ICLR 2023 Papers

Over the past few years afterward the birth of ResNet, skip connection has become the defacto standard for the design of modern architectures due to its widespread adoption, easy optimization, and proven performance. Prior work has explained the effectiveness of the skip connection mechanism from different perspectives. In this work, we deep dive into the model's behaviors with skip connections which can be formulated as a learnable Markov chain. An efficient Markov chain is preferred as it always maps the input data to the target domain in a better way. However, while a model is explained as a Markov chain, it is not guaranteed to be optimized following an efficient Markov chain by existing SGD-based optimizers prone to getting trapped in local optimal points. In order to move towards a more efficient Markov chain, we propose a simple routine of penal connection to make any residual-like model become a learnable Markov chain. Aside from that, the penal connection can also be viewed as a particular model regularization and can be easily implemented with one line of code in the most popular deep learning frameworks. The encouraging experimental results in multi-modal translation and image recognition empirically confirm our conjecture of the learnable Markov chain view and demonstrate the superiority of the proposed penal connection.

We introduce VA-DepthNet, a simple, effective, and accurate deep neural network approach for the single-image depth prediction (SIDP) problem. The proposed approach advocates using classical first-order variational constraints for this problem. While state-of-the-art deep neural network methods for SIDP learn the scene depth from images in a supervised setting, they often overlook the invaluable invariances and priors in the rigid scene space, such as the regularity of the scene. The paper's main contribution is to reveal the benefit of classical and well-founded variational constraints in the neural network design for the SIDP task. It is shown that imposing first-order variational constraints in the scene space together with popular encoder-decoder-based network architecture design provides excellent results for the supervised SIDP task. The imposed first-order variational constraint makes the network aware of the depth gradient in the scene space, i.e., regularity. The paper demonstrates the usefulness of the proposed approach via extensive evaluation and ablation analysis over several benchmark datasets, such as KITTI, NYU Depth V2, and SUN RGB-D. The VA-DepthNet at test time shows considerable improvements in depth prediction accuracy compared to the prior art and is accurate also at high-frequency regions in the scene space. At the time of writing this paper, our method---labeled as VA-DepthNet, when tested on the KITTI depth-prediction evaluation set benchmarks, shows state-of-the-art results, and is the top-performing published approach.

Scene Graph Generation (SGG) has already shown its great potential in various downstream tasks, but it comes at the price of a prohibitively expensive annotation process. To reduce the annotation cost, we propose using Active Learning (AL) for sampling the most informative data. However, directly porting current AL methods to the SGG task poses the following challenges: 1) unreliable uncertainty estimates, and 2) data bias problems. To deal with these challenges, we propose EDAL (\textbf{E}vidential Uncertainty and \textbf{D}iversity Guided Deep \textbf{A}ctive \textbf{L}earning), a novel AL framework tailored for the SGG task. For challenge 1), we start with Evidential Deep Learning (EDL) coupled with a global relationship mining approach to estimate uncertainty, which can effectively overcome the perturbations of open-set relationships and background-relationships to obtain reliable uncertainty estimates. To address challenge 2), we seek the diversity-based method and design the Context Blocking Module (CBM) and Image Blocking Module (IBM) to alleviate context-level bias and image-level bias, respectively. Experiments show that our AL framework can approach the performance of a fully supervised SGG model with only about $10\%$ annotation cost. Furthermore, our ablation studies indicate that introducing AL into the SGG will face many challenges not observed in other vision tasks that are successfully overcome by our new modules.

Recent text-to-image models have achieved impressive results. However, since they require large-scale datasets of text-image pairs, it is impractical to train them on new domains where data is scarce or not labeled. In this work, we propose using large-scale retrieval methods, in particular, efficient k-Nearest-Neighbors (kNN), which offers novel capabilities: (1) training a substantially small and efficient text-to-image diffusion model using only pre-trained multi-modal embeddings, but without an explicit text-image dataset, (2) generating out-of-distribution images by simply swapping the retrieval database at inference time, and (3) performing text-driven local semantic manipulations while preserving object identity. To demonstrate the robustness of our method, we apply our kNN approach on two state-of-the-art diffusion backbones, and show results on several different datasets. As evaluated by human studies and automatic metrics, our method achieves state-of-the-art results compared to existing approaches that train text-to-image generation models using images-only dataset.

