ICLR 2023 Papers

In the era of deep learning, transferring information from a pretrained network to a downstream task by finetuning has many benefits. The choice of task head plays an important role in fine-tuning, as the pretrained and downstream tasks are usually different. Although there exist many different designs for finetuning, a full understanding of when and why these algorithms work has been elusive. We analyze how the choice of task head controls feature adaptation and hence influences the downstream performance. By decomposing the feature's learning dynamics, we find the key aspect is the training accuracy and loss at the beginning of finetuning, which determines the "energy" available for the feature's adaptation. We identify a significant trend in the effect of changes in this initial energy on the resulting features after finetuning. Specifically, as the energy increases, the Euclidean and cosine distances between the resulting and original features increase, while their dot product (and the resulting features’ norm) first increases and then decreases. Inspired by this, we give several practical principles that lead to better downstream performance. We analytically prove this trend in an overparamterized linear setting and verify its applicability to different experimental settings.

In this paper we empirically investigate the conjecture from Entezari et al. (2021) which states that if permutation invariance is taken into account, then there should be no loss barrier to the linear interpolation between SGD solutions. We conduct our investigation using standard computer vision architectures trained on CIFAR-10 and ImageNet. First, we observe a general phenomenon in which interpolated deep networks suffer a collapse in the variance of their activations. We demonstrate that an appropriate rescaling of the pre-activations of the interpolated networks ameliorates this problem and significantly reduces the barrier. Second, by combining this with an algorithm for finding permutations based on maximizing correlations between the activations of matched neurons, we are able to reduce the interpolation barrier for a standard ResNet18 trained on CIFAR-10 to 1.5% absolute test error. We explore the interaction between our method and the choice of normalization layer, and demonstrate its robustness across a variety of architectures and training sets.

Image steganography is the process of concealing secret information in images through imperceptible changes. Recent work has formulated this task as a classic constrained optimization problem. In this paper, we argue that image steganography is inherently performed on the (elusive) manifold of natural images, and propose an iterative neural network trained to perform the optimization steps. In contrast to classical optimization methods like L-BFGS or projected gradient descent, we train the neural network to also stay close to the manifold of natural images throughout the optimization. We show that our learned neural optimization is faster and more reliable than classical optimization approaches. In comparison to previous state-of-the-art encoder-decoder based steganography methods, it reduces the recovery error rate by multiple orders of magnitude and achieves zero error up to 3 bits per pixel (bpp) without the need for error-correcting codes.

Modern semantic segmentation methods devote much effect to adjusting image feature representations to improve the segmentation performance in various ways, such as architecture design, attention mechnism, etc. However, almost all those methods neglect the particularity of class weights (in the classification layer) in segmentation models. In this paper, we notice that the class weights of categories that tend to share many adjacent boundary pixels lack discrimination, thereby limiting the performance. We call this issue Boundary-caused Class Weights Confusion (BCWC). We try to focus on this problem and propose a novel method named Embedded Conditional Random Field (E-CRF) to alleviate it. E-CRF innovatively fuses the CRF into the CNN network as an organic whole for more effective end-to-end optimization. The reasons are two folds. It utilizes CRF to guide the message passing between pixels in high-level features to purify the feature representation of boundary pixels, with the help of inner pixels belonging to the same object. More importantly, it enables optimizing class weights from both scale and direction during backpropagation. We make detailed theoretical analysis to prove it. Besides, superpixel is integrated into E-CRF and served as an auxiliary to exploit the local object prior for more reliable message passing. Finally, our proposed method yields impressive results on ADE20K, Cityscapes, and Pascal Context datasets.

Object rearrangement is a challenge for embodied agents because solving these tasks requires generalizing across a combinatorially large set of configurations of entities and their locations. Worse, the representations of these entities are unknown and must be inferred from sensory percepts. We present a hierarchical abstraction approach to uncover these underlying entities and achieve combinatorial generalization from unstructured visual inputs. By constructing a factorized transition graph over clusters of entity representations inferred from pixels, we show how to learn a correspondence between intervening on states of entities in the agent's model and acting on objects in the environment. We use this correspondence to develop a method for control that generalizes to different numbers and configurations of objects, which outperforms current offline deep RL methods when evaluated on simulated rearrangement tasks.

