Speeding Up Image Classifiers with Little Companions

Scaling up neural networks has been a key recipe to the success of large language and vision models. However, in practice, up-scaled models can be disproportionately costly in terms of computations, providing only marginal improvements in performance; for example, EfficientViT-L3-384 achieves <2% improvement on ImageNet-1K accuracy over the base L1-224 model, while requiring $14\times$ more multiply-accumulate operations (MACs). In this paper, we investigate scaling properties of popular families of neural networks for image classification, and find that scaled-up models mostly help with "difficult" samples. Decomposing the samples by difficulty, we develop a simple model-agnostic two-pass Little-Big algorithm that first uses a light-weight "little" model to make predictions of all samples, and only passes the difficult ones for the "big" model to solve. Good little companion achieve drastic MACs reduction for a wide variety of model families and scales. Without loss of accuracy or modification of existing models, our Little-Big models achieve MACs reductions of 76% for EfficientViT-L3-384, 81% for EfficientNet-B7-600, 71% for DeiT3-L-384 on ImageNet-1K. Little-Big also speeds up the InternImage-G-512 model by 62% while achieving 90% ImageNet-1K top-1 accuracy, serving both as a strong baseline and as a simple practical method for large model compression.

On the Use of Anchoring for Training Vision Models

Anchoring is a recent, architecture-agnostic principle for training deep neural networks that has been shown to significantly improve uncertainty estimation, calibration, and extrapolation capabilities. In this paper, we systematically explore anchoring as a general protocol for training vision models, providing fundamental insights into its training and inference processes and their implications for generalization and safety. Despite its promise, we identify a critical problem in anchored training that can lead to an increased risk of learning undesirable shortcuts, thereby limiting its generalization capabilities. To address this, we introduce a new anchored training protocol that employs a simple regularizer to mitigate this issue and significantly enhances generalization. We empirically evaluate our proposed approach across datasets and architectures of varying scales and complexities, demonstrating substantial performance gains in generalization and safety metrics compared to the standard training protocol.

Target-Aware Generative Augmentations for Single-Shot Adaptation

In this paper, we address the problem of adapting models from a source domain to a target domain, a task that has become increasingly important due to the brittle generalization of deep neural networks. While several test-time adaptation techniques have emerged, they typically rely on synthetic toolbox data augmentations in cases of limited target data availability. We consider the challenging setting of single-shot adaptation and explore the design of augmentation strategies. We argue that augmentations utilized by existing methods are insufficient to handle large distribution shifts, and hence propose a new approach SiSTA, which first fine-tunes a generative model from the source domain using a single-shot target, and then employs novel sampling strategies for curating synthetic target data. Using experiments on a variety of benchmarks, distribution shifts and image corruptions, we find that SiSTA produces significantly improved generalization over existing baselines in face attribute detection and multi-class object recognition. Furthermore, SiSTA performs competitively to models obtained by training on larger target datasets. Our codes can be accessed at https://github.com/Rakshith-2905/SiSTA.

Single-Shot Domain Adaptation via Target-Aware Generative Augmentation

The problem of adapting models from a source domain using data from any target domain of interest has gained prominence, thanks to the brittle generalization in deep neural networks. While several test-time adaptation techniques have emerged, they typically rely on synthetic data augmentations in cases of limited target data availability. In this paper, we consider the challenging setting of single-shot adaptation and explore the design of augmentation strategies. We argue that augmentations utilized by existing methods are insufficient to handle large distribution shifts, and hence propose a new approach SiSTA (Single-Shot Target Augmentations), which first fine-tunes a generative model from the source domain using a single-shot target, and then employs novel sampling strategies for curating synthetic target data. Using experiments with a state-of-the-art domain adaptation method, we find that SiSTA produces improvements as high as 20\% over existing baselines under challenging shifts in face attribute detection, and that it performs competitively to oracle models obtained by training on a larger target dataset.

Domain Alignment Meets Fully Test-Time Adaptation

A foundational requirement of a deployed ML model is to generalize to data drawn from a testing distribution that is different from training. A popular solution to this problem is to adapt a pre-trained model to novel domains using only unlabeled data. In this paper, we focus on a challenging variant of this problem, where access to the original source data is restricted. While fully test-time adaptation (FTTA) and unsupervised domain adaptation (UDA) are closely related, the advances in UDA are not readily applicable to TTA, since most UDA methods require access to the source data. Hence, we propose a new approach, CATTAn, that bridges UDA and FTTA, by relaxing the need to access entire source data, through a novel deep subspace alignment strategy. With a minimal overhead of storing the subspace basis set for the source data, CATTAn enables unsupervised alignment between source and target data during adaptation. Through extensive experimental evaluation on multiple 2D and 3D vision benchmarks (ImageNet-C, Office-31, OfficeHome, DomainNet, PointDA-10) and model architectures, we demonstrate significant gains in FTTA performance. Furthermore, we make a number of crucial findings on the utility of the alignment objective even with inherently robust models, pre-trained ViT representations and under low sample availability in the target domain.

