ICLR 2023 Papers

We prove that image classifiers are fundamentally sensitive to small perturbations in their inputs. Specifically, we show that given some image space of $n$-by-$n$ images, all but a tiny fraction of images in any image class induced over that space can be moved outside that class by adding some perturbation whose $p$-norm is $O(n^{1/\max{(p,1)}})$, as long as that image class takes up at most half of the image space. We then show that $O(n^{1/\max{(p,1)}})$ is asymptotically optimal. Finally, we show that an increase in the bit depth of the image space leads to a loss in robustness. We supplement our results with a discussion of their implications for vision systems.

Large pre-trained models exhibit distinct and complementary capabilities dependent on the data they are trained on. Language models such as GPT-3 are capable of textual reasoning but cannot understand visual information, while vision models such as DALL-E can generate photorealistic photos but fail to understand complex language descriptions. In this work, we propose a unified framework for composing ensembles of different pre-trained models -- combining the strengths of each individual model to solve various multimodal problems in a zero-shot manner. We use pre-trained models as "generators" or "scorers" and compose them via closed-loop iterative consensus optimization. The generator constructs proposals and the scorers iteratively provide feedback to refine the generated result. Such closed-loop communication enables models to correct errors caused by other models, significantly boosting performance on downstream tasks, e.g. improving accuracy on grade school math problems by 7.5%, without requiring any model finetuning. We demonstrate that consensus achieved by an ensemble of scorers outperforms the feedback of a single scorer, by leveraging the strengths of each expert model. Results show that the proposed method can be used as a general purpose framework for a wide range of zero-shot multimodal tasks, such as image generation, video question answering, mathematical reasoning, and robotic manipulation.

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.

Data augmentation has recently emerged as an essential component of modern training recipes for visual recognition tasks. However, data augmentation for video recognition has been rarely explored despite its effectiveness. Few existing augmentation recipes for video recognition naively extend the image augmentation methods by applying the same operations to the whole video frames. Our main idea is that the magnitude of augmentation operations for each frame needs to be changed over time to capture the real-world video's temporal variations. These variations should be generated as diverse as possible using fewer additional hyper-parameters during training. Through this motivation, we propose a simple yet effective video data augmentation framework, DynaAugment. The magnitude of augmentation operations on each frame is changed by an effective mechanism, Fourier Sampling that parameterizes diverse, smooth, and realistic temporal variations. DynaAugment also includes an extended search space suitable for video for automatic data augmentation methods. DynaAugment experimentally demonstrates that there are additional performance rooms to be improved from static augmentations on diverse video models. Specifically, we show the effectiveness of DynaAugment on various video datasets and tasks: large-scale video recognition (Kinetics-400 and Something-Something-v2), small-scale video recognition (UCF-101 and HMDB-51), fine-grained video recognition (Diving-48 and FineGym), video action segmentation on Breakfast, video action localization on THUMOS'14, and video object detection on MOT17Det.

It is expensive to collect training data for every possible domain that a vision model may encounter when deployed. We instead consider how simply $\textit{verbalizing}$ the training domain (e.g.``photos of birds'') as well as domains we want to extend to but do not have data for (e.g.``paintings of birds'') can improve robustness. Using a multimodal model with a joint image and language embedding space, our method $\textit{LADS}$ learns a transformation of the image embeddings from the source domain to each target domain, while preserving task relevant information. Without using any images from the target domain, we show that over the $\textit{extended}$ domain containing both source and target, $\textit{LADS}$ outperforms standard fine-tuning and ensemble approaches over a suite of 4 benchmarks targeting domain adaptation and dataset bias.

Asymmetric image retrieval aims to deploy compatible models on platforms of different resources to achieve a balance between computational efficiency and retrieval accuracy. The most critical issue is how to align the output features of different models. Despite the great progress, existing approaches apply strong constraints so that features or neighbor structures are strictly aligned across different models. However, such a one-to-one constraint is too strict to be well preserved for the query models with low capacity. Considering that the primary concern of the users is the rank of the returned images, we propose a generic rank preserving framework, which achieves feature compatibility and the order consistency between query and gallery models simultaneously. Specifically, we propose two alternatives to instantiate the framework. One realizes straightforward rank order preservation by directly preserving the consistency of the sorting results. To make sorting process differentiable, the Heaviside step function in sorting is approximated by the sigmoid function. The other aims to preserve a learnable monotonic mapping relationship between the returned similarity scores of query and gallery models. The mapped similarity scores of gallery model are considered as pseudo-supervision to guide the query model training. Extensive experiments on various large-scale datasets demonstrate the superiority of our two proposed methods.

