ICLR 2022 Papers

When transferring a pretrained model to a downstream task, two popular methods are full fine-tuning (updating all the model parameters) and linear probing (updating only the last linear layer---the "head"). It is well known that fine-tuning leads to better accuracy in-distribution (ID). However, in this paper, we find that fine-tuning can achieve worse accuracy than linear probing out-of-distribution (OOD) when the pretrained features are good and the distribution shift is large. On 10 distribution shift datasets (BREEDS-Living17, BREEDS-Entity30, DomainNet, CIFAR $\to$ STL, CIFAR-10.1, FMoW, ImageNetV2, ImageNet-R, ImageNet-A, ImageNet-Sketch), fine-tuning obtains on average 2% higher accuracy ID but 7% lower accuracy OOD than linear probing. We show theoretically that this tradeoff between ID and OOD accuracy arises even in a simple setting: fine-tuning overparameterized two-layer linear networks. We prove that the OOD error of fine-tuning is high when we initialize with a fixed or random head---this is because while fine-tuning learns the head, the lower layers of the neural network change simultaneously and distort the pretrained features. Our analysis suggests that the easy two-step strategy of linear probing then full fine-tuning (LP-FT), sometimes used as a fine-tuning heuristic, combines the benefits of both fine-tuning and linear probing. Empirically, LP-FT outperforms both fine-tuning and linear probing on the above datasets (1% better ID, 10% better OOD than full fine-tuning).

The nearest neighbor search (NNS) problem is widely studied in Euclidean space, and graph-based algorithms are known to outperform other approaches for this task. However, hyperbolic geometry often allows for better data representation in various domains, including graphs, words, and images. In this paper, we show that graph-based approaches are also well suited for hyperbolic geometry. From a theoretical perspective, we rigorously analyze the time and space complexity of graph-based NNS, assuming that an $n$-element dataset is uniformly distributed within a $d$-dimensional ball of radius $R$ in the hyperbolic space of curvature $-1$. Under some conditions on $R$ and $d$, we derive the time and space complexity of graph-based NNS and compare the obtained results with known guarantees for the Euclidean case. Interestingly, in the dense setting ($d \ll \log n$) and under some assumptions on the radius $R$, graph-based NNS has lower time complexity in the hyperbolic space. This agrees with our experiments: we consider datasets embedded in hyperbolic and Euclidean spaces and show that graph-based NNS can be more efficient in the hyperbolic space. We also demonstrate that graph-based methods outperform other existing baselines for hyperbolic NNS. Overall, our theoretical and empirical analysis suggests that graph-based NNS can be considered a default approach for similarity search in hyperbolic spaces.

Graph Neural Networks (GNNs) have achieved great success in various tasks, but their performance highly relies on a large number of labeled nodes, which typically requires considerable human effort. GNN-based Active Learning (AL) methods are proposed to improve the labeling efficiency by selecting the most valuable nodes to label. Existing methods assume an oracle can correctly categorize all the selected nodes and thus just focus on the node selection. However, such an exact labeling task is costly, especially when the categorization is out of the domain of individual expert (oracle). The paper goes further, presenting a soft-label approach to AL on GNNs. Our key innovations are: i) relaxed queries where a domain expert (oracle) only judges the correctness of the predicted labels (a binary question) rather than identifying the exact class (a multi-class question), and ii) new criteria of maximizing information gain propagation for active learner with relaxed queries and soft labels. Empirical studies on public datasets demonstrate that our method significantly outperforms the state-of-the-art GNN-based AL methods in terms of both accuracy and labeling cost.

A conventional approach to entity linking is to first find mentions in a given document and then infer their underlying entities in the knowledge base. A well-known limitation of this approach is that it requires finding mentions without knowing their entities, which is unnatural and difficult. We present a new model that does not suffer from this limitation called $\textbf{EntQA}$, which stands for $\mbox{\textbf{Ent}ity}$ linking as $\mbox{\textbf{Q}uestion}$ $\mbox{\textbf{A}nswering}$. EntQA first proposes candidate entities with a fast retrieval module, and then scrutinizes the document to find mentions of each candidate with a powerful reader module. Our approach combines progress in entity linking with that in open-domain question answering and capitalizes on pretrained models for dense entity retrieval and reading comprehension. Unlike in previous works, we do not rely on a mention-candidates dictionary or large-scale weak supervision. EntQA achieves strong results on the GERBIL benchmarking platform.

