ICML 2023 论文列表

We introduce a self-supervised learning (SSL) framework BEATs for general audio representation pre-training, where we optimize an acoustic tokenizer and an audio SSL model by iterations. Unlike the previous audio SSL models that employ reconstruction loss for pre-training, our audio SSL model is trained with the discrete label prediction task, where the labels are generated by a semantic-rich acoustic tokenizer. We propose an iterative pipeline to jointly optimize the tokenizer and the pre-trained model, aiming to abstract high-level semantics and discard the redundant details for audio. The experimental results demonstrate our acoustic tokenizers can generate discrete labels with rich audio semantics and our audio SSL models achieve state-of-the-art (SOTA) results across various audio classification benchmarks, even outperforming previous models that use more training data and model parameters significantly. Specifically, we set a new SOTA mAP 50.6% on AudioSet-2M without using any external data, and 98.1% accuracy on ESC-50. The code and pre-trained models are available at https://aka.ms/beats.

The training process of generative adversarial networks (GANs) is unstable and does not converge globally. In this paper, we examine the stability of GANs from the perspective of control theory and propose a universal higher-order noise-based controller called Brownian Motion Controller (BMC). Starting with the prototypical case of Dirac-GANs, we design a BMC to retrieve precisely the same but reachable optimal equilibrium. We theoretically prove that the training process of DiracGANs-BMC is globally exponential stable and derive bounds on the rate of convergence. Then we extend our BMC to normal GANs and provide implementation instructions on GANs-BMC. Our experiments show that our GANs-BMC effectively stabilizes GANs' training under StyleGANv2-ada frameworks with a faster rate of convergence, a smaller range of oscillation, and better performance in terms of FID score.

Differentially Private methods for training Deep Neural Networks (DNNs) have progressed recently, in particular with the use of massive batches and aggregated data augmentations for a large number of training steps. These techniques require much more computing resources than their non-private counterparts, shifting the traditional privacy-accuracy trade-off to a privacy-accuracy-compute trade-off and making hyper-parameter search virtually impossible for realistic scenarios. In this work, we decouple privacy analysis and experimental behavior of noisy training to explore the trade-off with minimal computational requirements. We first use the tools of Renyi Differential Privacy (RDP) to highlight that the privacy budget, when not overcharged, only depends on the total amount of noise (TAN) injected throughout training. We then derive scaling laws for training models with DP-SGD to optimize hyper-parameters with more than a $100\times$ reduction in computational budget. We apply the proposed method on CIFAR-10 and ImageNet and, in particular, strongly improve the state-of-the-art on ImageNet with a $+9$ points gain in top-1 accuracy for a privacy budget $\varepsilon=8$.

We propose a new dataset distillation algorithm using reparameterization and convexification of implicit gradients (RCIG), that substantially improves the state-of-the-art. To this end, we first formulate dataset distillation as a bi-level optimization problem. Then, we show how implicit gradients can be effectively used to compute meta-gradient updates. We further equip the algorithm with a convexified approximation that corresponds to learning on top of a frozen finite-width neural tangent kernel. Finally, we improve bias in implicit gradients by parameterizing the neural network to enable analytical computation of final-layer parameters given the body parameters. RCIG establishes the new state-of-the-art on a diverse series of dataset distillation tasks. Notably, with one image per class, on resized ImageNet, RCIG sees on average a 108% improvement over the previous state-of-the-art distillation algorithm. Similarly, we observed a 66% gain over SOTA on Tiny-ImageNet and 37% on CIFAR-100.

Activation compressed training provides a solution towards reducing the memory cost of training deep neural networks (DNNs). However, state-of-the-art work combines a search of quantization bit-width with the training, which makes the procedure complicated and less transparent. To this end, we propose a simple and effective method to compress DNN training. Our method is motivated by an instructive observation: DNN backward propagation mainly utilizes the low-frequency component (LFC) of the activation maps, while the majority of memory is for caching the high-frequency component (HFC) during the training. This indicates the HFC of activation maps is highly redundant and compressible, which inspires our proposed Dual Activation Precision (DIVISION). During the training, DIVISION preserves a high-precision copy of LFC and compresses the HFC into a light-weight copy with low numerical precision. This can significantly reduce the memory cost while maintaining a competitive model accuracy. Experiment results show DIVISION has better comprehensive performance than state-of-the-art methods, including over 10x compression of activation maps and competitive training throughput, without loss of model accuracy. The source code is available at https://github.com/guanchuwang/division.