Revisiting Deep Subspace Alignment for Unsupervised Domain Adaptation

Unsupervised domain adaptation (UDA) aims to transfer and adapt knowledge from a labeled source domain to an unlabeled target domain. Traditionally, subspace-based methods form an important class of solutions to this problem. Despite their mathematical elegance and tractability, these methods are often found to be ineffective at producing domain-invariant features with complex, real-world datasets. Motivated by the recent advances in representation learning with deep networks, this paper revisits the use of subspace alignment for UDA and proposes a novel adaptation algorithm that consistently leads to improved generalization. In contrast to existing adversarial training-based DA methods, our approach isolates feature learning and distribution alignment steps, and utilizes a primary-auxiliary optimization strategy to effectively balance the objectives of domain invariance and model fidelity. While providing a significant reduction in target data and computational requirements, our subspace-based DA performs competitively and sometimes even outperforms state-of-the-art approaches on several standard UDA benchmarks. Furthermore, subspace alignment leads to intrinsically well-regularized models that demonstrate strong generalization even in the challenging partial DA setting. Finally, the design of our UDA framework inherently supports progressive adaptation to new target domains at test-time, without requiring retraining of the model from scratch. In summary, powered by powerful feature learners and an effective optimization strategy, we establish subspace-based DA as a highly effective approach for visual recognition.

Improving Multi-Domain Generalization through Domain Re-labeling

Domain generalization (DG) methods aim to develop models that generalize to settings where the test distribution is different from the training data. In this paper, we focus on the challenging problem of multi-source zero-shot DG, where labeled training data from multiple source domains is available but with no access to data from the target domain. Though this problem has become an important topic of research, surprisingly, the simple solution of pooling all source data together and training a single classifier is highly competitive on standard benchmarks. More importantly, even sophisticated approaches that explicitly optimize for invariance across different domains do not necessarily provide non-trivial gains over ERM. In this paper, for the first time, we study the important link between pre-specified domain labels and the generalization performance. Using a motivating case-study and a new variant of a distributional robust optimization algorithm, GroupDRO++, we first demonstrate how inferring custom domain groups can lead to consistent improvements over the original domain labels that come with the dataset. Subsequently, we introduce a general approach for multi-domain generalization, MulDEns, that uses an ERM-based deep ensembling backbone and performs implicit domain re-labeling through a meta-optimization algorithm. Using empirical studies on multiple standard benchmarks, we show that MulDEns does not require tailoring the augmentation strategy or the training process specific to a dataset, consistently outperforms ERM by significant margins, and produces state-of-the-art generalization performance, even when compared to existing methods that exploit the domain labels.

MaAST: Map Attention with Semantic Transformersfor Efficient Visual Navigation

Visual navigation for autonomous agents is a core task in the fields of computer vision and robotics. Learning-based methods, such as deep reinforcement learning, have the potential to outperform the classical solutions developed for this task; however, they come at a significantly increased computational load. Through this work, we design a novel approach that focuses on performing better or comparable to the existing learning-based solutions but under a clear time/computational budget. To this end, we propose a method to encode vital scene semantics such as traversable paths, unexplored areas, and observed scene objects -- alongside raw visual streams such as RGB, depth, and semantic segmentation masks -- into a semantically informed, top-down egocentric map representation. Further, to enable the effective use of this information, we introduce a novel 2-D map attention mechanism, based on the successful multi-layer Transformer networks. We conduct experiments on 3-D reconstructed indoor PointGoal visual navigation and demonstrate the effectiveness of our approach. We show that by using our novel attention schema and auxiliary rewards to better utilize scene semantics, we outperform multiple baselines trained with only raw inputs or implicit semantic information while operating with an 80% decrease in the agent's experience.

Calibrate and Prune: Improving Reliability of Lottery Tickets Through Prediction Calibration

The hypothesis that sub-network initializations (lottery) exist within the initializations of over-parameterized networks, which when trained in isolation produce highly generalizable models, has led to crucial insights into network initialization and has enabled computationally efficient inferencing. In order to realize the full potential of these pruning strategies, particularly when utilized in transfer learning scenarios, it is necessary to understand the behavior of winning tickets when they might overfit to the dataset characteristics. In supervised and semi-supervised learning, prediction calibration is a commonly adopted strategy to handle such inductive biases in models. In this paper, we study the impact of incorporating calibration strategies during model training on the quality of the resulting lottery tickets, using several evaluation metrics. More specifically, we incorporate a suite of calibration strategies to different combinations of architectures and datasets, and evaluate the fidelity of sub-networks retrained based on winning tickets. Furthermore, we report the generalization performance of tickets across distributional shifts, when the inductive biases are explicitly controlled using calibration mechanisms. Finally, we provide key insights and recommendations for obtaining reliable lottery tickets, which we demonstrate to achieve improved generalization.

Invenio: Discovering Hidden Relationships Between Tasks/Domains Using Structured Meta Learning

Exploiting known semantic relationships between fine-grained tasks is critical to the success of recent model agnostic approaches. These approaches often rely on meta-optimization to make a model robust to systematic task or domain shifts. However, in practice, the performance of these methods can suffer, when there are no coherent semantic relationships between the tasks (or domains). We present Invenio, a structured meta-learning algorithm to infer semantic similarities between a given set of tasks and to provide insights into the complexity of transferring knowledge between different tasks. In contrast to existing techniques such as Task2Vec and Taskonomy, which measure similarities between pre-trained models, our approach employs a novel self-supervised learning strategy to discover these relationships in the training loop and at the same time utilizes them to update task-specific models in the meta-update step. Using challenging task and domain databases, under few-shot learning settings, we show that Invenio can discover intricate dependencies between tasks or domains, and can provide significant gains over existing approaches in terms of generalization performance. The learned semantic structure between tasks/domains from Invenio is interpretable and can be used to construct meaningful priors for tasks or domains.