ICLR 2023 Papers

Video generation, with the purpose of producing a sequence of frames, requires synthesizing consistent and persistent dynamic contents over time. This work investigates how to model the temporal relations for composing a video with arbitrary number of frames, from a few to even infinite, using generative adversarial networks (GANs). First, towards composing adjacent frames, we show that the alias-free operation for single image generation, together with adequately pre-learned knowledge, bring a smooth frame transition without harming the per-frame quality. Second, through incorporating a temporal shift module (TSM), which is originally designed for video understanding, into the discriminator, we manage to advance the generator in synthesizing more reasonable dynamics. Third, we develop a novel B-Spline based motion representation to ensure the temporal smoothness, and hence achieve infinite-length video generation, going beyond the frame number used in training. We evaluate our approach on a range of datasets and show substantial improvements over baselines on video generation. Code and models are publicly available at \url{https://genforce.github.io/StyleSV}.

We formally study how \emph{ensemble} of deep learning models can improve test accuracy, and how the superior performance of ensemble can be distilled into a single model using \emph{knowledge distillation}. We consider the challenging case where the ensemble is simply an average of the outputs of a few independently trained neural networks with the \emph{same} architecture, trained using the \emph{same} algorithm on the \emph{same} data set, and they only differ by the random seeds used in the initialization. We show that ensemble/knowledge distillation in \emph{deep learning} works very differently from traditional learning theory (such as boosting or NTKs). We develop a theory showing that when data has a structure we refer to as ``multi-view'', then ensemble of independently trained neural networks can provably improve test accuracy, and such superior test accuracy can also be provably distilled into a single model. Our result sheds light on how ensemble works in deep learning in a way that is completely different from traditional theorems, and how the ``dark knowledge'' is hidden in the outputs of the ensemble and can be used in distillation.

The transferability of adversarial perturbations between image models has been extensively studied. In this case, an attack is generated from a known surrogate \eg, the ImageNet trained model, and transferred to change the decision of an unknown (black-box) model trained on an image dataset. However, attacks generated from image models do not capture the dynamic nature of a moving object or a changing scene due to a lack of temporal cues within image models. This leads to reduced transferability of adversarial attacks from representation-enriched \emph{image} models such as Supervised Vision Transformers (ViTs), Self-supervised ViTs (\eg, DINO), and Vision-language models (\eg, CLIP) to black-box \emph{video} models. In this work, we induce dynamic cues within the image models without sacrificing their original performance on images. To this end, we optimize \emph{temporal prompts} through frozen image models to capture motion dynamics. Our temporal prompts are the result of a learnable transformation that allows optimizing for temporal gradients during an adversarial attack to fool the motion dynamics. Specifically, we introduce spatial (image) and temporal (video) cues within the same source model through task-specific prompts. Attacking such prompts maximizes the adversarial transferability from image-to-video and image-to-image models using the attacks designed for image models. As an example, an iterative attack launched from image model Deit-B with temporal prompts reduces generalization (top1 \% accuracy) of a video model by 35\% on Kinetics-400. Our approach also improves adversarial transferability to image models by 9\% on ImageNet w.r.t the current state-of-the-art approach. Our attack results indicate that the attacker does not need specialized architectures, \eg, divided space-time attention, 3D convolutions, or multi-view convolution networks for different data modalities. Image models are effective surrogates to optimize an adversarial attack to fool black-box models in a changing environment over time. Code is available at \url{https://bit.ly/3Xd9gRQ}

Despite the great success achieved in instructional action understanding by deep learning and mountainous data, deploying trained models to the unseen environment still remains a great challenge, since it requires strong generalizability of models from in-distribution training data to out-of-distribution (OOD) data. In this paper, we introduce a benchmark, named GAIN, to analyze the GeneralizAbility of INstructional action understanding models. In GAIN, we reassemble steps of existing instructional video training datasets to construct the OOD tasks and then collect the corresponding videos. We evaluate the generalizability of models trained on in-distribution datasets with the performance on OOD videos and observe a significant performance drop. We further propose a simple yet effective approach, which cuts off the excessive contextual dependency of action steps by performing causal inference, to provide a potential direction for enhancing the OOD generalizability. In the experiments, we show that this simple approach can improve several baselines on both instructional action segmentation and detection tasks. We expect the introduction of the GAIN dataset will promote future in-depth research on the generalization of instructional video understanding.