The recently proposed data augmentation TransMix employs attention labels to help visual transformers (ViT) achieve better robustness and performance. However, TransMix is deficient in two aspects: 1) The image cropping method of TransMix may not be suitable for vision transformer. 2) At the early stage of training, the model produces unreliable attention maps. TransMix uses unreliable attention maps to compute mixed attention labels that can affect the model. To address the aforementioned issues, we propose MaskMix and Progressive Attention Labeling (PAL) in image and label space, respectively. In detail, from the perspective of image space, we design MaskMix, which mixes two images based on a patch-like grid mask. In particular, the size of each mask patch is adjustable and is a multiple of the image patch size, which ensures each image patch comes from only one image and contains more global contents. From the perspective of label space, we design PAL, which utilizes a progressive factor to dynamically re-weight the attention weights of the mixed attention label. Finally, we combine MaskMix and Progressive Attention Labeling as our new data augmentation method, named MixPro. The experimental results show that our method can improve various ViT-based models at scales on ImageNet classification (73.8% top-1 accuracy based on DeiT-T for 300 epochs). After being pre-trained with MixPro on ImageNet, the ViT-based models also demonstrate better transferability to semantic segmentation, object detection, and instance segmentation. Furthermore, compared to TransMix, MixPro also shows stronger robustness on several benchmarks.

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 present a novel neural-networks-based algorithm to compute optimal transport maps and plans for strong and weak transport costs. To justify the usage of neural networks, we prove that they are universal approximators of transport plans between probability distributions. We evaluate the performance of our optimal transport algorithm on toy examples and on the unpaired image-to-image translation.

Self-distillation methods using Siamese networks are popular for self-supervised pre-training. DINO is one such method based on a cross-entropy loss between $K$-dimensional probability vectors, obtained by applying a softmax function to the dot product between representations and learnt prototypes. Given the fact that the learned representations are $L^2$-normalized, we show that DINO and its derivatives, such as iBOT, can be interpreted as a mixture model of von Mises-Fisher components. With this interpretation, DINO assumes equal precision for all components when the prototypes are also $L^2$-normalized. Using this insight we propose DINO-vMF, that adds appropriate normalization constants when computing the cluster assignment probabilities. Unlike DINO, DINO-vMF is stable also for the larger ViT-Base model with unnormalized prototypes. We show that the added flexibility of the mixture model is beneficial in terms of better image representations. The DINO-vMF pre-trained model consistently performs better than DINO on a range of downstream tasks. We obtain similar improvements for iBOT-vMF vs iBOT and thereby show the relevance of our proposed modification also for other methods derived from DINO.

We present a Lipschitz continuous Transformer, called LipsFormer, to pursue training stability both theoretically and empirically for Transformer-based models. In contrast to previous practical tricks that address training instability by learning rate warmup, layer normalization, attention formulation, and weight initialization, we show that Lipschitz continuity is a more essential property to ensure training stability. In LipsFormer, we replace unstable Transformer component modules with Lipschitz continuous counterparts: CenterNorm instead of LayerNorm, spectral initialization instead of Xavier initialization, scaled cosine similarity attention instead of dot-product attention, and weighted residual shortcut. We prove that these introduced modules are Lipschitz continuous and derive an upper bound on the Lipschitz constant of LipsFormer. Our experiments show that LipsFormer allows stable training of deep Transformer architectures without the need of careful learning rate tuning such as warmup, yielding a faster convergence and better generalization. As a result, on the ImageNet 1K dataset, LipsFormer-Tiny training for 100 epochs without learning rate warmup attains a top-1 accuracy of 81.6\% which is higher than Swin Transformer-Tiny training for 300 epochs with warmup. Moreover, LipsFormer-Tiny training for 300 epochs achieves a top-1 accuracy of 83.5\% with 4.7G FLOPs and 24M parameters.

Transformers flexibly operate over sets of real-valued vectors representing task-specific entities and their attributes, where each vector might encode one word-piece token and its position in a sequence, or some piece of information that carries no position at all. As set processors, transformers are at a disadvantage in reasoning over more general graph-structured data where nodes represent entities and edges represent relations between entities. To address this shortcoming, we generalize transformer attention to consider and update edge vectors in each transformer layer. We evaluate this relational transformer on a diverse array of graph-structured tasks, including the large and challenging CLRS Algorithmic Reasoning Benchmark. There, it dramatically outperforms state-of-the-art graph neural networks expressly designed to reason over graph-structured data. Our analysis demonstrates that these gains are attributable to relational attention's inherent ability to leverage the greater expressivity of graphs over sets.