In this paper, we present an effective and efficient structural inference approach by integrating a Reservoir Computing (RC) network into a Variational Auto-encoder-based (VAE-based) structural inference framework. With the help of Bi-level Optimization, the backbone VAE-based method follows the Information Bottleneck principle and infers a general adjacency matrix in its latent space; the RC net substitutes the partial role of the decoder and encourages the whole approach to perform further steps of gradient descent based on limited available data. The experimental results on various datasets including biological networks, simulated fMRI data, and physical simulations show the effectiveness and efficiency of our proposed method for structural inference, either with much fewer trajectories or with much shorter trajectories compared with previous works.

Mean-field games have been used as a theoretical tool to obtain an approximate Nash equilibrium for symmetric and anonymous $N$-player games. However, limiting applicability, existing theoretical results assume variations of a ``population generative model'', which allows arbitrary modifications of the population distribution by the learning algorithm. Moreover, learning algorithms typically work on abstract simulators with population instead of the $N$-player game. Instead, we show that $N$ agents running policy mirror ascent converge to the Nash equilibrium of the regularized game within $\widetilde{\mathcal{O}}(\varepsilon^{-2})$ samples from a single sample trajectory without a population generative model, up to a standard $\mathcal{O}(\frac{1}{\sqrt{N}})$ error due to the mean field. Taking a divergent approach from the literature, instead of working with the best-response map we first show that a policy mirror ascent map can be used to construct a contractive operator having the Nash equilibrium as its fixed point. We analyze single-path TD learning for $N$-agent games, proving sample complexity guarantees by only using a sample path from the $N$-agent simulator without a population generative model. Furthermore, we demonstrate that our methodology allows for independent learning by $N$ agents with finite sample guarantees.

We present unit scaling, a paradigm for designing deep learning models that simplifies the use of low-precision number formats. Training in FP16 or the recently proposed FP8 formats offers substantial efficiency gains, but can lack sufficient range for out-of-the-box training. Unit scaling addresses this by introducing a principled approach to model numerics: seeking unit variance of all weights, activations and gradients at initialisation. Unlike alternative methods, this approach neither requires multiple training runs to find a suitable scale nor has significant computational overhead. We demonstrate the efficacy of unit scaling across a range of models and optimisers. We further show that existing models can be adapted to be unit-scaled, training BERT-Large in FP16 and then FP8 with no degradation in accuracy.

Particle gradient descent, which uses particles to represent a probability measure and performs gradient descent on particles in parallel, is widely used to optimize functions of probability measures. This paper considers particle gradient descent with a finite number of particles and establishes its theoretical guarantees to optimize functions that are *displacement convex* in measures. Concretely, for Lipschitz displacement convex functions defined on probability over $R^d$, we prove that $O(1/\epsilon^2)$ particles and $O(d/\epsilon^4)$ iterations are sufficient to find the $\epsilon$-optimal solutions. We further provide improved complexity bounds for optimizing smooth displacement convex functions. An application of our results proves the conjecture of *no optimization-barrier up to permutation invariance*, proposed by Entezari et al. (2022), for specific two-layer neural networks with two-dimensional inputs uniformly drawn from unit circle.

Population Based Training (PBT) is an efficient hyperparameter optimization algorithm. PBT is a single-objective algorithm, but many real-world hyperparameter optimization problems involve two or more conflicting objectives. In this work, we therefore introduce a multi-objective version of PBT, MO-PBT. Our experiments on diverse multi-objective hyperparameter optimization problems (Precision/Recall, Accuracy/Fairness, Accuracy/Adversarial Robustness) show that MO-PBT outperforms random search, single-objective PBT, and the state-of-the-art multi-objective hyperparameter optimization algorithm MO-ASHA.

