ICLR 2023 Papers

We introduce semi-parametric inducing point networks (SPIN), a general-purpose architecture that can query the training set at inference time in a compute-efficient manner. Semi-parametric architectures are typically more compact than parametric models, but their computational complexity is often quadratic. In contrast, SPIN attains linear complexity via a cross-attention mechanism between datapoints inspired by inducing point methods. Querying large training sets can be particularly useful in meta-learning, as it unlocks additional training signal, but often exceeds the scaling limits of existing models. We use SPIN as the basis of the Inducing Point Neural Process, a probabilistic model which supports large contexts in meta-learning and achieves high accuracy where existing models fail. In our experiments, SPIN reduces memory requirements, improves accuracy across a range of meta-learning tasks, and improves state-of-the-art performance on an important practical problem, genotype imputation.

Fine-tuning pre-trained language models has become the prevalent paradigm for building downstream NLP models. Oftentimes fine-tuned models are readily available but their training data is not, due to data privacy or intellectual property concerns. This creates a barrier to fusing knowledge across individual models to yield a better single model. In this paper, we study the problem of merging individual models built on different training data sets to obtain a single model that performs well both across all data set domains and can generalize on out-of-domain data. We propose a data-less knowledge fusion method that merges models in their parameter space, guided by weights that minimize prediction differences between the merged model and the individual models. Over a battery of evaluation settings, we show that the proposed method significantly outperforms baselines such as Fisher-weighted averaging or model ensembling. Further, we find that our method is a promising alternative to multi-task learning that can preserve or sometimes improve over the individual models without access to the training data. Finally, model merging is more efficient than training a multi-task model, thus making it applicable to a wider set of scenarios.

Representation learning often plays a critical role in avoiding the curse of dimensionality in reinforcement learning. A representative class of algorithms exploits spectral decomposition of the stochastic transition dynamics to construct representations that enjoy strong theoretical properties in idealized settings. However, current spectral methods suffer from limited applicability because they are constructed for state-only aggregation and are derived from a policy-dependent transition kernel, without considering the issue of exploration. To address these issues, we propose an alternative spectral method, Spectral Decomposition Representation (SPEDER), that extracts a state-action abstraction from the dynamics without inducing spurious dependence on the data collection policy, while also balancing the exploration-versus-exploitation trade-off during learning. A theoretical analysis establishes the sample efficiency of the proposed algorithm in both the online and offline settings. In addition, an experimental investigation demonstrates superior performance over current state-of-the-art algorithms across several RL benchmarks.

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.

Learning logical rules is critical to improving reasoning in KGs. This is due to their ability to provide logical and interpretable explanations when used for predictions, as well as their ability to generalize to other tasks, domains, and data. While recent methods have been proposed to learn logical rules, the majority of these methods are either restricted by their computational complexity and can not handle the large search space of large-scale KGs, or show poor generalization when exposed to data outside the training set. In this paper, we propose an end-to-end neural model for learning compositional logical rules called NCRL. NCRL detects the best compositional structure of a rule body, and breaks it into small compositions in order to infer the rule head. By recurrently merging compositions in the rule body with a recurrent attention unit, NCRL finally predicts a single rule head. Experimental results show that NCRL learns high-quality rules, as well as being generalizable. Specifically, we show that NCRL is scalable, efficient, and yields state-of-the-art results for knowledge graph completion on large-scale KGs. Moreover, we test NCRL for systematic generalization by learning to reason on small-scale observed graphs and evaluating on larger unseen ones.

Guided diffusion is a technique for conditioning the output of a diffusion model at sampling time without retraining the network for each specific task. However, one drawback of diffusion models, whether they are guided or unguided, is their slow sampling process. Recent techniques can accelerate unguided sampling by applying high-order numerical methods to the sampling process when viewed as differential equations. On the contrary, we discover that the same techniques do not work for guided sampling, and little has been explored about its acceleration. This paper explores the culprit of this problem and provides a solution based on operator splitting methods, motivated by our key finding that classical high-order numerical methods are unsuitable for the conditional function. Our proposed method can re-utilize the high-order methods for guided sampling and can generate images with the same quality as a 250-step DDIM baseline using 32-58% less sampling time on ImageNet256. We also demonstrate usage on a wide variety of conditional generation tasks, such as text-to-image generation, colorization, inpainting, and super-resolution.