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.

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.

We present a comparative study on how and why contrastive learning (CL) and masked image modeling (MIM) differ in their representations and in their performance of downstream tasks. In particular, we demonstrate that self-supervised Vision Transformers (ViTs) have the following properties: (1) CL trains self-attentions to capture longer-range global patterns than MIM, such as the shape of an object, especially in the later layers of the ViT architecture. This CL property helps ViTs linearly separate images in their representation spaces. However, it also makes the self-attentions collapse into homogeneity for all query tokens and heads. Such homogeneity of self-attention reduces the diversity of representations, worsening scalability and dense prediction performance. (2) CL utilizes the low-frequency signals of the representations, but MIM utilizes high-frequencies. Since low- and high-frequency information respectively represent shapes and textures, CL is more shape-oriented and MIM more texture-oriented. (3) CL plays a crucial role in the later layers, while MIM mainly focuses on the early layers. Upon these analyses, we find that CL and MIM can complement each other and observe that even the simplest harmonization can help leverage the advantages of both methods. The code is available at https://github.com/naver-ai/cl-vs-mim.

What is an image, and how to extract latent features? Convolutional Networks (ConvNets) consider an image as organized pixels in a rectangular shape and extract features via convolutional operation in a local region; Vision Transformers (ViTs) treat an image as a sequence of patches and extract features via attention mechanism in a global range. In this work, we introduce a straightforward and promising paradigm for visual representation, which is called Context Clusters. Context clusters (CoCs) view an image as a set of unorganized points and extract features via a simplified clustering algorithm. In detail, each point includes the raw feature (e.g., color) and positional information (e.g., coordinates), and a simplified clustering algorithm is employed to group and extract deep features hierarchically. Our CoCs are convolution- and attention-free, only relying on clustering algorithm for spatial interaction. Owing to the simple design, we show CoCs endow gratifying interpretability via the visualization of the clustering process. Our CoCs aim at providing a new perspective on image and visual representation, which may enjoy broad applications in different domains and exhibit profound insights. Even though we are not targeting SOTA performance, COCs still achieve comparable or even better performance than ConvNets or ViTs on several benchmarks.

Few-shot prompting is a surprisingly powerful way to use Large Language Models (LLMs) to solve various tasks. However, this approach struggles as the task complexity increases or when the individual reasoning steps of the task themselves are hard to learn, especially when embedded in more complex tasks. To address this, we propose Decomposed Prompting, a new approach to solve complex tasks by decomposing them (via prompting) into simpler sub-tasks that can be delegated to a library of prompting-based LLMs dedicated to these sub-tasks. This modular structure allows each prompt to be optimized for its specific sub-task, further decomposed if necessary, and even easily replaced with more effective prompts, trained models, or symbolic functions if desired. We show that the flexibility and modularity of Decomposed Prompting allows it to outperform prior work on few-shot prompting using GPT3. On symbolic reasoning tasks, we can further decompose sub-tasks that are hard for LLMs into even simpler solvable sub-tasks. When the complexity comes from the input length, we can recursively decompose the task into the same task but with smaller inputs. We also evaluate our approach on textual multi-step reasoning tasks: on long-context multi-hop QA task, we can more effectively teach the sub-tasks via our separate sub-tasks prompts; and on open-domain multi-hop QA, we can incorporate a symbolic information retrieval within our decomposition framework, leading to improved performance on both tasks. Datasets, Code and Prompts available at https://github.com/allenai/DecomP.

Predicting the pose of objects from a single image is an important but difficult computer vision problem. Methods that predict a single point estimate do not predict the pose of objects with symmetries well and cannot represent uncertainty. Alternatively, some works predict a distribution over orientations in $\mathrm{SO}(3)$. However, training such models can be computation- and sample-inefficient. Instead, we propose a novel mapping of features from the image domain to the 3D rotation manifold. Our method then leverages $\mathrm{SO}(3)$ equivariant layers, which are more sample efficient, and outputs a distribution over rotations that can be sampled at arbitrary resolution. We demonstrate the effectiveness of our method at object orientation prediction, and achieve state-of-the-art performance on the popular PASCAL3D+ dataset. Moreover, we show that our method can model complex object symmetries, without any modifications to the parameters or loss function. Code is available at \url{https://dmklee.github.io/image2sphere}.