Abstraction is a desirable capability for deep learning models, which means to induce abstract concepts from concrete instances and flexibly apply them beyond the learning context. At the same time, there is a lack of clear understanding about both the presence and further characteristics of this capability in deep learning models. In this paper, we introduce a systematic probing framework to explore the abstraction capability of deep learning models from a transferability perspective. A set of controlled experiments are conducted based on this framework, providing strong evidence that two probed pre-trained language models (PLMs), T5 and GPT2, have the abstraction capability. We also conduct in-depth analysis, thus shedding further light: (1) the whole training phase exhibits a "memorize-then-abstract" two-stage process; (2) the learned abstract concepts are gathered in a few middle-layer attention heads, rather than being evenly distributed throughout the model; (3) the probed abstraction capabilities exhibit robustness against concept mutations, and are more robust to low-level/source-side mutations than high-level/target-side ones; (4) generic pre-training is critical to the emergence of abstraction capability, and PLMs exhibit better abstraction with larger model sizes and data scales.

The ability to predict future visual observations conditioned on past observations and motor commands can enable embodied agents to plan solutions to a variety of tasks in complex environments. This work shows that we can create good video prediction models by pre-training transformers via masked visual modeling. Our approach, named MaskViT, is based on two simple design decisions. First, for memory and training efficiency, we use two types of window attention: spatial and spatiotemporal. Second, during training, we mask a variable percentage of tokens instead of a fixed mask ratio. For inference, MaskViT generates all tokens via iterative refinement where we incrementally decrease the masking ratio following a mask scheduling function. On several datasets we demonstrate that MaskViT outperforms prior works in video prediction, is parameter efficient, and can generate high resolution videos ($256 \times $256). Further, we demonstrate the benefits of inference speedup (up to $512 \times$) due to iterative decoding by using MaskViT for planning on a real robot. Our work suggests that we can endow embodied agents with powerful predictive models by leveraging the general framework of masked visual modeling with minimal domain knowledge.

For unsupervised pretraining, mask-reconstruction pretraining (MRP) approaches, e.g. MAE and data2vec, randomly mask input patches and then reconstruct the pixels or semantic features of these masked patches via an auto-encoder. Then for a downstream task, supervised fine-tuning the pretrained encoder remarkably surpasses the conventional "supervised learning" (SL) trained from scratch. However, it is still unclear 1) how MRP performs semantic (feature) learning in the pretraining phase and 2) why it helps in downstream tasks. To solve these problems, we first theoretically show that on an auto-encoder of a two/one-layered convolution encoder/decoder, MRP can capture all discriminative semantics of each potential semantic class in the pretraining dataset. Then considering the fact that the pretraining dataset is of huge size and high diversity and thus covers most semantics in downstream dataset, in fine-tuning phase, the pretrained encoder can capture as much semantics as it can in downstream datasets, and would not lost these semantics with theoretical guarantees. In contrast, SL only randomly captures some semantics due to lottery ticket hypothesis. So MRP provably achieves better performance than SL on the classification tasks. Experimental results testify to our data assumptions and also our theoretical implications.

The backdoor attack, where the adversary uses inputs stamped with triggers (e.g., a patch) to activate pre-planted malicious behaviors, is a severe threat to Deep Neural Network (DNN) models. Trigger inversion is an effective way of identifying backdoor models and understanding embedded adversarial behaviors. A challenge of trigger inversion is that there are many ways of constructing the trigger. Existing methods cannot generalize to various types of triggers by making certain assumptions or attack-specific constraints. The fundamental reason is that existing work does not formally define the trigger and the inversion problem. This work formally defines and analyzes the trigger and the inversion problem. Then, it proposes a unified framework to invert backdoor triggers based on the formalization of triggers and the identified inner behaviors of backdoor models from our analysis. Our prototype UNICORN is general and effective in inverting backdoor triggers in DNNs. The code can be found at https://github.com/RU-System-Software-and-Security/UNICORN.

