Publications
Understanding Hallucinations in Diffusion Models through Mode Interpolation
Sumukh K Aithal, Pratyush Maini, Zachary C Lipton, J Zico Kolter
In Neural Information Processing Systems (NeurIPS) 2024.
In ICML 2024 Workshop on Geometry-grounded Representation Learning and Generative Modeling and DMLR Workshop. [Paper] [Code]
TLDR
Mode interpolation leads to hallucinations in diffusion models.Abstract
Colloquially speaking, image generation models based upon diffusion processes are frequently said to exhibit “hallucinations”—samples that could never occur in the training data. But where do such hallucinations come from? In this paper, we study a particular failure mode in diffusion models, which we term mode interpolation. Specifically, we find that diffusion models smoothly “interpolate” between nearby data modes in the training set to generate samples that are completely outside the support of the original training distribution; this phenomenon leads diffusion models to generate artifacts that never existed in real data (i.e., hallucinations). We systematically study the reasons for, and the manifestation of this phenomenon. Through experiments on 1D and 2D Gaussians, we show how a discontinuous loss landscape in the diffusion model’s decoder leads to a region where any smooth approximation will cause such hallucinations. Through experiments on artificial datasets with various shapes, we show how hallucination leads to the generation of combinations of shapes that never existed. We extend the validity of mode interpolation in real-world datasets by explaining the unexpected generation of images with additional or missing fingers similar to those produced by popular text-to-image generative models. Finally, we show that diffusion models in fact know when they go out of support and hallucinate. This is captured by the high variance in the trajectory of the generated sample towards the final few backward sampling steps. Using a simple metric to capture this variance, we can remove over 95% of hallucinations at generation time while retaining 96% of in-support samples in the synthetic datasets. We conclude our exploration by showing the implications of such hallucination (and its removal) on the collapse (and stabilization) of recursive training on synthetic data with experiments on MNIST and a 2D Gaussians dataset. We release our code at https://github.com/locuslab/diffusion-model-hallucination.Leveraging the Third Dimension in Contrastive Learning
Sumukh K Aithal, Anirudh Goyal, Alex Lamb, Yoshua Bengio, Michael Mozer
In NeurIPS 2022 Workshop: Self-Supervised Learning - Theory and Practice. [Paper]
TLDR
Depth signal improves self-supervised learning.Abstract
Self-Supervised Learning (SSL) methods operate on unlabeled data to learn robust representations useful for downstream tasks. Most SSL methods rely on augmentations obtained by transforming the 2D image pixel map. These augmentations ignore the fact that biological vision takes place in an immersive three-dimensional, temporally contiguous environment, and that low-level biological vision relies heavily on depth cues. Using a signal provided by a pretrained state-of-the-art monocular RGB-to-depth model (the Depth Prediction Transformer, Ranftl et al., 2021), we explore two distinct approaches to incorporating depth signals into the SSL framework. First, we evaluate contrastive learning using an RGB+depth input representation. Second, we use the depth signal to generate novel views from slightly different camera positions, thereby producing a 3D augmentation for contrastive learning. We evaluate these two approaches on three different SSL methods—BYOL, SimSiam, and SwAV—using ImageNette (10 class subset of ImageNet), ImageNet-100 and ImageNet-1k datasets. We find that both approaches to incorporating depth signals improve the robustness and generalization of the baseline SSL methods, though the first approach (with depth-channel concatena- tion) is superior. For instance, BYOL with the additional depth channel leads to an increase in downstream classification accuracy from 85.3% to 88.0% on ImageNette and 84.1% to 87.0% on ImageNet-C.Escaping Saddle Points for Effective Generalization on Class-Imbalanced Data
Harsh Rangwani*, Sumukh K Aithal*, Mayank Mishra, R. Venkatesh Babu
In Neural Information Processing Systems (NeurIPS) 2022. [Paper] [Code] [Slides] [Poster]
TLDR
Tail class loss landscape converges to a saddle point in imbalanced datasets and SAM can effectively escape from these solutions.Abstract
Real-world datasets exhibit imbalances of varying types and degrees. Several techniques based on re-weighting and margin adjustment of loss are often used to enhance the performance of neural networks, particularly on minority classes. In this work, we analyze the class-imbalanced learning problem by examining the loss landscape of neural networks trained with re-weighting and margin based techniques. Specifically, we examine the spectral density of Hessian of class-wise loss, through which we observe that the network weights converges to a saddle point in the loss landscapes of minority classes. Following this observation, we also find that optimization methods designed to escape from saddle points can be effectively used to improve generalization on minority classes. We further theoretically and empirically demonstrate that Sharpness-Aware Minimization (SAM), a recent technique that encourages convergence to a flat minima, can be effectively used to escape saddle points for minority classes. Using SAM results in a 6.2% increase in accuracy on the minority classes over the state-of-the-art Vector Scaling Loss, leading to an overall average increase of 4% across imbalanced datasets. The code is available at https://github.com/val-iisc/Saddle-LongTail.A Closer Look at Smoothness at Domain Adversarial Training
