative methods that require a fully formed $B$. We propose a cheap stochastic iterative method that solves the optimization problem while having access only to a random estimates of $B$. Our method does not enforce the constraint in every iteration; instead, it produces iterations that converge to critical points on the generalized Stiefel manifold defined in expectation. The method has lower per-iteration cost, requires only matrix multiplications, and has the same convergence rates as its Riemannian optimization counterparts that require the full matrix $B$. Experiments demonstrate its effectiveness in various machine learning applications involving generalized orthogonality constraints, including CCA, ICA, and the GEVP.

Deep Learning · Large Language Models

Gautier Dagan, Gabriel Synnaeve, Baptiste Roziere

Tokenization is an understudied and often neglected component of modern LLMs. Most published works use a single tokenizer for all experiments, often borrowed from another model, without performing ablations or analysis to optimize tokenization. Moreover, the tokenizer is generally kept unchanged when fine-tuning a base model. In this paper, we show that the size, pre-tokenization regular expression, and training data of a tokenizer can significantly impact the model's generation speed, effective context size, memory usage, and downstream performance. We train specialized Byte-Pair Encoding code tokenizers, and conduct extensive ablations on the impact of tokenizer design on the performance of LLMs for code generation tasks such as HumanEval and MBPP, and provide recommendations for tokenizer hyper-parameters selection and switching the tokenizer in a pre-trained LLM. We perform our experiments on models trained from scratch and from pre-trained models, verifying their applicability to a wide range of use-cases. We find that when fine-tuning on more than 50 billion tokens, we can specialize the tokenizer of a pre-trained LLM to obtain large gains in generation speed and effective context size.

Deep Learning · Large Language Models

Xiao Zhang, Miao Li, Ji Wu

Language models can learn sophisticated language understanding skills from fitting raw text. They also unselectively learn useless corpus statistics and biases, especially during finetuning on domain-specific corpora. In this paper, we propose a simple modification to causal language modeling called conditional finetuning, which performs language modeling conditioned on a context. We show that a context can "explain away" certain corpus statistics and make the model avoid learning them. In this fashion, conditional finetuning achieves selective learning from a corpus, learning knowledge useful for downstream tasks while avoiding learning useless corpus statistics like topic biases. This selective learning effect leads to less forgetting and better stability-plasticity tradeoff in domain finetuning, potentially benefitting lifelong learning with language models.

Deep Learning · Everything Else

Jiaqi Zhang, Joel Jennings, Agrin Hilmkil, Nick Pawlowski, Cheng Zhang, Chao Ma

Foundation models have brought changes to the landscape of machine learning, demonstrating sparks of human-level intelligence across a diverse array of tasks. However, a gap persists in complex tasks such as causal inference, primarily due to challenges associated with intricate reasoning steps and high numerical precision requirements. In this work, we take a first step towards building causally-aware foundation models for treatment effect estimations. We propose a novel, theoretically justified method called Causal Inference with Attention (CInA), which utilizes multiple unlabeled datasets to perform self-supervised causal learning, and subsequently enables zero-shot causal inference on unseen tasks with new data. This is based on our theoretical results that demonstrate the primal-dual connection between optimal covariate balancing and self-attention, facilitating zero-shot causal inference through the final layer of a trained transformer-type architecture. We demonstrate empirically that CInA effectively generalizes to out-of-distribution datasets and various real-world datasets, matching or even surpassing traditional per-dataset methodologies. These results provide compelling evidence that our method has the potential to serve as a stepping stone for the development of causal foundation models.

Miscellaneous Aspects of Machine Learning · Causality

Meyer Scetbon, Joel Jennings, Agrin Hilmkil, Cheng Zhang, Chao Ma

We propose a novel formalism for describing Structural Causal Models (SCMs) as fixed-point problems on causally ordered variables, eliminating the need for Directed Acyclic Graphs (DAGs), and establish the weakest known conditions for their unique recovery given the topological ordering (TO). Based on this, we design a two-stage causal generative model that first infers in a zero-shot manner a valid TO from observations, and then learns the generative SCM on the ordered variables. To infer TOs, we propose to amortize the learning of TOs on synthetically generated datasets by sequentially predicting the leaves of graphs seen during training. To learn SCMs, we design a transformer-based architecture that exploits a new attention mechanism enabling the modeling of causal structures, and show that this parameterization is consistent with our formalism. Finally, we conduct an extensive evaluation of each method individually, and show that when combined, our model outperforms various baselines on generated out-of-distribution problems.