The goal of meta-learning is to learn a good prior model from a collection of tasks such that the learned prior is able to adapt quickly to new tasks without accessing many data from the new tasks. A common practice in meta-learning is to perform a train-validation split on each task, where the training set is used for adapting the model parameter to that specific task and the validation set is used for learning a prior model that is shared across all tasks. Despite its success and popularity in multitask learning and few-shot learning, the understanding of the train-validation split is still limited, especially when the neural network models are used. In this paper, we study the benefit of train-validation split for classification problems with neural network models trained by gradient descent. We prove that the train-validation split is necessary to learn a good prior model when the noise in the training sample is large, while the train-train method fails. We validate our theory by conducting experiment on both synthetic and real datasets. To the best of our knowledge, this is the first work towards the theoretical understanding of train-validation split in meta-learning with neural networks.

Multi-view projection methods have demonstrated promising performance on 3D understanding tasks like 3D classification and segmentation. However, it remains unclear how to combine such multi-view methods with the widely available 3D point clouds. Previous methods use unlearned heuristics to combine features at the point level. To this end, we introduce the concept of the multi-view point cloud (Voint cloud), representing each 3D point as a set of features extracted from several view-points. This novel 3D Voint cloud representation combines the compactness of 3D point cloud representation with the natural view-awareness of multi-view representation. Naturally, we can equip this new representation with convolutional and pooling operations. We deploy a Voint neural network (VointNet) to learn representations in the Voint space. Our novel representation achieves state-of-the-art performance on 3D classification, shape retrieval, and robust 3D part segmentation on standard benchmarks ( ScanObjectNN, ShapeNet Core55, and ShapeNet Parts). Further analysis shows that VointNet improves the robustness to occlusion compared to other methods.

We propose a new task to benchmark scene understanding of embodied agents: Situated Question Answering in 3D Scenes (SQA3D). Given a scene context (e.g., 3D scan), SQA3D requires the tested agent to first understand its situation (position, orientation, etc.) in the 3D scene as described by text, then reason about its surrounding environment and answer a question under that situation. Based upon 650 scenes from ScanNet, we provide a dataset centered around 6.8k unique situations, along with 20.4k descriptions and 33.4k diverse reasoning questions for these situations. These questions examine a wide spectrum of reasoning capabilities for an intelligent agent, ranging from spatial relation comprehension to commonsense understanding, navigation, and multi-hop reasoning. SQA3D imposes a significant challenge to current multi-modal especially 3D reasoning models. We evaluate various state-of-the-art approaches and find that the best one only achieves an overall score of 47.20%, while amateur human participants can reach 90.06%. We believe SQA3D could facilitate future embodied AI research with stronger situation understanding and reasoning capability.

Deep learning vision systems are widely deployed across applications where reliability is critical. However, even today's best models can fail to recognize an object when its pose, lighting, or background varies. While existing benchmarks surface examples challenging for models, they do not explain why such mistakes arise. To address this need, we introduce ImageNet-X—a set of sixteen human annotations of factors such as pose, background, or lighting the entire ImageNet-1k validation set as well as a random subset of 12k training images. Equipped with ImageNet-X, we investigate 2,200 current recognition models and study the types of mistakes as a function of model’s (1) architecture, e.g. transformer vs. convolutional, (2) learning paradigm, e.g. supervised vs. self-supervised, and (3) training procedures, e.g., data augmentation. Regardless of these choices, we find models have consistent failure modes across ImageNet-X categories. We also find that while data augmentation can improve robustness to certain factors, they induce spill-over effects to other factors. For example, color-jitter augmentation improves robustness to color and brightness, but surprisingly hurts robustness to pose. Together, these insights suggest to advance the robustness of modern vision models, future research should focus on collecting additional data and understanding data augmentation schemes. Along with these insights, we release a toolkit based on ImageNet-X to spur further study into the mistakes image recognition systems make.

We consider a setting that a model needs to adapt to a new domain under distribution shifts, given that only unlabeled test samples from the new domain are accessible at test time. A common idea in most of the related works is constructing pseudo-labels for the unlabeled test samples and applying gradient descent (GD) to a loss function with the pseudo-labels. Recently, Goyal et al. (2022) propose conjugate labels, which is a new kind of pseudo-labels for self-training at test time. They empirically show that the conjugate label outperforms other ways of pseudo-labeling on many domain adaptation benchmarks. However, provably showing that GD with conjugate labels learns a good classifier for test-time adaptation remains open. In this work, we aim at theoretically understanding GD with hard and conjugate labels for a binary classification problem. We show that for square loss, GD with conjugate labels converges to an $\epsilon$-optimal predictor under a Gaussian model for any arbitrarily small $\epsilon$, while GD with hard pseudo-labels fails in this task. We also analyze them under different loss functions for the update. Our results shed lights on understanding when and why GD with hard labels or conjugate labels works in test-time adaptation.