Harsh Rangwani*, Sumukh K Aithal*, Mayank Mishra, Arihant Jain, R. Venkatesh Babu
In International Conference on Machine Learning (ICML) 2022. [Paper] [Code] [Slides] [Poster]
TLDR
Smooth Minima with respect to task loss leads to effective generalization on the target domain.Abstract
Domain adversarial training has been ubiquitous for achieving invariant representations and is used widely for various domain adaptation tasks. In recent times, methods converging to smooth optima have shown improved generalization for supervised learning tasks like classification. In this work, we analyze the effect of smoothness enhancing formulations on domain adversarial training, the objective of which is a combination of task loss (eg. classification, regression etc.) and adversarial terms. We find that converging to a smooth minima with respect to (w.r.t.) task loss stabilizes the adversarial training leading to better performance on target domain. In contrast to task loss, our analysis shows that converging to smooth minima w.r.t. adversarial loss leads to sub-optimal generalization on the target domain. Based on the analysis, we introduce the Smooth Domain Adversarial Training (SDAT) procedure, which effectively enhances the performance of existing domain adversarial methods for both classification and object detection tasks. Our analysis also provides insight into the extensive usage of SGD over Adam in the community for domain adversarial training.S3VAADA: Submodular Subset Selection for Virtual Adversarial Active Domain Adaptation
Harsh Rangwani, Arihant Jain*, Sumukh K Aithal*, R. Venkatesh Babu
In International Conference on Computer Vision (ICCV) 2021. [Project Website]
TLDR
Informative sample selection and effective adaptation through S3VAADA can lead to effective gains by using small amount of labeled target data.Abstract
Unsupervised domain adaptation (DA) methods have focused on achieving maximal performance through aligning features from source and target domains without using labeled data in the target domain. Whereas, in the real-world scenario’s it might be feasible to get labels for a small proportion of target data. In these scenarios, it is important to select maximally-informative samples to label and find an effective way to combine them with the existing knowledge from source data. Towards achieving this, we propose S3VAADA which i) introduces a novel submodular criterion to select a maximally informative subset to label and ii) enhances a cluster-based DA procedure through novel improvements to effectively utilize all the available data for improving generalization on target. Our approach consistently outperforms the competing state-of-the-art approaches on datasets with varying degrees of domain shifts.Methods and Analysis of The First Competition in Predicting Generalization of Deep Learning
Yiding Jiang, Parth Natekar, Manik Sharma, Sumukh K Aithal, Dhruva Kashyap, Natarajan Subramanyam, Carlos Lassance, Daniel M. Roy, Gintare Karolina Dziugaite, Suriya Gunasekar, Isabelle Guyon, Pierre Foret, Scott Yak, Hossein Mobahi, Behnam Neyshabur, Samy Bengio
PMLR: NeurIPS 2020 Competition and Demonstration Track, 2020 [Paper]
TLDR
A summary of the solutions of the top-three teams in the PGDL Competition.Abstract
Deep learning has been recently successfully applied to an ever larger number of problems, ranging from pattern recognition to complex decision making. However, several concerns have been raised, including guarantees of good generalization, which is of foremost importance. Despite numerous attempts, conventional statistical learning approaches fall short of providing a satisfactory explanation on why deep learning works. In a competition hosted at the Thirty-Fourth Conference on Neural Information Processing Systems (NeurIPS 2020), we invited the community to design robust and general complexity measures that can accurately predict the generalization of models. In this paper, we describe the competition design, the protocols, and the solutions of the top-three teams at the competition in details. In addition, we discuss the outcomes, common failure modes, and potential future directions for the competition.Robustness to Augmentations as a Generalization Metric
Sumukh Aithal K*, Dhruva Kashyap *, Natarajan Subramanyam
1st Runner Up in Predicting Generalization in Deep Learning, NeurIPS 2020 Competition Track [Paper] [Video] [Code]
TLDR
In this work, we developed a simple yet effective method to predict the generalization performance of a model by using the concept that models that are robust to augmentations are more generalizable than those which are not.Abstract
Generalization is the ability of a model to predict on unseen domains and is a fundamental task in machine learning. Several generalization bounds, both theoretical and empirical have been proposed but they do not provide tight bounds. In this work, we propose a simple yet effective method to predict the generalization performance of a model by using the concept that models that are robust to augmentations are more generalizable than those which are not. We experiment with several augmentations and composition of augmentations to check the generalization capacity of a model. We also provide a detailed motivation behind the proposed method. The proposed generalization metric is calculated based on the change in the model’s output after augmenting the input. The proposed method was the first runner up solution for the competition "Predicting Generalization in Deep Learning".Domain Shift in Capsule Networks
Rajath S*, Sumukh Aithal K*, Natarajan Subramanyam
10th International Conference on Pattern Recognition Applications and Methods (ICPRAM 2021) [Paper]