We introduce the concept of provably robust adversarial examples for deep neural networks – connected input regions constructed from standard adversarial examples which are guaranteed to be robust to a set of real-world perturbations (such as changes in pixel intensity and geometric transformations). We present a novel method called PARADE for generating these regions in a scalable manner which works by iteratively refining the region initially obtained via sampling until a refined region is certified to be adversarial with existing state-of-the-art verifiers. At each step, a novel optimization procedure is applied to maximize the region's volume under the constraint that the convex relaxation of the network behavior with respect to the region implies a chosen bound on the certification objective. Our experimental evaluation shows the effectiveness of PARADE: it successfully finds large provably robust regions including ones containing $\approx 10^{573}$ adversarial examples for pixel intensity and $\approx 10^{599}$ for geometric perturbations. The provability enables our robust examples to be significantly more effective against state-of-the-art defenses based on randomized smoothing than the individual attacks used to construct the regions.

We introduce the maximum $n$-times coverage problem that selects $k$ overlays to maximize the summed coverage of weighted elements, where each element must be covered at least $n$ times. We also define the min-cost $n$-times coverage problem where the objective is to select the minimum set of overlays such that the sum of the weights of elements that are covered at least $n$ times is at least $\tau$. Maximum $n$-times coverage is a generalization of the multi-set multi-cover problem, is NP-complete, and is not submodular. We introduce two new practical solutions for $n$-times coverage based on integer linear programming and sequential greedy optimization. We show that maximum $n$-times coverage is a natural way to frame peptide vaccine design, and find that it produces a pan-strain COVID-19 vaccine design that is superior to 29 other published designs in predicted population coverage and the expected number of peptides displayed by each individual's HLA molecules.

Implicit shape models are promising 3D representations for modeling arbitrary locations, with Signed Distance Functions (SDFs) particularly suitable for clear mesh surface reconstruction. Existing approaches for single object reconstruction impose supervision signals based on the loss of the signed distance value from all locations in a scene, posing difficulties when extending to real-world scenarios. The spatial gradient of the signed distance field, rather than the SDF value itself, has not been typically employed as a source of supervision for single-view reconstruction, in part due to the difficulties of differentiable sampling a spatial gradient from the feature map. In this study, we derive a novel closed-form gradient sampling solution for Differentialble Gradient Sampling (DGS) that enables backpropagation of the loss of the spatial gradient back to the feature map pixels, thus allowing the imposition of the loss efficiently on the spatial gradient. As a result, we achieve high-quality single view indoor scene reconstruction results learning directly from a real-world scanned dataset (e.g. ScannetV2). Our model also performs well when generalizing to unseen images downloaded directly from the internet (Fig. 1). We comfortably advanced the state-of-the-art results with several established datasets including ShapeNet and ScannetV2; extensive quantitative analysis confirmed that our proposed DGS module plays an essential role in achieving this performance improvement. Full codes are available in MaskedURL.

Time series signal analysis plays an essential role in many applications, e.g., activity recognition and healthcare monitoring. Recently, features extracted with deep neural networks (DNNs) have shown to be more effective than conventional hand-crafted ones. However, most existing solutions rely solely on the network to extract information carried in the raw signal, regardless of its inherent physical and statistical properties, leading to sub-optimal performance particularly under a limited amount of training data. In this work, we propose a novel tree-structured wavelet neural network for time series signal analysis, namely \emph{T-WaveNet}, taking advantage of an inherent property of various types of signals, known as the \emph{dominant frequency range}. Specifically, with \emph{T-WaveNet}, we first conduct frequency spectrum energy analysis of the signals to get a set of dominant frequency subbands. Then, we construct a tree-structured network that iteratively decomposes the input signal into various frequency subbands with similar energies. Each node on the tree is built with an invertible neural network (INN) based wavelet transform unit. Such a disentangled representation learning method facilitates a more effective extraction of the discriminative features, as demonstrated with the comprehensive experiments on various real-life time series classification datasets.

Training very deep neural networks is still an extremely challenging task. The common solution is to use shortcut connections and normalization layers, which are both crucial ingredients in the popular ResNet architecture. However, there is strong evidence to suggest that ResNets behave more like ensembles of shallower networks than truly deep ones. Recently, it was shown that deep vanilla networks (i.e.~networks without normalization layers or shortcut connections) can be trained as fast as ResNets by applying certain transformations to their activation functions. However, this method (called Deep Kernel Shaping) isn't fully compatible with ReLUs, and produces networks that overfit significantly more than ResNets on ImageNet. In this work, we rectify this situation by developing a new type of transformation that is fully compatible with a variant of ReLUs -- Leaky ReLUs. We show in experiments that our method, which introduces negligible extra computational cost, achieves validation accuracies with deep vanilla networks that are competitive with ResNets (of the same width/depth), and significantly higher than those obtained with the Edge of Chaos (EOC) method. And unlike with EOC, the validation accuracies we obtain do not get worse with depth.