Thomas Pouplin, Alan Jeffares, Nabeel Seedat, M van der Schaar

Constructing valid prediction intervals rather than point estimates is a well-established approach for uncertainty quantification in the regression setting. Models equipped with this capacity output an interval of values in which the ground truth target will fall with some prespecified probability. This is an essential requirement in many real-world applications where simple point predictions' inability to convey the magnitude and frequency of errors renders them insufficient for high-stakes decisions. Quantile regression is a leading approach for obtaining such intervals via the empirical estimation of quantiles in the (non-parametric) distribution of outputs. This method is simple, computationally inexpensive, interpretable, assumption-free, and effective. However, it does require that the specific quantiles being learned are chosen a priori. This results in (a) intervals that are arbitrarily symmetric around the median which is sub-optimal for realistic skewed distributions, or (b) learning an excessive number of intervals. In this work, we propose Relaxed Quantile Regression (RQR), a direct alternative to quantile regression based interval construction that removes this arbitrary constraint whilst maintaining its strengths. We demonstrate that this added flexibility results in intervals with an improvement in desirable qualities (e.g. mean width) whilst retaining the essential coverage guarantees of quantile regression.

Deep Learning · Theory

Yajie Bao, Michael Crawshaw, Mingrui Liu

Local steps are crucial for Federated Learning (FL) algorithms and have witnessed great empirical success in reducing communication costs and improving the generalization performance of deep neural networks. However, there are limited studies on the effect of local steps on heterogeneous FL. A few works investigate this problem from the optimization perspective. Woodworth et al. (2020a) showed that the iteration complexity of Local SGD, the most popular FL algorithm, is dominated by the baseline mini-batch SGD, which does not show the benefits of local steps. In addition, Levy (2023) proposed a new local update method that provably benefits over mini-batch SGD. However, in the same setting, there is still no work analyzing the effects of local steps to generalization in a heterogeneous FL setting. Motivated by our experimental findings where Local SGD learns more distinguishing features than parallel SGD, this paper studies the generalization benefits of local steps from a feature learning perspective. We propose a novel federated data model that exhibits a new form of data heterogeneity, under which we show that a convolutional neural network (CNN) trained by GD with *global* updates will miss some pattern-related features, while the network trained by GD with *local* updates can learn all features in polynomial time. Consequently, local steps help CNN generalize better in our data model. In a different parameter setting, we also prove that Local GD with *one-shot* model averaging can learn all features and generalize well in all clients. Our experimental results also confirm the benefits of local steps in improving test accuracy on real-world data.

Deep Learning · Generative Models and Autoencoders

Ben Lonnqvist, Zhengqing Wu, Michael Herzog

Humans are able to segment images effortlessly without supervision using perceptual grouping. Here, we propose a counter-intuitive computational approach to solving unsupervised perceptual grouping and segmentation: that they arise *because* of neural noise, rather than in spite of it. We (1) mathematically demonstrate that under realistic assumptions, neural noise can be used to separate objects from each other; (2) that adding noise in a DNN enables the network to segment images even though it was never trained on any segmentation labels; and (3) that segmenting objects using noise results in segmentation performance that aligns with the perceptual grouping phenomena observed in humans, and is sample-efficient. We introduce the Good Gestalt (GG) datasets --- six datasets designed to specifically test perceptual grouping, and show that our DNN models reproduce many important phenomena in human perception, such as illusory contours, closure, continuity, proximity, and occlusion. Finally, we (4) show that our model improves performance on our GG datasets compared to other tested unsupervised models by $24.9$%. Together, our results suggest a novel unsupervised segmentation method requiring few assumptions, a new explanation for the formation of perceptual grouping, and a novel potential benefit of neural noise.

Theory · Probabilistic Methods

Aditya Chattopadhyay, Benjamin Haeffele, Rene Vidal, Donald Geman

In many applications like experimental design, group testing, and medical diagnosis, the state of a random variable $Y$ is revealed by successively observing the outcomes of binary tests about $Y$. New tests are selected adaptively based on the history of outcomes observed so far. If the number of states of $Y$ is finite, the process ends when $Y$ can be predicted with a desired level of confidence or all available tests have been used. Finding the strategy that minimizes the expected number of tests needed to predict $Y$ is virtually impossible in most real applications. Therefore, the commonly used strategy is the greedy heuristic of Information Maximization (InfoMax) that selects tests sequentially in order of information gain. Despite its widespread use, existing guarantees on its performance are often vacuous when compared to its empirical efficiency. In this paper, for the first time to the best of our knowledge, we establish tight non-vacuous bounds on InfoMax's performance. Our analysis is based on the assumption that at any iteration of the greedy strategy, there is always a binary test available whose conditional probability of being 'true', given the history, is within $\delta$ units of one-half. This assumption is motivated by practical applications where the available set of tests often satisfies this property for modest values of $\delta$, say, ${0.1 \leq \delta \leq 0.4}$. Specifically, we analyze two distinct scenarios: (i) all tests are functions of $Y$, and (ii) test outcomes are corrupted by a binary symmetric channel. For both cases, our bounds guarantee the near-optimal performance of InfoMax for modest $\delta$ values. It requires only a small multiplicative factor of the entropy of $Y$, in terms of the average number of tests needed to make accurate predictions.