ICLR 2023 Papers

Mixup, which creates synthetic training instances by linearly interpolating random sample pairs, is a simple and yet effective regularization technique to boost the performance of deep models trained with SGD. In this work, we report a previously unobserved phenomenon in Mixup raining: on a number of standard datasets, the performance of Mixup-trained models starts to decay after training for a large number of epochs, giving rise to a U-shaped generalization curve. This behavior is further aggravated when the size of original dataset is reduced. To help understand such a behavior of Mixup, we show theoretically that Mixup training may introduce undesired data-dependent label noises to the synthesized data. Via analyzing a least-square regression problem with a random feature model, we explain why noisy labels may cause the U-shaped curve to occur: Mixup improves generalization through fitting the clean patterns at the early training stage, but as training progresses, Mixup becomes over-fitting to the noise in the synthetic data. Extensive experiments are performed on a variety of benchmark datasets, validating this explanation.

Meta-learning has arisen as a successful method for improving training performance by training over many similar tasks, especially with deep neural networks (DNNs). However, the theoretical understanding of when and why overparameterized models such as DNNs can generalize well in meta-learning is still limited. As an initial step towards addressing this challenge, this paper studies the generalization performance of overfitted meta-learning under a linear regression model with Gaussian features. In contrast to a few recent studies along the same line, our framework allows the number of model parameters to be arbitrarily larger than the number of features in the ground truth signal, and hence naturally captures the overparameterized regime in practical deep meta-learning. We show that the overfitted min $\ell_2$-norm solution of model-agnostic meta-learning (MAML) can be beneficial, which is similar to the recent remarkable findings on "benign overfitting" and "double descent" phenomenon in the classical (single-task) linear regression. However, due to the uniqueness of meta-learning such as task-specific gradient descent inner training and the diversity/fluctuation of the ground-truth signals among training tasks, we find new and interesting properties that do not exist in single-task linear regression. We first provide a high-probability upper bound (under reasonable tightness) on the generalization error, where certain terms decrease when the number of features increases. Our analysis suggests that benign overfitting is more significant and easier to observe when the noise and the diversity/fluctuation of the ground truth of each training task are large. Under this circumstance, we show that the overfitted min $\ell_2$-norm solution can achieve an even lower generalization error than the underparameterized solution.

A number of machine learning models have been proposed with the goal of achieving systematic generalization: the ability to reason about new situations by combining aspects of previous experiences. These models leverage compositional architectures which aim to learn specialized modules dedicated to structures in a task that can be composed to solve novel problems with similar structures. While the compositionality of these architectures is guaranteed by design, the modules specializing is not. Here we theoretically study the ability of network modules to specialize to useful structures in a dataset and achieve systematic generalization. To this end we introduce a minimal space of datasets motivated by practical systematic generalization benchmarks. From this space of datasets we present a mathematical definition of systematicity and study the learning dynamics of linear neural modules when solving components of the task. Our results shed light on the difficulty of module specialization, what is required for modules to successfully specialize, and the necessity of modular architectures to achieve systematicity. Finally, we confirm that the theoretical results in our tractable setting generalize to more complex datasets and non-linear architectures.

While machine learning models rapidly advance the state-of-the-art on various real-world tasks, out-of-domain (OOD) generalization remains a challenging problem given the vulnerability of these models to spurious correlations. We propose a balanced mini-batch sampling strategy to transform a biased data distribution into a spurious-free balanced distribution, based on the invariance of the underlying causal mechanisms for the data generation process. We argue that the Bayes optimal classifiers trained on such balanced distribution are minimax optimal across a diverse enough environment space. We also provide an identifiability guarantee of the latent variable model of the proposed data generation process, when utilizing enough train environments. Experiments are conducted on DomainBed, demonstrating empirically that our method obtains the best performance across 20 baselines reported on the benchmark.