Recent studies have revealed that deep neural networks (DNNs) are vulnerable to backdoor attacks, where attackers embed hidden backdoors in the DNN model by poisoning a few training samples. The attacked model behaves normally on benign samples, whereas its prediction will be maliciously changed when the backdoor is activated. We reveal that poisoned samples tend to cluster together in the feature space of the attacked DNN model, which is mostly due to the end-to-end supervised training paradigm. Inspired by this observation, we propose a novel backdoor defense via decoupling the original end-to-end training process into three stages. Specifically, we first learn the backbone of a DNN model via \emph{self-supervised learning} based on training samples without their labels. The learned backbone will map samples with the same ground-truth label to similar locations in the feature space. Then, we freeze the parameters of the learned backbone and train the remaining fully connected layers via standard training with all (labeled) training samples. Lastly, to further alleviate side-effects of poisoned samples in the second stage, we remove labels of some `low-credible' samples determined based on the learned model and conduct a \emph{semi-supervised fine-tuning} of the whole model. Extensive experiments on multiple benchmark datasets and DNN models verify that the proposed defense is effective in reducing backdoor threats while preserving high accuracy in predicting benign samples. Our code is available at \url{https://github.com/SCLBD/DBD}.

Language models typically need to be trained or finetuned in order to acquire new knowledge, which involves updating their weights. We instead envision language models that can simply read and memorize new data at inference time, thus acquiring new knowledge immediately. In this work, we extend language models with the ability to memorize the internal representations of past inputs. We demonstrate that an approximate $k$NN lookup into a non-differentiable memory of recent (key, value) pairs improves language modeling across various benchmarks and tasks, including generic webtext (C4), math papers (arXiv), books (PG-19), code (Github), as well as formal theorems (Isabelle). We show that the performance steadily improves when we increase the size of memory up to 262K tokens. On benchmarks including code and mathematics, we find that the model is capable of making use of newly defined functions and theorems during test time.

Visual search, recommendation, and contrastive similarity learning power technologies that impact billions of users worldwide. Modern model architectures can be complex and difficult to interpret, and there are several competing techniques one can use to explain a search engine's behavior. We show that the theory of fair credit assignment provides a unique axiomatic solution that generalizes several existing recommendation- and metric-explainability techniques in the literature. Using this formalism, we show when existing approaches violate "fairness" and derive methods that sidestep these shortcomings and naturally handle counterfactual information. More specifically, we show existing approaches implicitly approximate second-order Shapley-Taylor indices and extend CAM, GradCAM, LIME, SHAP, SBSM, and other methods to search engines. These extensions can extract pairwise correspondences between images from trained opaque-box models. We also introduce a fast kernel-based method for estimating Shapley-Taylor indices that require orders of magnitude fewer function evaluations to converge. Finally, we show that these game-theoretic measures yield more consistent explanations for image similarity architectures.

We present a novel approach to use curricula to identify principles by which a system learns. Previous work in curriculum learning has focused on how curricula can be designed to improve learning of a model on particular tasks. We consider the inverse problem: what can a curriculum tell us about how a learning system acquired a task? Using recurrent neural networks (RNNs) and models of common experimental neuroscience tasks, we demonstrate that curricula can be used to differentiate learning principles using target-based and a representation-based loss functions as use cases. In particular, we compare the performance of RNNs using target-based learning rules versus those using representational learning rules on three different curricula in the context of two tasks. We show that the learned state-space trajectories of RNNs trained by these two learning rules under all curricula tested are indistinguishable. However, by comparing learning times during different curricula, we can disambiguate the learning rules and challenge traditional approaches of interrogating learning systems. Although all animals in neuroscience lab settings are trained by curriculum-based procedures called shaping, almost no behavioral or neural data are collected or published on the relative successes or training times under different curricula. Our results motivate the systematic collection and curation of data during shaping by demonstrating curriculum learning in RNNs as a tool to probe and differentiate learning principles used by biological systems, over conventional statistical analyses of learned state spaces.

To train machine learning models that are robust to distribution shifts in the data, distributionally robust optimization (DRO) has been proven very effective. However, the existing approaches to learning a distributionally robust model either require solving complex optimization problems such as semidefinite programming or a first-order method whose convergence scales linearly with the number of data samples-- which hinders their scalability to large datasets. In this paper, we show how different variants of DRO are simply instances of a finite-sum composite optimization for which we provide scalable methods. We also provide empirical results that demonstrate the effectiveness of our proposed algorithm with respect to the prior art in order to learn robust models from very large datasets.