Segmentation uncertainty models predict a distribution over plausible segmentations for a given input, which they learn from the annotator variation in the training set. However, in practice these annotations can differ systematically in the way they are generated, for example through the use of different labeling tools. This results in datasets that contain both data variability and differing label styles. In this paper, we demonstrate that applying state-of-the-art segmentation uncertainty models on such datasets can lead to model bias caused by the different label styles. We present an updated modelling objective conditioning on labeling style for aleatoric uncertainty estimation, and modify two state-of-the-art-architectures for segmentation uncertainty accordingly. We show with extensive experiments that this method reduces label style bias, while improving segmentation performance, increasing the applicability of segmentation uncertainty models in the wild. We curate two datasets, with annotations in different label styles, which we will make publicly available along with our code upon publication.

We present Phenaki, a model capable of realistic video synthesis given a sequence of textual prompts. Generating videos from text is particularly challenging due to the computational cost, limited quantities of high quality text-video data and variable length of videos. To address these issues, we introduce a new causal model for learning video representation which compresses the video to a small discrete tokens representation. This tokenizer is auto-regressive in time, which allows it to work with video representations of different length. To generate video tokens from text we are using a bidirectional masked transformer conditioned on pre-computed text tokens. The generated video tokens are subsequently de-tokenized to create the actual video. To address data issues, we demonstrate how joint training on a large corpus of image-text pairs as well as a smaller number of video-text examples can result in generalization beyond what is available in the video datasets. Compared to the previous video generation methods, Phenaki can generate arbitrary long videos conditioned on a sequence of prompts (i.e. time variable text or story) in open domain. To the best of our knowledge, this is the first time a paper studies generating videos from time variable prompts.

Current reinforcement learning (RL) often suffers when solving a challenging exploration problem where the desired outcomes or high rewards are rarely observed. Even though curriculum RL, a framework that solves complex tasks by proposing a sequence of surrogate tasks, shows reasonable results, most of the previous works still have difficulty in proposing curriculum due to the absence of a mechanism for obtaining calibrated guidance to the desired outcome state without any prior domain knowledge. To alleviate it, we propose an uncertainty \& temporal distance-aware curriculum goal generation method for the outcome-directed RL via solving a bipartite matching problem. It could not only provide precisely calibrated guidance of the curriculum to the desired outcome states but also bring much better sample efficiency and geometry-agnostic curriculum goal proposal capability compared to previous curriculum RL methods. We demonstrate that our algorithm significantly outperforms these prior methods in a variety of challenging navigation tasks and robotic manipulation tasks in a quantitative and qualitative way.

Image restoration aims to reconstruct the latent sharp image from its corrupted counterpart. Besides dealing with this long-standing task in the spatial domain, a few approaches seek solutions in the frequency domain in consideration of the large discrepancy between spectra of sharp/degraded image pairs. However, these works commonly utilize transformation tools, e.g., wavelet transform, to split features into several frequency parts, which is not flexible enough to select the most informative frequency component to recover. In this paper, we exploit a multi-branch and content-aware module to decompose features into separate frequency subbands dynamically and locally, and then accentuate the useful ones via channel-wise attention weights. In addition, to handle large-scale degradation blurs, we propose an extremely simple decoupling and modulation module to enlarge the receptive field via global and window-based average pooling. Integrating two developed modules into a U-Net backbone, the proposed Selective Frequency Network (SFNet) performs favorably against state-of-the-art algorithms on five image restoration tasks, including single-image defocus deblurring, image dehazing, image motion deblurring, image desnowing, and image deraining.

In this work, our objective is to adapt a Deep generative model trained on a large-scale source dataset to multiple target domains with scarce data. Specifically, we focus on adapting a pre-trained Generative Adversarial Network (GAN) to a target domain without re-training the generator. Our method draws the motivation from the fact that out-of-distribution samples can be `embedded' onto the latent space of a pre-trained source-GAN. We propose to train a small latent-generation network during the inference stage, each time a batch of target samples is to be generated. These target latent codes are fed to the source-generator to obtain novel target samples. Despite using the same small set of target samples and the source generator, multiple independent training episodes of the latent-generation network results in the diversity of the generated target samples. Our method, albeit simple, can be used to generate data from multiple target distributions using a generator trained on a single source distribution. We demonstrate the efficacy of our surprisingly simple method in generating multiple target datasets with only a single source generator and a few target samples.