Most approaches for self-supervised learning (SSL) are optimised on curated balanced datasets, e.g. ImageNet, despite the fact that natural data usually exhibits long-tail distributions. In this paper, we analyse the behaviour of one of the most popular variants of SSL, i.e. contrastive methods, on imbalanced data. In particular, we investigate the role of the temperature parameter $\tau$ in the contrastive loss, by analysing the loss through the lens of average distance maximisation, and find that a large $\tau$ emphasises group-wise discrimination, whereas a small $\tau$ leads to a higher degree of instance discrimination. While $\tau$ has thus far been treated exclusively as a constant hyperparameter, in this work, we propose to employ a dynamic $\tau$ and show that a simple cosine schedule can yield significant improvements in the learnt representations. Such a schedule results in a constant `task switching' between an emphasis on instance discrimination and group-wise discrimination and thereby ensures that the model learns both group-wise features, as well as instance-specific details. Since frequent classes benefit from the former, while infrequent classes require the latter, we find this method to consistently improve separation between the classes in long-tail data without any additional computational cost.

Despite the widespread use of unsupervised models, very few methods are designed to explain them. Most explanation methods explain a scalar model output. However, unsupervised models output representation vectors, the elements of which are not good candidates to explain because they lack semantic meaning. To bridge this gap, recent works defined a scalar explanation output: a dot product-based similarity in the representation space to the sample being explained (i.e., an explicand). Although this enabled explanations of unsupervised models, the interpretation of this approach can still be opaque because similarity to the explicand's representation may not be meaningful to humans. To address this, we propose contrastive corpus similarity, a novel and semantically meaningful scalar explanation output based on a reference corpus and a contrasting foil set of samples. We demonstrate that contrastive corpus similarity is compatible with many post-hoc feature attribution methods to generate COntrastive COrpus Attributions (COCOA) and quantitatively verify that features important to the corpus are identified. We showcase the utility of COCOA in two ways: (i) we draw insights by explaining augmentations of the same image in a contrastive learning setting (SimCLR); and (ii) we perform zero-shot object localization by explaining the similarity of image representations to jointly learned text representations (CLIP).

Privacy and security concerns in real-world applications have led to the development of adversarially robust federated models. However, the straightforward combination between adversarial training and federated learning in one framework can lead to the undesired robustness deterioration. We discover that the attribution behind this phenomenon is that the generated adversarial data could exacerbate the data heterogeneity among local clients, making the wrapped federated learning perform poorly. To deal with this problem, we propose a novel framework called Slack Federated Adversarial Training (SFAT), assigning the client-wise slack during aggregation to combat the intensified heterogeneity. Theoretically, we analyze the convergence of the proposed method to properly relax the objective when combining federated learning and adversarial training. Experimentally, we verify the rationality and effectiveness of SFAT on various benchmarked and real-world datasets with different adversarial training and federated optimization methods. The code is publicly available at: https://github.com/ZFancy/SFAT.

Offline reinforcement learning suffers from out-of-distribution issue and extrapolation error. Most methods penalize the out-of-distribution state-action pairs or regularize the trained policy towards the behavior policy but cannot guarantee to get rid of extrapolation error. We propose In-sample Actor Critic (IAC) which utilizes sampling-importance resampling to execute in-sample policy evaluation. IAC only uses the target Q-values of the actions in the dataset to evaluate the trained policy, thus avoiding extrapolation error. The proposed method performs unbiased policy evaluation and has a lower variance than importance sampling in many cases. Empirical results show that IAC obtains competitive performance compared to the state-of-the-art methods on Gym-MuJoCo locomotion domains and much more challenging AntMaze domains.

Minimum Description Length (MDL) provides a framework and an objective for principled model evaluation. It formalizes Occam's Razor and can be applied to data from non-stationary sources. In the prequential formulation of MDL, the objective is to minimize the cumulative next-step log-loss when sequentially going through the data and using previous observations for parameter estimation. It thus closely resembles a continual- or online-learning problem. In this study, we evaluate approaches for computing prequential description lengths for image classification datasets with neural networks. Considering the computational cost, we find that online-learning with rehearsal has favorable performance compared to the previously widely used block-wise estimation. We propose forward-calibration to better align the models predictions with the empirical observations and introduce replay-streams, a minibatch incremental training technique to efficiently implement approximate random replay while avoiding large in-memory replay buffers. As a result, we present description lengths for a suite of image classification datasets that improve upon previously reported results by large margins.