Preference-based reinforcement learning (RL) has shown potential for teaching agents to perform the target tasks without a costly, pre-defined reward function by learning the reward with a supervisor’s preference between the two agent behaviors. However, preference-based learning often requires a large amount of human feedback, making it difficult to apply this approach to various applications. This data-efficiency problem, on the other hand, has been typically addressed by using unlabeled samples or data augmentation techniques in the context of supervised learning. Motivated by the recent success of these approaches, we present SURF, a semi-supervised reward learning framework that utilizes a large amount of unlabeled samples with data augmentation. In order to leverage unlabeled samples for reward learning, we infer pseudo-labels of the unlabeled samples based on the confidence of the preference predictor. To further improve the label-efficiency of reward learning, we introduce a new data augmentation that temporally crops consecutive subsequences from the original behaviors. Our experiments demonstrate that our approach significantly improves the feedback-efficiency of the state-of-the-art preference-based method on a variety of locomotion and robotic manipulation tasks.

Missingness, or the absence of features from an input, is a concept fundamental to many model debugging tools. However, in computer vision, pixels cannot simply be removed from an image. One thus tends to resort to heuristics such as blacking out pixels, which may in turn introduce bias into the debugging process. We study such biases and, in particular, show how transformer-based architectures can enable a more natural implementation of missingness, which side-steps these issues and improves the reliability of model debugging in practice.

Vision transformer (ViT) has recently shown its strong capability in achieving comparable results to convolutional neural networks (CNNs) on image classification. However, vanilla ViT simply inherits the same architecture from the natural language processing directly, which is often not optimized for vision applications. Motivated by this, in this paper, we propose a new architecture that adopts the pyramid structure and employ novel regional-to-local attention rather than global self-attention in vision transformers. More specifically, our model first generates regional tokens and local tokens from an image with different patch sizes, where each regional token is associated with a set of local tokens based on the spatial location. The regional-to-local attention includes two steps: first, the regional self-attention extracts global information among all regional tokens and then the local self-attention exchanges the information among one regional token and the associated local tokens via self-attention. Therefore, even though local self-attention confines the scope in a local region but it can still receive global information. Extensive experiments on four vision tasks, including image classification, object and keypoint detection, semantics segmentation and action recognition, show that our approach outperforms or is on par with state-of-the-art ViT variants including many concurrent works. Our source codes and models are available at \url{https://github.com/IBM/RegionViT}.

This paper studies some under-explored connections between personalized recommendation and marketing systems. Obviously, these two systems are different, in two main ways. Firstly, personalized item-recommendation (ItemRec) is user-centric, whereas marketing recommends the best user-state segments (UserRec) on behalf of its item providers. (We treat different temporal states of the same user as separate marketing opportunities.) To overcome this difference, we realize a novel connection to Marked-Temporal Point Processes (MTPPs), where we view both problems as different projections from a unified temporal intensity model for all user-item pairs. Correspondingly, we derive Recurrent Intensity Models (RIMs) to extend from recurrent ItemRec models with minimal changes. The second difference between recommendation and marketing is in the temporal domains where they operate. While recommendation demands immediate responses in real-time, marketing campaigns are often long-term, setting goals to cover a given percentage of all opportunities for a given item in a given period of time. We formulate both considerations into a constrained optimization problem we call online match (OnlnMtch) and derive a solution we call Dual algorithm. Simply put, Dual modifies the real-time ItemRec scores such that the marketing constraints can be met with least compromises in user-centric utilities. Finally, our connections between recommendation and marketing may lead to novel applications. We run experiments where we use marketing as an alternative to cold-start item exploration, by setting a minimal-exposure constraint for every item in the audience base. Our experiments are available at \url{https://github.com/awslabs/recurrent-intensity-model-experiments}

We introduce Network Augmentation (NetAug), a new training method for improving the performance of tiny neural networks. Existing regularization techniques (e.g., data augmentation, dropout) have shown much success on large neural networks by adding noise to overcome over-fitting. However, we found these techniques hurt the performance of tiny neural networks. We argue that training tiny models are different from large models: rather than augmenting the data, we should augment the model, since tiny models tend to suffer from under-fitting rather than over-fitting due to limited capacity. To alleviate this issue, NetAug augments the network (reverse dropout) instead of inserting noise into the dataset or the network. It puts the tiny model into larger models and encourages it to work as a sub-model of larger models to get extra supervision, in addition to functioning as an independent model. At test time, only the tiny model is used for inference, incurring zero inference overhead. We demonstrate the effectiveness of NetAug on image classification and object detection. NetAug consistently improves the performance of tiny models, achieving up to 2.2% accuracy improvement on ImageNet. On object detection, achieving the same level of performance, NetAug requires 41% fewer MACs on Pascal VOC and 38% fewer MACs on COCO than the baseline.