In many retrieval systems the original high dimensional data (e.g., images) is mapped to a lower dimensional feature through a learned embedding model. The task of retrieving the most similar data from a gallery set to a given query data is performed through similarity comparison on features. When the embedding model is updated, it might produce features that are not comparable/compatible with features already in the gallery computed with the old model. Subsequently, all features in the gallery need to be re-computed using the new embedding model -- a computationally expensive process called backfilling. Recently, compatible representation learning methods have been proposed to avoid back-filling. Despite their relative success, there is an inherent trade-off between new model performance and its compatibility with the old model. In this work, we introduce FastFill: a compatible model update process using feature alignment and policy based partial backfilling to promptly elevate retrieval performance. We show that previous backfilling strategies suffer from decreased performance and demonstrate the importance of both the training objective and the ordering in online partial backfilling. We propose a new training method for feature alignment between old and new embedding models using uncertainty estimation. Compared to previous works, we obtain significantly improved backfilling results on a variety of datasets: mAP on ImageNet (+4.4%), Places-365 (+2.7%), and VGG-Face2 (+1.3%). Further, we demonstrate that when updating a biased model with FastFill, the minority subgroup accuracy gap promptly vanishes with a small fraction of partial backfilling.

In this work, we present CogVideo, a 9B-parameter transformer for text-to-video generation. The CogVideo model has been trained by inheriting a pretrained text-to-image model, CogView2, which significantly reduces the training cost and alleviates the problem of scarcity and weak relevance. We also propose a multi-frame-rate training strategy for better aligning text and video clips. CogVideo achieves state-of-the-art performance in machine evaluation and outperforms publicly available models by a large margin in human evaluation. Its codes and model are also publicly available at https://github.com/THUDM/CogVideo.

We propose a novel \underline{e}dge guided \underline{g}enerative \underline{a}dversarial \underline{n}etwork with \underline{c}ontrastive learning (ECGAN) for the challenging semantic image synthesis task. Although considerable improvement has been achieved, the quality of synthesized images is far from satisfactory due to three largely unresolved challenges. 1) The semantic labels do not provide detailed structural information, making it difficult to synthesize local details and structures. 2) The widely adopted CNN operations such as convolution, down-sampling, and normalization usually cause spatial resolution loss and thus cannot fully preserve the original semantic information, leading to semantically inconsistent results (e.g., missing small objects). 3) Existing semantic image synthesis methods focus on modeling ``local'' semantic information from a single input semantic layout. However, they ignore ``global'' semantic information of multiple input semantic layouts, i.e., semantic cross-relations between pixels across different input layouts. To tackle 1), we propose to use edge as an intermediate representation which is further adopted to guide image generation via a proposed attention guided edge transfer module. Edge information is produced by a convolutional generator and introduces detailed structure information. To tackle 2), we design an effective module to selectively highlight class-dependent feature maps according to the original semantic layout to preserve the semantic information. To tackle 3), inspired by current methods in contrastive learning, we propose a novel contrastive learning method, which aims to enforce pixel embeddings belonging to the same semantic class to generate more similar image content than those from different classes. Doing so can capture more semantic relations by explicitly exploring the structures of labeled pixels from multiple input semantic layouts. Experiments on three challenging datasets show that our ECGAN achieves significantly better results than state-of-the-art methods.

A number of problems in learning can be formulated in terms of the basic primitive of sampling $k$ elements out of a universe of $n$ elements. This subset sampling operation cannot directly be included in differentiable models and approximations are essential. Current approaches take an \emph{order sampling} approach to sampling subsets and depend on differentiable approximations of the Top-$k$ operator for selecting the largest $k$ elements from a set. We present a simple alternative method for sampling subsets based on \emph{conditional Poisson sampling}. Unlike order sampling approaches, the parallel complexity of the proposed method is independent of the subset size which makes the method scalable to large subset sizes. We adapt the procedure to make it efficient and amenable to discrete gradient approximations for use in differentiable models. Furthermore, the method also allows the subset size parameter $k$ to be differentiable. We demonstrate our approach on model explanation, image sub-sampling and stochastic $k$-nearest neighbor tasks outperforming existing methods in accuracy, efficiency and scalability.

Deep neural networks trained end-to-end to map a measurement of a (noisy) image to a clean image perform excellent for a variety of linear inverse problems. Current methods are only trained on a few hundreds or thousands of images as opposed to the millions of examples deep networks are trained on in other domains. In this work, we study whether major performance gains are expected from scaling up the training set size. We consider image denoising, accelerated magnetic resonance imaging, and super-resolution and empirically determine the reconstruction quality as a function of training set size, while simultaneously scaling the network size. For all three tasks we find that an initially steep power-law scaling slows significantly already at moderate training set sizes. Interpolating those scaling laws suggests that even training on millions of images would not significantly improve performance. To understand the expected behavior, we analytically characterize the performance of a linear estimator learned with early stopped gradient descent. The result formalizes the intuition that once the error induced by learning the signal model is small relative to the error floor, more training examples do not improve performance.

