We propose CLIP-Lite, an information efficient method for visual representation learning by feature alignment with textual annotations. Compared to the previously proposed CLIP model, CLIP-Lite requires only one negative image-text sample pair for every positive image-text sample during the optimization of its contrastive learning objective. We accomplish this by taking advantage of an information efficient lower-bound to maximize the mutual information between the two input modalities. This allows CLIP-Lite to be trained with significantly reduced amounts of data and batch sizes while obtaining better performance than CLIP at the same scale. We evaluate CLIP-Lite by pretraining on the COCO-Captions dataset and testing transfer learning to other datasets. CLIP-Lite obtains a +14.0$%$ mAP absolute gain in performance on Pascal VOC classification, and a +22.1$%$ top-1 accuracy gain on ImageNet, while being comparable or superior to other, more complex, text-supervised models. CLIP-Lite is also superior to CLIP on image and text retrieval, zero-shot classification, and visual grounding. Finally, we show that CLIP-Lite can leverage language semantics to encourage bias-free visual representations that can be used in downstream tasks. Implementation: https://github.com/4m4n5/CLIP-Lite

Videos are a rich source for self-supervised learning (SSL) of visual representations due to the presence of natural temporal transformations of objects. However, current methods typically randomly sample video clips for learning, which results in an imperfect supervisory signal. In this work, we propose PreViTS, an SSL framework that utilizes an unsupervised tracking signal for selecting clips containing the same object, which helps better utilize temporal transformations of objects. PreViTS further uses the tracking signal to spatially constrain the frame regions to learn from and trains the model to locate meaningful objects by providing supervision on Grad-CAM attention maps. To evaluate our approach, we train a momentum contrastive (MoCo) encoder on VGG-Sound and Kinetics-400 datasets with PreViTS. Training with PreViTS outperforms representations learnt by contrastive strategy alone on video downstream tasks, obtaining state-of-the-art performance on action classification. PreViTS helps learn feature representations that are more robust to changes in background and context, as seen by experiments on datasets with background changes. Our experiment also demonstrates various visual transformation invariance captured by our model. Learning from large-scale videos with PreViTS could lead to more accurate and robust visual feature representations.

Recent advances in self-supervised learning (SSL) have largely closed the gap with supervised ImageNet pretraining. Despite their success these methods have been primarily applied to unlabeled ImageNet images, and show marginal gains when trained on larger sets of uncurated images. We hypothesize that current SSL methods perform best on iconic images, and struggle on complex scene images with many objects. Analyzing contrastive SSL methods shows that they have poor visual grounding and receive poor supervisory signal when trained on scene images. We propose Contrastive Attention-Supervised Tuning (CAST) to overcome these limitations. CAST uses unsupervised saliency maps to intelligently sample crops, and to provide grounding supervision via a Grad-CAM attention loss. Experiments on COCO show that CAST significantly improves the features learned by SSL methods on scene images, and further experiments show that CAST-trained models are more robust to changes in backgrounds. Our code is available at https://github.com/salesforce/CAST/.

Sameer Dharur, Purva Tendulkar, Dhruv Batra, Devi Parikh, Ramprasaath R. Selvaraju. Proceedings of the 2021 Conference of the North American Chapter of the Association for Computational Linguistics: Human Language Technologies. 2021.

Large-scale vision and language representation learning has shown promising improvements on various vision-language tasks. Most existing methods employ a transformer-based multimodal encoder to jointly model visual tokens (region-based image features) and word tokens. Because the visual tokens and word tokens are unaligned, it is challenging for the multimodal encoder to learn image-text interactions. In this paper, we introduce a contrastive loss to ALign the image and text representations BEfore Fusing (ALBEF) them through cross-modal attention, which enables more grounded vision and language representation learning. Unlike most existing methods, our method does not require bounding box annotations nor high-resolution images. In order to improve learning from noisy web data, we propose momentum distillation, a self-training method which learns from pseudo-targets produced by a momentum model. We provide a theoretical analysis of ALBEF from a mutual information maximization perspective, showing that different training tasks can be interpreted as different ways to generate views for an image-text pair. ALBEF achieves state-of-the-art performance on multiple downstream vision-language tasks. On image-text retrieval, ALBEF outperforms methods that are pre-trained on orders of magnitude larger datasets. On VQA and NLVR$^2$, ALBEF achieves absolute improvements of 2.37% and 3.84% compared to the state-of-the-art, while enjoying faster inference speed. Code and models are available at https://github.com/salesforce/ALBEF.