The studies on adversarial attacks and defenses have greatly improved the robustness of Deep Neural Networks (DNNs). Most advanced approaches have been overwhelmingly designed for continuous data such as images. However, these achievements are still hard to be generalized to categorical data. To bridge this gap, we propose a novel framework, Probabilistic Categorical Adversarial Attack (or PCAA). It transfers the discrete optimization problem of finding categorical adversarial examples to a continuous problem that can be solved via gradient-based methods. We analyze the optimality (attack success rate) and time complexity of PCAA to demonstrate its significant advantage over current search-based attacks. More importantly, through extensive empirical studies, we demonstrate that the well-established defenses for continuous data, such as adversarial training and TRADES, can be easily accommodated to defend DNNs for categorical data.

We study semiparametrically efficient estimation in off-policy evaluation (OPE) where the underlying Markov decision process (MDP) is linear with a known feature map. We characterize the variance lower bound for regular estimators in the linear MDP setting and propose an efficient estimator whose variance achieves that lower bound. Consistency and asymptotic normality of our estimator are established under mild conditions, which merely requires the only infinite-dimensional nuisance parameter to be estimated at a $n^{-1/4}$ convergence rate. We also construct an asymptotically valid confidence interval for statistical inference and conduct simulation studies to validate our results. To our knowledge, this is the first work that concerns efficient estimation in the presence of a known structure of MDPs in the OPE literature.

We present the OMG-CMDP! algorithm for regret minimization in adversarial Contextual MDPs. The algorithm operates under the minimal assumptions of realizable function class and access to online least squares and log loss regression oracles. Our algorithm is efficient (assuming efficient online regression oracles), simple and robust to approximation errors. It enjoys an $\widetilde{O}(H^{2.5} \sqrt{ T|S||A| ( \mathcal{R}\_{TH}(\mathcal{O}) + H \log(\delta^{-1}) )})$ regret guarantee, with $T$ being the number of episodes, $S$ the state space, $A$ the action space, $H$ the horizon and $\mathcal{R}\_{TH}(\mathcal{O}) = \mathcal{R}\_{TH}(\mathcal{O}\_{sq}^\mathcal{F}) + \mathcal{R}\_{TH}(\mathcal{O}\_{log}^\mathcal{P})$ is the sum of the square and log-loss regression oracles' regret, used to approximate the context-dependent rewards and dynamics, respectively. To the best of our knowledge, our algorithm is the first efficient rate optimal regret minimization algorithm for adversarial CMDPs that operates under the minimal standard assumption of online function approximation.

It is often useful to compactly summarize important properties of model parameters and training data so that they can be used later without storing and/or iterating over the entire dataset. As a specific case, we consider estimating the Function Space Distance (FSD) over a training set, i.e. the average discrepancy between the outputs of two neural networks. We propose a Linearized Activation Function TRick (LAFTR) and derive an efficient approximation to FSD for ReLU neural networks. The key idea is to approximate the architecture as a linear network with stochastic gating. Despite requiring only one parameter per unit of the network, our approach outcompetes other parametric approximations with larger memory requirements. Applied to continual learning, our parametric approximation is competitive with state-of-the-art nonparametric approximations, which require storing many training examples. Furthermore, we show its efficacy in estimating influence functions accurately and detecting mislabeled examples without expensive iterations over the entire dataset.

While leveraging additional training data is well established to improve adversarial robustness, it incurs the unavoidable cost of data collection and the heavy computation to train models. To mitigate the costs, we propose *Guided Adversarial Training * (GAT), a novel adversarial training technique that exploits auxiliary tasks under a limited set of training data. Our approach extends single-task models into multi-task models during the min-max optimization of adversarial training, and drives the loss optimization with a regularization of the gradient curvature across multiple tasks. GAT leverages two types of auxiliary tasks: self-supervised tasks, where the labels are generated automatically, and domain-knowledge tasks, where human experts provide additional labels. Experimentally, under limited data, GAT increases the robust accuracy on CIFAR-10 up to four times (from 11% to 42% robust accuracy) and the robust AUC of CheXpert medical imaging dataset from 50% to 83%. On the full CIFAR-10 dataset, GAT outperforms eight state-of-the-art adversarial training strategies. Our large study across five datasets and six tasks demonstrates that task augmentation is an efficient alternative to data augmentation, and can be key to achieving both clean and robust performances.