Recent advances in generative adversarial networks (GANs) have demonstrated the capabilities of generating stunning photo-realistic portrait images. While some prior works have applied such image GANs to unconditional 2D portrait video generation and static 3D portrait synthesis, there are few works successfully extending GANs for generating 3D-aware portrait videos. In this work, we propose PV3D, the first generative framework that can synthesize multi-view consistent portrait videos. Specifically, our method extends the recent static 3D-aware image GAN to the video domain by generalizing the 3D implicit neural representation to model the spatio-temporal space. To introduce motion dynamics into the generation process, we develop a motion generator by stacking multiple motion layers to generate motion features via modulated convolution. To alleviate motion ambiguities caused by camera/human motions, we propose a simple yet effective camera condition strategy for PV3D, enabling both temporal and multi-view consistent video generation. Moreover, PV3D introduces two discriminators for regularizing the spatial and temporal domains to ensure the plausibility of the generated portrait videos. These elaborated designs enable PV3D to generate 3D-aware motion-plausible portrait videos with high-quality appearance and geometry, significantly outperforming prior works. As a result, PV3D is able to support downstream applications such as static portrait animation and view-consistent motion editing. Code and models are available at https://showlab.github.io/pv3d.

Recent text-to-image generation models have shown promising results in generating high-fidelity photo-realistic images. Though the results are astonishing to human eyes, how applicable these generated images are for recognition tasks remains under-explored. In this work, we extensively study whether and how synthetic images generated from state-of-the-art text-to-image generation models can be used for image recognition tasks, and focus on two perspectives: synthetic data for improving classification models in the data-scare settings (i.e. zero-shot and few-shot), and synthetic data for large-scale model pre-training for transfer learning. We showcase the powerfulness and shortcomings of synthetic data from existing generative models, and propose strategies for better applying synthetic data for recognition tasks. Code: https://github.com/CVMI-Lab/SyntheticData.

We propose Make-A-Video -- an approach for directly translating the tremendous recent progress in Text-to-Image (T2I) generation to Text-to-Video (T2V). Our intuition is simple: learn what the world looks like and how it is described from paired text-image data, and learn how the world moves from unsupervised video footage. Make-A-Video has three advantages: (1) it accelerates training of the T2V model (it does not need to learn visual and multimodal representations from scratch), (2) it does not require paired text-video data, and (3) the generated videos inherit the vastness (diversity in aesthetic, fantastical depictions, etc.) of today's image generation models. We design a simple yet effective way to build on T2I models with novel and effective spatial-temporal modules. First, we decompose the full temporal U-Net and attention tensors and approximate them in space and time. Second, we design a spatial temporal pipeline to generate high resolution and frame rate videos with a video decoder, interpolation model and two super resolution models that can enable various applications besides T2V. In all aspects, spatial and temporal resolution, faithfulness to text, and quality, Make-A-Video sets the new state-of-the-art in text-to-video generation, as determined by both qualitative and quantitative measures.

Visualization methods based on the nearest neighbor graph, such as t-SNE or UMAP, are widely used for visualizing high-dimensional data. Yet, these approaches only produce meaningful results if the nearest neighbors themselves are meaningful. For images represented in pixel space this is not the case, as distances in pixel space are often not capturing our sense of similarity and therefore neighbors are not semantically close. This problem can be circumvented by self-supervised approaches based on contrastive learning, such as SimCLR, relying on data augmentation to generate implicit neighbors, but these methods do not produce two-dimensional embeddings suitable for visualization. Here, we present a new method, called t-SimCNE, for unsupervised visualization of image data. T-SimCNE combines ideas from contrastive learning and neighbor embeddings, and trains a parametric mapping from the high-dimensional pixel space into two dimensions. We show that the resulting 2D embeddings achieve classification accuracy comparable to the state-of-the-art high-dimensional SimCLR representations, thus faithfully capturing semantic relationships. Using t-SimCNE, we obtain informative visualizations of the CIFAR-10 and CIFAR-100 datasets, showing rich cluster structure and highlighting artifacts and outliers.

In text generation, models that generate text from scratch one token at a time are currently the dominant paradigm. Despite being performant, these models lack the ability to revise existing text, which limits their usability in many practical scenarios. We look to address this, with DiffusER (Diffusion via Edit-based Reconstruction), a new edit-based generative model for text based on denoising diffusion models -- a class of models that use a Markov chain of denoising steps to incrementally generate data. DiffusER is not only a strong generative model in general, rivalling autoregressive models on several tasks spanning machine translation, summarization, and style transfer; it can also perform other varieties of generation that standard autoregressive models are not well-suited for. For instance, we demonstrate that DiffusER makes it possible for a user to condition generation on a prototype, or an incomplete sequence, and continue revising based on previous edit steps.