Tejas Gokhale

A comprehensive overview of the state-of-the-art methods in multimodal retrieval, generation, and retrieval-augmented generation.

Series Title: Synthesis Lectures on Computer Vision

Hardcover ISBN: 978-3-031-57816-8; Softcover ISBN978-3-031-57818-2; eBook ISBN: 978-3-031-57816-8

We identify that the tendency of existing pooling layers in CNNs to pass larger signals to subsequent layers is a major factor that's strongly correlated with the lack of shift invariance in CNNs. Based on this finding, we design a new pooling operator Translation-Invariant Polyphase Sampling (TIPS) and two regularizations on the intermediate feature maps to learn translation-invariant representations. TIPS results in consistent and architecture-agnostic improvements in accuracy and four measures of shift invariance, across multiple image classification and segmentation benchmarks.

A method for generating ``hard'' negative captions via in-context learning and synthesizing corresponding negative images with text-to-image generators. A novel contrastive pre-training strategy that leverages these hard negative captions and images in an alternating fashion to train CLIP. Our method "TripletCLIP" enhances the compositional capabilities of CLIP as well as improvements in zero-shot image classification and image retrieval.

We improve spatial understanding of T2I models by creating the SPRIGHT dataset by recaptioning 6M images from widely used vision datasets. Finetuning T2I models with just 500 images from SPRIGHT leads to a large improvement in T2I spatial understanding performance, across several evaluation benchmarks such as T2I-CompBench, VISOR, and GenEval.

Traditional generative Text-to-Image models struggle to generate images that faithfully represent the spatial relationships mentioned in the input prompt. We develop REVISION, an efficient rendering pipeline that enables a training-free, guidance-based mechanism to address this shortcoming. REVISION takes the objects and their spatial relationships parsed from the given input prompt and synthesizes an image in Blender, placing the respective object assets at coordinates corresponding to the parsed spatial relationship. Given a user-provided input prompt T, we synthesize an image using REVISION and use it to guide existing T2I pipelines such as Stable Diffusion or ControlNet to obtain a spatially accurate output.

The motivation of this work is to analyze the efficacy of language guidance for low-level non-semantic computer vision tasks. We focus on depth estimation and find that language-guided depth estimators benefit only from scene-level language information and counter-intuitively, are worse off when presented with sentences that describe 3D spatial relationships in the scene. With an increasing number of methods using language for depth estimation, our findings highlight the opportunities and pitfalls that require careful consideration for effective deployment in real-world settings.

Two new quantitative measures ICV and IDD to describe domain shifts in terms of consistency of classes within one domain and similarity between two stylistic domains. New dataset: SuperMarioDomains (SMD) incorporating unique features of consistent classes of video game scenes across stylistic domains in video game graphics that are dissimilar to ImageNet1K.

Textual inversion models have the potential to learn novel concepts from a small number of example images. We quantify this concept learning ability with ConceptBed: a dataset that contains 284 unique visual concepts and 33K concept compositions, and CCD (Concept Confidence Deviation): an evaluation metric uses the confidence of oracle concept classifiers to measure the alignment between generated images and concepts contained in ground truth images.

ABA draws on the strengths of adversarial learning and Bayesian neural networks to guide the generation of diverse data augmentations — these synthesized image domains aid the classifier in generalizing to several types of domain shift including style shift, subpopulation shift, and domain shift in the medical imaging setting. ABA outperforms all previous state-of-the-art methods, including pre-specified augmentations, pixel-based and convolutional-based augmentations.

This dissertation contributes to the reliability of machine learning models from several perspectives including the development of robust training algorithms to mitigate the risks of such failures, construction of new datasets that provide a new perspective on capabilities of vision models, and the design of evaluation metrics for re-calibrating the perception of performance improvements.

Knowledge retrieval with multi-modal queries, i.e., queries containing information split across image and text inputs, a challenging task that differs from previous work on cross-modal retrieval. A new dataset called ReMuQ, a new pretraining task for learning knowledge retrieval with multimodal queries, and a retriever model "ReViz" that can directly process input text and images to retrieve relevant knowledge in an end-to-end fashion without being dependent on intermediate modules such as object detectors or caption generators.

A data poisoning attack that confounds ML models without any manipulation of the image or label, achieved by simply leveraging the most confounding natural samples found within the training data itself. We show the efficacy of this novel attack in offline as well as continual learning (CL) settings in image classification, thereby exposing a previously undetected vulnerability of image classifiers.