We study the implicit regularization of gradient descent towards structured sparsity via a novel neural reparameterization, which we call a diagonally grouped linear neural network. We show the following intriguing property of our reparameterization: gradient descent over the squared regression loss, without any explicit regularization, biases towards solutions with a group sparsity structure. In contrast to many existing works in understanding implicit regularization, we prove that our training trajectory cannot be simulated by mirror descent. We analyze the gradient dynamics of the corresponding regression problem in the general noise setting and obtain minimax-optimal error rates. Compared to existing bounds for implicit sparse regularization using diagonal linear networks, our analysis with the new reparameterization shows improved sample complexity. In the degenerate case of size-one groups, our approach gives rise to a new algorithm for sparse linear regression. Finally, we demonstrate the efficacy of our approach with several numerical experiments.

Scaling transformers has led to significant breakthroughs in many domains, leading to a paradigm in which larger versions of existing models are trained and released on a periodic basis. New instances of such models are typically trained completely from scratch, despite the fact that they are often just scaled-up versions of their smaller counterparts. How can we use the implicit knowledge in the parameters of smaller, extant models to enable faster training of newer, larger models? This paper describes an approach for accelerating transformer training by learning to grow pretrained transformers, where we learn to linearly map the parameters of the smaller model to initialize the larger model. For tractable learning, we factorize the linear transformation as a composition of (linear) width- and depth-growth operators, and further employ a Kronecker factorization of these growth operators to encode architectural knowledge. Extensive experiments across both language and vision transformers demonstrate that our learned Linear Growth Operator (LiGO) can save up to 50% computational cost of training from scratch, while also consistently outperforming strong baselines that also reuse smaller pretrained models to initialize larger models.

Soft threshold pruning is among the cutting-edge pruning methods with state-of-the-art performance. However, previous methods either perform aimless searching on the threshold scheduler or simply set the threshold trainable, lacking theoretical explanation from a unified perspective. In this work, we reformulate soft threshold pruning as an implicit optimization problem solved using the Iterative Shrinkage-Thresholding Algorithm (ISTA), a classic method from the fields of sparse recovery and compressed sensing. Under this theoretical framework, all threshold tuning strategies proposed in previous studies of soft threshold pruning are concluded as different styles of tuning $L_1$-regularization term. We further derive an optimal threshold scheduler through an in-depth study of threshold scheduling based on our framework. This scheduler keeps $L_1$-regularization coefficient stable, implying a time-invariant objective function from the perspective of optimization. In principle, the derived pruning algorithm could sparsify any mathematical model trained via SGD. We conduct extensive experiments and verify its state-of-the-art performance on both Artificial Neural Networks (ResNet-50 and MobileNet-V1) and Spiking Neural Networks (SEW ResNet-18) on ImageNet datasets. On the basis of this framework, we derive a family of pruning methods, including sparsify-during-training, early pruning, and pruning at initialization. The code is available at https://github.com/Yanqi-Chen/LATS.

Large language models (LLMs) transfer well to new tasks out-of-the-box simply given a natural language prompt that demonstrates how to perform the task and no additional training. Prompting is a brittle process wherein small modifications to the prompt can cause large variations in the model predictions, and therefore significant effort is dedicated towards designing a painstakingly crafted "perfect prompt" for a task. To mitigate the high degree of effort, we instead ask whether collecting multiple decent, yet imperfect, prompts and aggregating them can lead to a high quality prompting strategy. Our observations motivate our proposed method, Ask Me Anything (AMA). We first develop an understanding of the effective prompt formats, finding question-answering (QA) prompts, which encourage open-ended generation ("Who went to the park?") tend to outperform those that restrict the model outputs ("John went to the park. True or False?"). AMA recursively uses the LLM to transform task inputs to the effective QA format. AM generates multiple questions per input and applies these prompts to collect several noisy "votes" for the input's true label. We find the prompts have varying accuracies and dependencies and thus propose to use weak supervision, a procedure for combining the noisy predictions, to produce the final predictions. We evaluate AMA across open-source model families (EleutherAI, BLOOM, OPT, and T0) and sizes (125M-175B parameters), demonstrating an average performance lift of 10.2\% over the few-shot baseline. This simple strategy enables the open-source GPT-J-6B model to match and exceed the performance of few-shot GPT3-175B on 15 of 20 popular benchmarks. Averaged across these tasks, the GPT-J-6B model outperforms few-shot GPT3-175B. We release our code here: https://github.com/HazyResearch/ama_prompting.