Existing VQA datasets contain questions with varying levels of complexity. While the majority of questions in these datasets require perception for recognizing existence, properties, and spatial relationships of entities, a significant portion of questions pose challenges that correspond to reasoning tasks - tasks that can only be answered through a synthesis of perception and knowledge about the world, logic and / or reasoning. Analyzing performance across this distinction allows us to notice when existing VQA models have consistency issues - they answer the reasoning questions correctly but fail on associated low-level perception questions. For example, in Figure 1, models answer the complex reasoning question “Is the banana ripe enough to eat?” correctly, but fail on the associated perception question “Are the bananas mostly green or yellow?” indicating that the model likely answered the reasoning question correctly but for the wrong reason. We quantify the extent to which this phenomenon occurs by creating a new Reasoning split of the VQA dataset and collecting VQAintrospect, a new dataset1 which currently consists of 200K new perception questions which serve as sub questions corresponding to the set of perceptual tasks needed to effectively answer the complex reasoning questions in the Reasoning split. Our evaluation shows that state-of-the-art VQA models have comparable performance in answering perception and reasoning questions, but suffer from consistency problems. To address this shortcoming, we propose an approach called Sub-Question Importance-aware Network Tuning (SQuINT), which encourages the model to attend to the same parts of the image when answering the reasoning question and the perception sub question. We show that SQuINT improves model consistency by ~7%, also marginally improving performance on the Reasoning questions in VQA, while also displaying better attention maps.

Many vision and language models suffer from poor visual grounding -- often falling back on easy-to-learn language priors rather than basing their decisions on visual concepts in the image. In this work, we propose a generic approach called Human Importance-aware Network Tuning (HINT) that effectively leverages human demonstrations to improve visual grounding. HINT encourages deep networks to be sensitive to the same input regions as humans. Our approach optimizes the alignment between human attention maps and gradient-based network importances -- ensuring that models learn not just to look at but rather rely on visual concepts that humans found relevant for a task when making predictions. We apply HINT to Visual Question Answering and Image Captioning tasks, outperforming top approaches on splits that penalize over-reliance on language priors (VQA-CP and robust captioning) using human attention demonstrations for just 6% of the training data.

A single image captures the appearance and position of multiple entities in a scene as well as their complex interactions. As a consequence, natural language grounded in visual contexts tends to be diverse---with utterances differing as focus shifts to specific objects, interactions, or levels of detail. Recently, neural sequence models such as RNNs and LSTMs have been employed to produce visually-grounded language. Beam Search, the standard work-horse for decoding sequences from these models, is an approximate inference algorithm that decodes the top-B sequences in a greedy left-to-right fashion. In practice, the resulting sequences are often minor rewordings of a common utterance, failing to capture the multimodal nature of source images. To address this shortcoming, we propose Diverse Beam Search (DBS), a diversity promoting alternative to BS for approximate inference. DBS produces sequences that are significantly different from each other by incorporating diversity constraints within groups of candidate sequences during decoding; moreover, it achieves this with minimal computational or memory overhead. We demonstrate that our method improves both diversity and quality of decoded sequences over existing techniques on two visually-grounded language generation tasks---image captioning and visual question generation---particularly on complex scenes containing diverse visual content. We also show similar improvements at language-only machine translation tasks, highlighting the generality of our approach.

We are interested in counting the number of instances of object classes in natural, everyday images. Previous counting approaches tackle the problem in restricted domains such as counting pedestrians in surveillance videos. Counts can also be estimated from outputs of other vision tasks like object detection. In this work, we build dedicated models for counting designed to tackle the large variance in counts, appearances, and scales of objects found in natural scenes. Our approach is inspired by the phenomenon of subitizing – the ability of humans to make quick assessments of counts given a perceptual signal, for small count values. Given a natural scene, we employ a divide and conquer strategy while incorporating context across the scene to adapt the subitizing idea to counting. Our approach offers consistent improvements over numerous baseline approaches for counting on the PASCAL VOC 2007 and COCO datasets. Subsequently, we study how counting can be used to improve object detection. We then show a proof of concept application of our counting methods to the task of Visual Question Answering, by studying the how many? questions in the VQA and COCO-QA datasets.

We propose a technique for producing `visual explanations' for decisions from a large class of Convolutional Neural Network (CNN)-based models, making them more transparent. Our approach - Gradient-weighted Class Activation Mapping (Grad-CAM), uses the gradients of any target concept (say logits for `dog' or even a caption), flowing into the final convolutional layer to produce a coarse localization map highlighting the important regions in the image for predicting the concept. Unlike previous approaches, Grad- CAM is applicable to a wide variety of CNN model-families: (1) CNNs with fully-connected layers (e.g. VGG), (2) CNNs used for structured outputs (e.g. captioning), (3) CNNs used in tasks with multi-modal inputs (e.g. visual question answering) or reinforcement learning, without architectural changes or re-training. We combine Grad-CAM with existing fine-grained visualizations to create a high-resolution class-discriminative visualization, Guided Grad-CAM, and apply it to image classification, image captioning, and visual question answering (VQA) models, including ResNet-based architectures. In the context of image classification models, our visualizations (a) lend insights into failure modes of these models (showing that seemingly unreasonable predictions have reasonable explanations), (b) outperform previous methods on the ILSVRC-15 weakly-supervised localization task, (c) are more faithful to the underlying model, and (d) help achieve model generalization by identifying dataset bias. For image captioning and VQA, our visualizations show even non-attention based models can localize inputs. Finally, we design and conduct human studies to measure if Grad-CAM explanations help users establish appropriate trust in predictions from deep networks and show that Grad-CAM helps untrained users successfully discern a `stronger' deep network from a `weaker' one even when both make identical predictions. Our code is available at https: //github.com/ramprs/grad-cam/ along with a demo on CloudCV [2] and video at youtu.be/COjUB9Izk6E.