Test-time adaptation (TTA) aims to adapt a trained classifier using online unlabeled test data only, without any information related to the training procedure. Most existing TTA methods adapt the trained classifier using the classifier's prediction on the test data as pseudo-label. However, under test-time domain shift, accuracy of the pseudo labels cannot be guaranteed, and thus the TTA methods often encounter performance degradation at the adapted classifier. To overcome this limitation, we propose a novel test-time adaptation method, called Test-time Adaptation via Self-Training with nearest neighbor information (TAST), which is composed of the following procedures: (1) adds trainable adaptation modules on top of the trained feature extractor; (2) newly defines a pseudo-label distribution for the test data by using the nearest neighbor information; (3) trains these modules only a few times during test time to match the nearest neighbor-based pseudo label distribution and a prototype-based class distribution for the test data; and (4) predicts the label of test data using the average predicted class distribution from these modules. The pseudo-label generation is based on the basic intuition that a test data and its nearest neighbor in the embedding space are likely to share the same label under the domain shift. By utilizing multiple randomly initialized adaptation modules, TAST extracts useful information for the classification of the test data under the domain shift, using the nearest neighbor information. TAST showed better performance than the state-of-the-art TTA methods on two standard benchmark tasks, domain generalization, namely VLCS, PACS, OfficeHome, and TerraIncognita, and image corruption, particularly CIFAR-10/100C.

We provide a convergence analysis of gradient descent for the problem of agnostically learning a single ReLU function under Gaussian distributions. Unlike prior work that studies the setting of zero bias, we consider the more challenging scenario when the bias of the ReLU function is non-zero. Our main result establishes that starting from random initialization, in a polynomial number of iterations gradient descent outputs, with high probability, a ReLU function that achieves an error that is within a constant factor of the optimal i.e., it is guaranteed to achieve an error of $O(OPT)$, where $OPT$ is the error of the best ReLU function. This is a significant improvement over existing guarantees for gradient descent, which only guarantee error of $O(\sqrt{d \cdot OPT})$ even in the zero-bias case (Frei et al., 2020). We also provide finite sample guarantees, and obtain similar guarantees for a broader class of marginal distributions beyond Gaussians.

Federated learning (FL) is a subfield of machine learning where multiple clients try to collaboratively learn a model over a network under communication constraints. We consider finite-sum federated optimization under a second-order function similarity condition and strong convexity, and propose two new algorithms: SVRP and Catalyzed SVRP. This second-order similarity condition has grown popular recently, and is satisfied in many applications including distributed statistical learning and differentially private empirical risk minimization. The first algorithm, SVRP, combines approximate stochastic proximal point evaluations, client sampling, and variance reduction. We show that SVRP is communication efficient and achieves superior performance to many existing algorithms when function similarity is high enough. Our second algorithm, Catalyzed SVRP, is a Catalyst-accelerated variant of SVRP that achieves even better performance and uniformly improves upon existing algorithms for federated optimization under second-order similarity and strong convexity. In the course of analyzing these algorithms, we provide a new analysis of the Stochastic Proximal Point Method (SPPM) that might be of independent interest. Our analysis of SPPM is simple, allows for approximate proximal point evaluations, does not require any smoothness assumptions, and shows a clear benefit in communication complexity over ordinary distributed stochastic gradient descent.

Federated learning aims to train models collaboratively across different clients without sharing data for privacy considerations. However, one major challenge for this learning paradigm is the data heterogeneity problem, which refers to the discrepancies between the local data distributions among various clients. To tackle this problem, we first study how data heterogeneity affects the representations of the globally aggregated models. Interestingly, we find that heterogeneous data results in the global model suffering from severe dimensional collapse, in which representations tend to reside in a lower-dimensional space instead of the ambient space. Moreover, we observe a similar phenomenon on models locally trained on each client and deduce that the dimensional collapse on the global model is inherited from local models. In addition, we theoretically analyze the gradient flow dynamics to shed light on how data heterogeneity result in dimensional collapse for local models. To remedy this problem caused by the data heterogeneity, we propose FedDecorr, a novel method that can effectively mitigate dimensional collapse in federated learning. Specifically, FedDecorr applies a regularization term during local training that encourages different dimensions of representations to be uncorrelated. FedDecorr, which is implementation-friendly and computationally-efficient, yields consistent improvements over baselines on standard benchmark datasets. Code: https://github.com/bytedance/FedDecorr.