We report a surprising finding that, although recent state-of-the-art T2I models exhibit high image quality, they are severely limited in their ability to generate multiple objects or the specified spatial relations such as left/right/above/below. We introduce a metric called VISOR to quantify spatial reasoning performance. VISOR can be used off-the-shelf with any text-to-image model. We construct and make available SR2D, a dataset which contains sentences that describe spatial relationships (left/right/above/below) between a pair of commonly occurring objects.

ALT discovers diverse and adversarial transformations using an image-to-image neural network with learnable weights. ALT improves the state-of-the-art single domain generalization performance on three benchmarks and is significantly better than pixel-wise adversarial training and standard data augmentation techniques.

Although the imaging pipeline is unable to capture many physical properties of objects (eg. mass and coefficient of friction), these properties can be estimated by utilizing cues introduced by collisions. We introduce a new dataset (CRIPP-VQA) for reasoning about the implicit physical properties of objects from videos. The dataset contains videos of objects in motion, annotated with hypothetical/counterfactual questions about the effect of actions (removing/adding/replacing objects) and questions about planning (performing actions to reach a goal).

In this paper, we introduce a benchmark for covariate shift detection (CSD), that builds upon and complements previous work on domain generalization. We find that existing novelty detection methods designed for OOD benchmarks perform worse than simple confidence-based methods on our CSD benchmark. We propose Domain Interpolation Sensitivity (DIS), based on the simple hypothesis that interpolation between the test input and randomly sampled inputs from the training domain, offers sufficient information to distinguish between the training domain and unseen domains under covariate shift.

Natural Language Inference (NLI) under three low-data settings (with missing labels; with missing labels and hypothesis; and with missing labels, hypotheses, and premises). A procedural data generation approach that leverages a set of sentence transformations to collect PHL (Premise, Hypothesis, Label) triplets for training NLI models, bypassing the need for human-annotated training data. State-of-the-art results under all three "unsupervised" settings.

SDRO: a distributed robust optimization method that operates with linguistic transformations of sentence inputs, SISP: a suit of semantics-inverting (SI) and semantics-preserving (SP) linguistic transformations, and an ensembling technique for vision-and-language inference.

In this work, we conduct a comprehensive study of common data modification strategies and evaluate not only their in-domain and OOD performance, but also their adversarial robustness (AR). This work serves as an empirical study towards understanding the relationship between generalizing to unseen domains and defending against adversarial perturbations.

We present a debiased dataset for the Person Centric Visual Grounding (PCVG) task. For instance, in many cases the first name in the sentence corresponds to the largest bounding box, or the sequence of names in the sentence corresponds to an exact left-to-right order in the image). The debiased dataset offers the PCVG task a more practical baseline for reliable benchmarking and future improvements.

We seek to improve information retrieval (IR) using neural retrievers (NR) in the biomedical domain, using a three-pronged approach. (1) a template-based question generation method, (2) two novel pre-training tasks that are closely aligned to the downstream task of information retrieval, (3) the ``Poly-DPR'' model which encodes each context into multiple context vectors.

VQA models trained with two additional objectives: object centroid estimation and relative position estimation, lead to improved performance on spatial reasoning questions (in GQA) in fully supervised and few shot settings as well as improved O.O.D. generalization.

We show that models can be trained without any human-annotated Q-A pairs, but only with images and associated text captions. Our experiments suggest gains on benchmark with shifted priors (VQA-CP) over baselines which use full supervision from human-authored QA data.

Unsupervised Reading Comprehension method that operates directly on a single test passage. Synthetic QA pairs are generated from the passage, and models are trained on these. When a new human-authored test question appears, models infer answers better than previous unsupervised methods.

An adversarial training approach which learns to generate new samples so as to maximize exposure of the classifier to attributes-space. Studies robustness to semantic shifts that are beyond L-p norm perturbations, on 3 types of naturally occurring perturbations — object-related shifts, geometric transformations, and common image corruptions.

MUTANT is a training paradigm that exposes VQA models to perceptually similar, yet semantically distinct mutations of the input image or question. We use a pairwise consistency loss between answers to original and mutant inputs as a regularization, along with an answer embedding NCE loss. MUTANTimproves generalization of VQA models under Changing Priors.

Actions in videos are inherently linked to latent social and commonsense aspects. We present the first work on generating commonsense captions directly from videos, to describe latent intentions, attributes, and effects of humans in videos. Additionally we explore the use of open-ended video-based commonsense question answering (V2C-QA) as a way to enrich our captions.

VQA models struggle at negation, antonyms, conjunction, disjunction! We show a capability of answering logically composed questions with our novel modules and datasets, while retaining performance on VQA data.