Machine learning models fail to perform when facing out-of-distribution (OOD) domains, a challenging task known as domain generalization (DG). In this work, we develop a novel DG training strategy, we call PGrad, to learn a robust gradient direction, improving models' generalization ability on unseen domains. The proposed gradient aggregates the principal directions of a sampled roll-out optimization trajectory that measures the training dynamics across all training domains. PGrad gradient design forces the DG training to ignore domain-dependent noise signals and updates all training domains with a robust direction covering main components of parameter dynamics. We further improve PGrad via bijection-based computational refinement and directional plus length-based calibrations. Our theoretical proof connects PGrad to the spectral analysis of Hessian in training neural networks. Experiments on DomainBed and WILDS benchmarks demonstrate that our approach effectively enables robust DG optimization and leads to smoothly decreased loss curves. Empirically, PGrad achieves competitive results across seven datasets, demonstrating its efficacy across both synthetic and real-world distributional shifts.

The field of Natural Language Processing (NLP) has experienced a dramatic leap in capabilities with the recent introduction of huge Language Models (LMs). Despite this success, natural language problems that involve several compounded steps are still practically unlearnable, even by the largest LMs. This complies with experimental failures for end-to-end learning of composite problems that were demonstrated in a variety of domains. An effective mitigation is to introduce intermediate supervision for solving sub-tasks of the compounded problem. Recently, several works have demonstrated high gains by taking a straightforward approach for incorporating intermediate supervision in compounded natural language problems: the sequence-to-sequence LM is fed with an augmented input, in which the decomposed tasks' labels are simply concatenated to the original input. In this paper, we prove a positive learning result that motivates these recent efforts. We show that when concatenating intermediate supervision to the input and training a sequence-to-sequence model on this modified input, unlearnable composite problems can become learnable. We show that this is true for any family of tasks which on the one hand, are unlearnable, and on the other hand, can be decomposed into a polynomial number of simple sub-tasks, each of which depends only on $O(1)$ previous sub-task results. Beyond motivating contemporary empirical efforts for incorporating intermediate supervision in sequence-to-sequence language models, our positive theoretical result is the first of its kind in the landscape of results on the benefits of intermediate supervision for neural-network learning: Until now, all theoretical results on the subject are negative, i.e., show cases where learning is impossible without intermediate supervision, while our result is positive, showing that learning is facilitated in the presence of intermediate supervision.

In domain generalization (DG), most existing methods aspire to fine-tune a specific pretrained model through novel DG algorithms. In this paper, we propose an alternative direction, i.e., to efficiently leverage a pool of pretrained models without fine-tuning. Through extensive empirical and theoretical evidence, we demonstrate that (1) pretrained models have possessed generalization to some extent while there is no single best pretrained model across all distribution shifts, and (2) out-of-distribution (OOD) generalization error depends on the fitness between the pretrained model and unseen test distributions. This analysis motivates us to incorporate diverse pretrained models and to dispatch the best matched models for each OOD sample by means of recommendation techniques. To this end, we propose SIMPLE, a specialized model-sample matching method for domain generalization. First, the predictions of pretrained models are adapted to the target domain by a linear label space transformation. A matching network aware of model specialty is then proposed to dynamically recommend proper pretrained models to predict each test sample. The experiments on DomainBed show that our method achieves significant performance improvements (up to 12.2% for individual dataset and 3.9% on average) compared to state-of-the-art (SOTA) methods and further achieves 6.1% gain via enlarging the pretrained model pool. Moreover, our method is highly efficient and achieves more than 1000 times training speedup compared to the conventional DG methods with fine-tuning a pretrained model. Code and supplemental materials are available at https://seqml.github.io/simple.

Generalization analysis of learning algorithms often builds on a critical assumption that training examples are independently and identically distributed, which is often violated in practical problems such as time series prediction. In this paper, we use algorithmic stability to study the generalization performance of learning algorithms with $\psi$-mixing data, where the dependency between observations weakens over time. We show uniformly stable algorithms guarantee high-probability generalization bounds of the order $O(1/\sqrt{n})$ (within a logarithmic factor), where $n$ is the sample size. We apply our general result to specific algorithms including regularization schemes, stochastic gradient descent and localized iterative regularization, and develop excess population risk bounds for learning with $\psi$-mixing data. Our analysis builds on a novel moment bound for weakly-dependent random variables on a $\varphi$-mixing sequence and a novel error decomposition of generalization error.