The optimal transportation map finds the most economical way to transport one probability measure to another, and it has been applied in a broad range of applications in machine learning and computer vision. By the Brenier theory, computing the optimal transport map is equivalent to solving a Monge-Amp\`ere equation, which is highly non-linear. Therefore, the computation of optimal transportation maps is intrinsically challenging. In this work, we propose a novel and powerful method, the FFT-OT (fast Fourier transform-optimal transport), to compute the 3-dimensional OT problems. The method is based on several key ideas: first, the Monge-Amp\`ere equation is linearized to a sequence of linear elliptic PDEs with spacial and temporal variant coefficients; second, the obliqueness property of optimal transportation maps is reformulated as a Neumann boundary condition; and third, the variant coefficient elliptic PDEs are approximated by constant coefficient elliptic PDEs and solved by FFT on GPUs. We also prove that the algorithm converges linearly, namely the approximation error decreases exponentially fast. Experimental results show that the FFT-OT algorithm is more than a hundred times faster than the conventional methods based on the convex geometry. Furthermore, the method can be directly applied for sampling from complex 3D density functions in machine learning and magnifying the volumetric data in medical imaging.

Partial label learning (PLL) is an important weakly supervised learning problem, where each training instance is associated with a set of candidate labels that include both the true label and additional noisy labels. Most existing PLL methods assume the candidate noisy labels are randomly chosen, which hardly holds in real-world learning scenarios. In this paper, we consider a more realistic PLL scenario with competitive label noise that is more difficult to distinguish from the true label than the random label noise. We propose a novel Mutual Learning based PLL approach named ML-PLL to address this challenging problem. ML-PLL learns a prediction network based classifier and a class-prototype based classifier cooperatively through interactive mutual learning and label correction. Moreover, we use a transformation network to model the association relationships between the true label and candidate labels, and learn it together with the prediction network to match the observed candidate labels in the training data and enhance label correction. Extensive experiments are conducted on several benchmark PLL datasets, and the proposed ML-PLL approach demonstrates state-of-the-art performance for partial label learning.

Existing techniques for training language models can be misaligned with the truth: if we train models with imitation learning, they may reproduce errors that humans make; if we train them to generate text that humans rate highly, they may output errors that human evaluators can't detect. We propose circumventing this issue by directly finding latent knowledge inside the internal activations of a language model in a purely unsupervised way. Specifically, we introduce a method for accurately answering yes-no questions given only unlabeled model activations. It works by finding a direction in activation space that satisfies logical consistency properties, such as that a statement and its negation have opposite truth values. We show that despite using no supervision and no model outputs, our method can recover diverse knowledge represented in large language models: across 6 models and 10 question-answering datasets, it outperforms zero-shot accuracy by 4\% on average. We also find that it cuts prompt sensitivity in half and continues to maintain high accuracy even when models are prompted to generate incorrect answers. Our results provide an initial step toward discovering what language models know, distinct from what they say, even when we don't have access to explicit ground truth labels.

This paper proposes a novel batch normalization strategy for test-time adaptation. Recent test-time adaptation methods heavily rely on the modified batch normalization, i.e., transductive batch normalization (TBN), which calculates the mean and the variance from the current test batch rather than using the running mean and variance obtained from the source data, i.e., conventional batch normalization (CBN). Adopting TBN that employs test batch statistics mitigates the performance degradation caused by the domain shift. However, re-estimating normalization statistics using test data depends on impractical assumptions that a test batch should be large enough and be drawn from i.i.d. stream, and we observed that the previous methods with TBN show critical performance drop without the assumptions. In this paper, we identify that CBN and TBN are in a trade-off relationship and present a new test-time normalization (TTN) method that interpolates the statistics by adjusting the importance between CBN and TBN according to the domain-shift sensitivity of each BN layer. Our proposed TTN improves model robustness to shifted domains across a wide range of batch sizes and in various realistic evaluation scenarios. TTN is widely applicable to other test-time adaptation methods that rely on updating model parameters via backpropagation. We demonstrate that adopting TTN further improves their performance and achieves state-of-the-art performance in various standard benchmarks.