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.

Deploying machine learning models on new tasks is a major challenge due to differences in distributions of the train (source) data and the new (target) data. However, the training data likely captures some of the properties of the new task. We consider the problem of reweighing the training samples to gain insights into the distribution of the target task. Specifically, we formulate a distribution shift model based on the exponential tilt assumption and learn train data importance weights minimizing the KL divergence between labeled train and unlabeled target datasets. The learned train data weights can then be used for downstream tasks such as target performance evaluation, fine-tuning, and model selection. We demonstrate the efficacy of our method on Waterbirds and Breeds benchmarks.

Recent vision transformer based video models mostly follow the ``image pre-training then finetuning" paradigm and have achieved great success on multiple video benchmarks. However, fully finetuning such a video model could be computationally expensive and unnecessary, given the pre-trained image transformer models have demonstrated exceptional transferability. In this work, we propose a novel method to Adapt pre-trained Image Models (AIM) for efficient video understanding. By freezing the pre-trained image model and adding a few lightweight Adapters, we introduce spatial adaptation, temporal adaptation and joint adaptation to gradually equip an image model with spatiotemporal reasoning capability. We show that our proposed AIM can achieve competitive or even better performance than prior arts with substantially fewer tunable parameters on four video action recognition benchmarks. Thanks to its simplicity, our method is also generally applicable to different image pre-trained models, which has the potential to leverage more powerful image foundation models in the future. The project webpage is https://adapt-image-models.github.io/.

In machine learning, we traditionally evaluate the performance of a single model, averaged over a collection of test inputs. In this work, we propose a new approach: we measure the performance of a collection of models when evaluated at *single input point*. Specifically, we study a point's *profile*: the relationship between models' average performance on the test distribution and their pointwise performance on this individual point. We find that profiles can yield new insights into the structure of both models and data---in and out-of-distribution. For example, we empirically show that real data distributions consist of points with qualitatively different profiles. On one hand, there are ``compatible'' points with strong correlation between the pointwise and average performance. On the other hand, there are points with weak and even *negative* correlation: cases where improving overall model accuracy actually *hurts* performance on these inputs. As an application, we use profiles to construct a dataset we call CIFAR-10-NEG: a subset of CINIC-10 such that for standard models, accuracy on CIFAR-10-NEG is *negatively correlated* with CIFAR-10 accuracy. Illustrating for the first time an OOD dataset that completely inverts ``accuracy-on-the-line'' (Miller et al., 2021).

Real-world machine learning applications often involve deploying neural networks to domains that are not seen in the training time. Hence, we need to understand the extrapolation of \textit{nonlinear} models---under what conditions on the distributions and function class, models can be guaranteed to extrapolate to new test distributions. The question is very challenging because even two-layer neural networks cannot be guaranteed to extrapolate outside the support of the training distribution without further assumptions on the domain shift. This paper makes some initial steps towards analyzing the extrapolation of nonlinear models for structured domain shift. We primarily consider settings where the \textit{marginal} distribution of each coordinate of the data (or subset of coordinates) do not shift significantly across the training and test distributions, but the joint distribution may have a much bigger shift. We prove that the family of nonlinear models of the form $f(x)=\sum f_i(x_i)$, where $f_i$ is an \emph{arbitrary} function on the subset of features $x_i$, can extrapolate to unseen distributions, if the covariance of the features is well-conditioned. To the best of our knowledge, this is the first result that goes beyond linear models and the bounded density ratio assumption, even though the assumptions on the distribution shift and function class are stylized.

Self-supervised learning has significantly improved the performance of many NLP tasks. However, how can self-supervised learning discover useful features, and why is it better than traditional approaches such as probabilistic models are still largely unknown. In this paper, we focus on the context of topic modeling and highlight a key advantage of self-supervised learning - when applied to data generated by topic models, self-supervised learning can be oblivious to the specific model, and hence is less susceptible to model misspecification. In particular, we prove that commonly used self-supervised objectives based on reconstruction or contrastive samples can both recover useful posterior information for general topic models. Empirically, we show that the same objectives can perform on par with posterior inference using the correct model, while outperforming posterior inference using misspecified models.