We study the problem of training a two-layer neural network (NN) of arbitrary width using stochastic gradient descent (SGD) where the input $\boldsymbol{x}\in \mathbb{R}^d$ is Gaussian and the target $y \in \mathbb{R}$ follows a multiple-index model, i.e., $y=g(\langle\boldsymbol{u_1},\boldsymbol{x}\rangle,...,\langle\boldsymbol{u_k},\boldsymbol{x}\rangle)$ with a noisy link function $g$. We prove that the first-layer weights in the NN converge to the $k$-dimensional principal subspace spanned by the vectors $\boldsymbol{u_1},...,\boldsymbol{u_k}$ of the true model, when online SGD with weight decay is used for training. This phenomenon has several important consequences when $k \ll d$. First, by employing uniform convergence on this smaller subspace, we establish a generalization error bound of $\mathcal{O}(\sqrt{{kd}/{T}})$ after $T$ iterations of SGD, which is independent of the width of the NN. We further demonstrate that, by recovering the principal direction, SGD-trained ReLU NNs can learn a single-index target of the form $y=f(\langle\boldsymbol{u},\boldsymbol{x}\rangle) + \epsilon$ with a sample complexity linear in $d$ (up to log factors), where $f$ is a monotonic function with at most polynomial growth, and $\epsilon$ is the noise. This is in contrast to the known $d^{\Omega(p)}$ samples required to learn any degree $p$ polynomial in the kernel regime, and shows that SGD-trained NNs can outperform the Neural Tangent Kernel at initialization. Finally, we establish compressibility guarantees for NNs using that SGD produces an approximately rank-$k$ first-layer weight matrix.

The great success of deep learning heavily relies on increasingly larger training data, which comes at a price of huge computational and infrastructural costs. This poses crucial questions that, do all training data contribute to model's performance? How much does each individual training sample or a sub-training-set affect the model's generalization, and how to construct the smallest subset from the entire training data as a proxy training set without significantly sacrificing the model's performance? To answer these, we propose dataset pruning, an optimization-based sample selection method that can (1) examine the influence of removing a particular set of training samples on model's generalization ability with theoretical guarantee, and (2) construct the smallest subset of training data that yields strictly constrained generalization gap. The empirically observed generalization gap of dataset pruning is substantially consistent with our theoretical expectations. Furthermore, the proposed method prunes 40% training examples on the CIFAR-10 dataset, halves the convergence time with only 1.3% test accuracy decrease, which is superior to previous score-based sample selection methods.

Due to the significant computational challenge of training large-scale graph neural networks (GNNs), various sparse learning techniques have been exploited to reduce memory and storage costs. Examples include graph sparsification that samples a subgraph to reduce the amount of data aggregation and model sparsification that prunes the neural network to reduce the number of trainable weights. Despite the empirical successes in reducing the training cost while maintaining the test accuracy, the theoretical generalization analysis of sparse learning for GNNs remains elusive. To the best of our knowledge, this paper provides the first theoretical characterization of joint edge-model sparse learning from the perspective of sample complexity and convergence rate in achieving zero generalization error. It proves analytically that both sampling important nodes and pruning neurons with lowest-magnitude can reduce the sample complexity and improve convergence without compromising the test accuracy. Although the analysis is centered on two-layer GNNs with structural constraints on data, the insights are applicable to more general setups and justified by both synthetic and practical citation datasets.