Disentangled learning representations have promising utility in many applications, but they currently suffer from serious reliability issues. We present Gaussian Channel Autoencoder (GCAE), a method which achieves reliable disentanglement via scalable non-parametric density estimation of the latent space. GCAE avoids the curse of dimensionality of density estimation by disentangling subsets of its latent space with the Dual Total Correlation (DTC) metric, thereby representing its high-dimensional latent joint distribution as a collection of many low-dimensional conditional distributions. In our experiments, GCAE achieves highly competitive and reliable disentanglement scores compared with state-of-the-art baselines.

Bias is a common problem inherent in recommender systems, which is entangled with users' preferences and poses a great challenge to unbiased learning. For debiasing tasks, the doubly robust (DR) method and its variants show superior performance due to the double robustness property, that is, DR is unbiased when either imputed errors or learned propensities are accurate. However, our theoretical analysis reveals that DR usually has a large variance. Meanwhile, DR would suffer unexpectedly large bias and poor generalization caused by inaccurate imputed errors and learned propensities, which usually occur in practice. In this paper, we propose a principled approach that can effectively reduce the bias and variance simultaneously for existing DR approaches when the error imputation model is misspecified. In addition, we further propose a novel semi-parametric collaborative learning approach that decomposes imputed errors into parametric and nonparametric parts and updates them collaboratively, resulting in more accurate predictions. Both theoretical analysis and experiments demonstrate the superiority of the proposed methods compared with existing debiasing methods.

Despite recent successes with neural models for sign language translation (SLT), translation quality still lags behind spoken languages because of the data scarcity and modality gap between sign video and text. To address both problems, we investigate strategies for cross-modality representation sharing for SLT. We propose SLTUNET, a simple unified neural model designed to support multiple SLTrelated tasks jointly, such as sign-to-gloss, gloss-to-text and sign-to-text translation. Jointly modeling different tasks endows SLTUNET with the capability to explore the cross-task relatedness that could help narrow the modality gap. In addition, this allows us to leverage the knowledge from external resources, such as abundant parallel data used for spoken-language machine translation (MT). We show in experiments that SLTUNET achieves competitive and even state-of-the-art performance on PHOENIX-2014T and CSL-Daily when augmented with MT data and equipped with a set of optimization techniques. We further use the DGS Corpus for end-to-end SLT for the first time. It covers broader domains with a significantly larger vocabulary, which is more challenging and which we consider to allow for a more realistic assessment of the current state of SLT than the former two. Still, SLTUNET obtains improved results on the DGS Corpus. Code is available at https://github.com/bzhangGo/sltunet.

There is growing interest in understanding how real brains may approximate gradients and how gradients can be used to train neuromorphic chips. However, neither real brains nor neuromorphic chips can perfectly follow the loss gradient, so parameter updates would necessarily use gradient estimators that have some variance and/or bias. Therefore, there is a need to understand better how variance and bias in gradient estimators impact learning dependent on network and task properties. Here, we show that variance and bias can impair learning on the training data, but some degree of variance and bias in a gradient estimator can be beneficial for generalization. We find that the ideal amount of variance and bias in a gradient estimator are dependent on several properties of the network and task: the size and activity sparsity of the network, the norm of the gradient, and the curvature of the loss landscape. As such, whether considering biologically-plausible learning algorithms or algorithms for training neuromorphic chips, researchers can analyze these properties to determine whether their approximation to gradient descent will be effective for learning given their network and task properties.

Sequential data naturally have different lengths in many domains, with some very long sequences. As an important modeling tool, neural attention should capture long-range interaction in such sequences. However, most existing neural attention models admit only short sequences, or they have to employ chunking or padding to enforce a constant input length. Here we propose a simple neural network building block called ChordMixer which can model the attention for long sequences with variable lengths. Each ChordMixer block consists of a position-wise rotation layer without learnable parameters and an element-wise MLP layer. Repeatedly applying such blocks forms an effective network backbone that mixes the input signals towards the learning targets. We have tested ChordMixer on the synthetic adding problem, long document classification, and DNA sequence-based taxonomy classification. The experiment results show that our method substantially outperforms other neural attention models.