Decoupling representation learning and classifier learning has been shown to be effective in classification with long-tailed data. There are two main ingredients in constructing a decoupled learning scheme; 1) how to train the feature extractor for representation learning so that it provides generalizable representations and 2) how to re-train the classifier that constructs proper decision boundaries by handling class imbalances in long-tailed data. In this work, we first apply Stochastic Weight Averaging (SWA), an optimization technique for improving the generalization of deep neural networks, to obtain better generalizing feature extractors for long-tailed classification. We then propose a novel classifier re-training algorithm based on stochastic representation obtained from the SWA-Gaussian, a Gaussian perturbed SWA, and a self-distillation strategy that can harness the diverse stochastic representations based on uncertainty estimates to build more robust classifiers. Extensive experiments on CIFAR10/100-LT, ImageNet-LT, and iNaturalist-2018 benchmarks show that our proposed method improves upon previous methods both in terms of prediction accuracy and uncertainty estimation.

The information-theoretic framework promises to explain the predictive power of neural networks. In particular, the information plane analysis, which measures mutual information (MI) between input and representation as well as representation and output, should give rich insights into the training process. This approach, however, was shown to strongly depend on the choice of estimator of the MI. The problem is amplified for deterministic networks if the MI between input and representation is infinite. Thus, the estimated values are defined by the different approaches for estimation, but do not adequately represent the training process from an information-theoretic perspective. In this work, we show that dropout with continuously distributed noise ensures that MI is finite. We demonstrate in a range of experiments that this enables a meaningful information plane analysis for a class of dropout neural networks that is widely used in practice.

Recurrent neural network (RNN) and self-attention mechanism (SAM) are the de facto methods to extract spatial-temporal information for temporal graph learning. Interestingly, we found that although both RNN and SAM could lead to a good performance, in practice neither of them is always necessary. In this paper, we propose GraphMixer, a conceptually and technically simple architecture that consists of three components: (1) a link-encoder that is only based on multi-layer perceptrons (MLP) to summarize the information from temporal links, (2) a node-encoder that is only based on neighbor mean-pooling to summarize node information, and (3) an MLP-based link classifier that performs link prediction based on the outputs of the encoders. Despite its simplicity, GraphMixer attains an outstanding performance on temporal link prediction benchmarks with faster convergence and better generalization performance. These results motivate us to rethink the importance of simpler model architecture.

We consider the problem of computing the $1$-Wasserstein distance $\mathcal{W}(\mu,\nu)$ between two $d$-dimensional discrete distributions $\mu$ and $\nu$ whose support lie within the unit hypercube. There are several algorithms that estimate $\mathcal{W}(\mu,\nu)$ within an additive error of $\varepsilon$. However, when $\mathcal{W}(\mu,\nu)$ is small, the additive error $\varepsilon$ dominates, leading to noisy results. Consider any additive approximation algorithm with execution time $T(n,\varepsilon)$. We propose an algorithm that runs in $O(T(n,\varepsilon/d) \log n)$ time and boosts the accuracy of estimating $\mathcal{W}(\mu,\nu)$ from $\varepsilon$ to an expected additive error of $\min\{\varepsilon, (d\log_{\sqrt{d}/\varepsilon} n)\mathcal{W}(\mu,\nu)\}$. For the special case where every point in the support of $\mu$ and $\nu$ has a mass of $1/n$ (also called the Euclidean Bipartite Matching problem), we describe an algorithm to boost the accuracy of any additive approximation algorithm from $\varepsilon$ to an expected additive error of $\min\{\varepsilon, (d\log\log n)\mathcal{W}(\mu,\nu)\}$ in $O(T(n, \varepsilon/d)\log\log n)$ time.

Out-of-distribution (OOD) detection is a critical task for reliable machine learning. Recent advances in representation learning give rise to distance-based OOD detection, where testing samples are detected as OOD if they are relatively far away from the centroids or prototypes of in-distribution (ID) classes. However, prior methods directly take off-the-shelf contrastive losses that suffice for classifying ID samples, but are not optimally designed when test inputs contain OOD samples. In this work, we propose CIDER, a novel representation learning framework that exploits hyperspherical embeddings for OOD detection. CIDER jointly optimizes two losses to promote strong ID-OOD separability: a dispersion loss that promotes large angular distances among different class prototypes, and a compactness loss that encourages samples to be close to their class prototypes. We analyze and establish the unexplored relationship between OOD detection performance and the embedding properties in the hyperspherical space, and demonstrate the importance of dispersion and compactness. CIDER establishes superior performance, outperforming the latest rival by 13.33% in FPR95. Code is available at https://github.com/deeplearning-wisc/cider.