Text-to-image (T2I) diffusion models have become prominent tools for generating high-fidelity images from text prompts. However, when trained on unfiltered internet data, these models can produce unsafe, incorrect, or stylis-tically undesirable images that are not aligned with human preferences. To address this, recent approaches have incor-porated human preference datasets to fine-tune T2I mod-els or to optimize reward functions that capture these pref-erences. Although effective, these methods are vulnerable to reward hacking, where the model overfits to the reward function, leading to a loss of diversity in the generated images. In this paper, we prove the inevitability of reward hacking and study natural regularization techniques like KL divergence and LoRA scaling, and their limitations for diffusion models. We also introduce Annealed Importance Guidance (AIG), an inference-time regularization inspired by Annealed Importance Sampling, which retains the di-versity of the base model while achieving Pareto-Optimal reward-diversity tradeoffs. Our experiments demonstrate the benefits of AIG for Stable Diffusion models, striking the optimal balance between reward optimization and im-age diversity. Furthermore, a user study confirms that AIG improves diversity and quality of generated images across different model architectures and reward functions.

Deep Learning in Image Registration (DLIR) methods have been tremendously successful in image registration due to their speed and ability to incorporate weak label supervision at training time. However, existing DLIR methods forego many of the benefits and invariances of optimization methods. The lack of a task-specific inductive bias in DLIR methods leads to suboptimal performance, especially in the presence of domain shift. Our method aims to bridge this gap between statistical learning and optimization by explicitly incorporating optimization as a layer in a deep network. A deep network is trained to predict multi-scale dense feature images that are registered using a black box iterative optimization solver. This optimal warp is then used to minimize image and label alignment errors. By implicitly differentiating end-to-end through an iterative optimization solver, we explicitly exploit invariances of the correspondence matching problem induced by the optimization, while learning registration and label-aware features, and guaranteeing the warp functions to be a local minima of the registration objective in the feature space. Our framework shows excellent performance on in-domain datasets, and is agnostic to domain shift such as anisotropy and varying intensity profiles. For the first time, our method allows switching between arbitrary transformation representations (free-form to diffeomorphic) at test time with zero retraining. End-to-end feature learning also facilitates interpretability of features and arbitrary test-time regularization, which is not possible with existing DLIR methods.

Classical optimization and learning-based methods are the two reigning paradigms in deformable image registration. While optimization-based methods boast generalizability across modalities and robust performance, learning-based methods promise peak performance, incorporating weak supervision and amortized optimization. However, the exact conditions for either paradigm to perform well over the other are shrouded and not explicitly outlined in the existing literature. In this paper, we make an explicit correspondence between the mutual information of the distribution of per-pixel intensity and labels, and the performance of classical registration methods. This strong correlation hints to the fact that architectural designs in learning-based methods is unlikely to affect this correlation, and therefore, the performance of learning-based methods. This hypothesis is thoroughly validated with state-of-the-art classical and learning-based methods. However, learning-based methods with weak supervision can perform high-fidelity intensity and label registration, which is not possible with classical methods. Next, we show that this high-fidelity feature learning does not translate to invariance to domain shift, and learning-based methods are sensitive to such changes in the data distribution. We reassess and recalibrate performance expectations from classical and DLIR methods under access to label supervision, training time, and its generalization capabilities under minor domain shifts.

Correctness of instance segmentation constitutes counting the number of objects, correctly localizing all predictions and classifying each localized prediction. Average Precision is the de-facto metric used to measure all these constituents of segmentation. However, this metric does not penalize duplicate predictions in the high-recall range, and cannot distinguish instances that are localized correctly but categorized incorrectly. This weakness has inadvertently led to network designs that achieve significant gains in AP but also introduce a large number of false positives. We therefore cannot rely on AP to choose a model that provides an optimal tradeoff between false positives and high recall. To resolve this dilemma, we review alternative metrics in the literature and propose two new measures to explicitly measure the amount of both spatial and categorical duplicate predictions. We also propose a Semantic Sorting and NMS module to remove these duplicates based on a pixel occupancy matching scheme. Experiments show that modern segmentation networks have significant gains in AP, but also contain a considerable amount of duplicates. Our Semantic Sorting and NMS can be added as a plug-and-play module to mitigate hedged predictions and preserve AP.

To build an agent providing assistance to human rescuers in an urban search and rescue task, it is crucial to understand not only human actions but also human beliefs that may influence the decision to take these actions. Developing data-driven models to predict a rescuer’s strategies for navigating the environment and triaging victims requires costly data collection and training for each new environment of interest. Transfer learning approaches can be used to mitigate this challenge, allowing a model trained on a source environment/task to generalize to a previously unseen target environment/task with few training examples. In this paper, we investigate transfer learning (a) from a source environment with smaller number of types of injured victims to one with larger number of victim injury classes and (b) from a smaller and simpler environment to a larger and more complex one for navigation strategy. Inspired by hierarchical organization of human spatial cognition, we used graph division to represent spatial knowledge, and Transfer Learning Diffusion Convo-lutional Recurrent Neural Network (TL-DCRNN), a spatial and temporal graph-based recurrent neural network suitable for transfer learning, to predict navigation. To abstract the rescue strategy from a rescuer’s field-of-view stream, we used attention-based LSTM networks. We experimented on various transfer learning scenarios and evaluated the performance using mean average error. Results indicated our assistant agent can improve predictive accuracy and learn target tasks faster when equipped with transfer learning methods.

Current approaches for semantic image inpainting rely on deep neural networks (DNNs) that learn under full supervision, i.e., using a training set comprising pairs of (i) corrupted images with holes and (ii) corresponding uncorrupted images. However, for several real-world applications, obtaining large sets of uncorrupted images is challenging or infeasible. Current methods also rely on adversarial training involving min-max optimization that is prone to instability during learning. We propose a novel self-supervised image-inpainting DNN framework that can learn in both completely unsupervised and semi-supervised modes. Moreover, our DNN learning formulation bypasses adversarial training and, thereby, lends itself to more stable training. Results on the publicly available CelebA dataset show that our method, even when learning unsupervisedly, outperforms the state of the art that learns with full supervision.