We consider the problem of training a deep neural network on a given classification task, e.g., ImageNet-1K (IN1K), so that it excels at both the training task as well as at other (future) transfer tasks. These two seemingly contradictory properties impose a trade-off between improving the model’s generalization and maintaining its performance on the original task. Models trained with self-supervised learning tend to generalize better than their supervised counterparts for transfer learning; yet, they still lag behind supervised models on IN1K. In this paper, we propose a supervised learning setup that leverages the best of both worlds. We extensively analyze supervised training using multi-scale crops for data augmentation and an expendable projector head, and reveal that the design of the projector allows us to control the trade-off between performance on the training task and transferability. We further replace the last layer of class weights with class prototypes computed on the fly using a memory bank and derive two models: t-ReX that achieves a new state of the art for transfer learning and outperforms top methods such as DINO and PAWS on IN1K, and t-ReX* that matches the highly optimized RSB-A1 model on IN1K while performing better on transfer tasks. Code and pretrained models: https://europe.naverlabs.com/t-rex

The \emph{stability} of learning algorithms to changes in the training sample has been actively studied as a powerful proxy for reasoning about generalization. Recently, exponential generalization and excess risk bounds with near-optimal rates have been obtained under the stringent and distribution-free notion of uniform stability~\citep{bousquet2020sharper,klochkov2021stability}. In the meanwhile, under the notion of $L_q$-stability, which is weaker and distribution dependent, exponential generalization bounds are also available yet so far only with sub-optimal rates. Therefore, a fundamental question we would like to address in this paper is whether it is possible to derive near-optimal exponential generalization bounds for $L_q$-stable learning algorithms. As the core contribution of the present work, we give an affirmative answer to this question by developing strict analogues of the near-optimal generalization and risk bounds of uniformly stable algorithms for $L_q$-stable algorithms. Further, we demonstrate the power of our improved $L_q$-stability and generalization theory by applying it to derive strong sparse excess risk bounds, under mild conditions, for computationally tractable sparsity estimation algorithms such as Iterative Hard Thresholding (IHT).

We propose a generic module named Indirection Layer (InLay), which leverages indirection and data internal relationships to effectively construct symbolic indirect representations to improve out-of-distribution generalization capabilities of various neural architectures. InLay receives data input in the form of a sequence of objects, treats it as a complete weighted graph whose vertices are the objects and edge weights are scalars representing relationships between vertices. The input is first mapped via indirection to a symbolic graph with data-independent and trainable vertices. This symbolic graph is then propagated, resulting in new vertex features whose indirection will be used for prediction steps afterward. Theoretically, we show that the distances between indirection representations are bounded by the distances between corresponding graphs, implying that unseen samples with very different surface statistics can still be close in the representation space to the seen samples if they share similar internal relationships. We demonstrate that InLay is consistently effective in improving out-of-distribution generalization throughout a comprehensive suite of experiments, including IQ problems, distorted image classification, and few-shot domain adaptation NLP classification. We also conduct ablation studies to verify different design choices of InLay.

Human languages enable robust generalization, letting us leverage our prior experience to communicate about novel meanings. This is partly due to language being compositional, where the meaning of a whole expression is a function of its parts. Natural languages also exhibit extensive variation, encoding meaning predictably enough to enable generalization without limiting speakers to one and only one way of expressing something. Previous work looking at the languages that emerge between neural networks in a communicative task has shown languages that enable robust communication and generalization reliably emerge. Despite this those languages score poorly on existing measures of compositionality leading to claims that a language's degree of compositionality has little bearing on how well it can generalise. We argue that the languages that emerge between networks are in fact straightforwardly compositional, but with a degree of natural language-like variation that can obscure their compositionality from existing measures. We introduce 4 measures of linguistic variation and show that early in training measures of variation correlate with generalization performance, but that this effect goes away over time as the languages that emerge become regular enough to generalize robustly. Like natural languages, emergent languages appear able to support a high degree of variation while retaining the generalizability we expect from compositionality. In an effort to decrease the variability of emergent languages we show how reducing a model's capacity results in greater regularity, in line with claims about factors shaping the emergence of regularity in human language.