Recent trends in language modeling have focused on increasing performance through scaling, and have resulted in an environment where training language models is out of reach for most researchers and practitioners. While most in the community are asking how to push the limits of extreme computation, we ask the opposite question: How far can we get with a single GPU in just one day? We investigate the downstream performance achievable with a transformer-based language model trained completely from scratch with masked language modeling for a single day on a single consumer GPU. Aside from re-analyzing nearly all components of the pretraining pipeline for this scenario and providing a modified pipeline with performance close to BERT, we investigate why scaling down is hard, and which modifications actually improve performance in this scenario. We provide evidence that even in this constrained setting, performance closely follows scaling laws observed in large-compute settings. Through the lens of scaling laws, we categorize a range of recent improvements to training and architecture and discuss their merit and practical applicability (or lack thereof) for the limited compute setting. We provide code to reproduce all experiments at github.com/JonasGeiping/cramming .

We introduce a value-based RL agent, which we call BBF, that achieves super-human performance in the Atari 100K benchmark. BBF relies on scaling the neural networks used for value estimation, as well as a number of other design choices that enable this scaling in a sample-efficient manner. We conduct extensive analyses of these design choices and provide insights for future work. We end with a discussion about updating the goalposts for sample-efficient RL research on the ALE. We make our code and data publicly available at https://github.com/google-research/google-research/tree/master/bigger_better_faster.

Evidential deep learning, built upon belief theory and subjective logic, offers a principled and computationally efficient way to turn a deterministic neural network uncertainty-aware. The resultant evidential models can quantify fine-grained uncertainty using the learned evidence. To ensure theoretically sound evidential models, the evidence needs to be non-negative, which requires special activation functions for model training and inference. This constraint often leads to inferior predictive performance compared to standard softmax models, making it challenging to extend them to many large-scale datasets. To unveil the real cause of this undesired behavior, we theoretically investigate evidential models and identify a fundamental limitation that explains the inferior performance: existing evidential activation functions create *zero evidence regions*, which prevent the model to learn from training samples falling into such regions. A deeper analysis of evidential activation functions based on our theoretical underpinning inspires the design of a novel regularizer that effectively alleviates this fundamental limitation. Extensive experiments over many challenging real-world datasets and settings confirm our theoretical findings and demonstrate the effectiveness of our proposed approach.

Training data and model sizes are increasing exponentially. One way to reduce training time and resources is to train with a carefully selected subset of the full dataset. Prior work uses the gradient signals obtained during a warm-up or ``pre-training" phase over the full dataset, for determining the core subset; if the pre-training phase is too small, the gradients obtained are chaotic and unreliable. As a result, the pre-training phase itself incurs significant time/resource overhead, and prior work has not gone beyond hyperparameter search to reduce pre-training time. Our work explicitly aims to reduce this $\textbf{pre-training tax}$ in gradient-based subset training. We develop a principled, scalable approach for pre-training in a distributed setup. Our approach is $\textit{lightweight}$ and $\textit{minimizes communication}$ between distributed worker nodes. It is the first to utilize the concept of model-soup based distributed training $\textit{at initialization}$. The key idea is to minimally train an ensemble of models on small, disjointed subsets of the data; we further employ data-driven sparsity and data augmentation for local worker training to boost ensemble diversity. The centralized model, obtained at the end of pre-training by merging the per-worker models, is found to offer stabilized gradient signals to select subsets, on which the main model is further trained. We have validated the effectiveness of our method through extensive experiments on CIFAR-10/100, and ImageNet, using ResNet and WideResNet models. For example, our approach is shown to achieve $\textbf{15.4$\times$}$ pre-training speedup and $\textbf{2.8$\times$}$ end-to-end speedup on CIFAR10 and ResNet18 without loss of accuracy. The code is at https://github.com/moonbucks/LiPT.git.