Towards Multimodal In-Context Learning for Vision & Language Models

Inspired by the emergence of Large Language Models (LLMs) that can truly understand human language, significant progress has been made in aligning other, non-language, modalities to be `understandable' by an LLM, primarily via converting their samples into a sequence of embedded language-like tokens directly fed into the LLM (decoder) input stream. However, so far limited attention has been given to transferring (and evaluating) one of the core LLM capabilities to the emerging VLMs, namely the In-Context Learning (ICL) ability, or in other words to guide VLMs to desired target downstream tasks or output structure using in-context image+text demonstrations. In this work, we dive deeper into analyzing the capabilities of some of the state-of-the-art VLMs to follow ICL instructions, discovering them to be somewhat lacking. We discover that even models that underwent large-scale mixed modality pre-training and were implicitly guided to make use of interleaved image and text information (intended to consume helpful context from multiple images) under-perform when prompted with few-shot (ICL) demonstrations, likely due to their lack of `direct' ICL instruction tuning. To test this conjecture, we propose a simple, yet surprisingly effective, strategy of extending a common VLM alignment framework with ICL support, methodology, and curriculum. We explore, analyze, and provide insights into effective data mixes, leading up to a significant 21.03% (and 11.3% on average) ICL performance boost over the strongest VLM baselines and a variety of ICL benchmarks. We also contribute new benchmarks for ICL evaluation in VLMs and discuss their advantages over the prior art.

Efficient Rehearsal Free Zero Forgetting Continual Learning using Adaptive Weight Modulation

Artificial neural networks encounter a notable challenge known as continual learning, which involves acquiring knowledge of multiple tasks over an extended period. This challenge arises due to the tendency of previously learned weights to be adjusted to suit the objectives of new tasks, resulting in a phenomenon called catastrophic forgetting. Most approaches to this problem seek a balance between maximizing performance on the new tasks and minimizing the forgetting of previous tasks. In contrast, our approach attempts to maximize the performance of the new task, while ensuring zero forgetting. This is accomplished by creating a task-specific modulation parameters for each task. Only these would be learnable parameters during learning of consecutive tasks. Through comprehensive experimental evaluations, our model demonstrates superior performance in acquiring and retaining novel tasks that pose difficulties for other multi-task models. This emphasizes the efficacy of our approach in preventing catastrophic forgetting while accommodating the acquisition of new tasks

Top-Down Processing: Top-Down Network Combines Back-Propagation with Attention

Early neural network models relied exclusively on bottom-up processing going from the input signals to higher-level representations. Many recent models also incorporate top-down networks going in the opposite direction. Top-down processing in deep learning models plays two primary roles: learning and directing attention. These two roles are accomplished in current models through distinct mechanisms. While top-down attention is often implemented by extending the model's architecture with additional units that propagate information from high to low levels of the network, learning is typically accomplished by an external learning algorithm such as back-propagation. In the current work, we present an integration of the two functions above, which appear unrelated, using a single unified mechanism. We propose a novel symmetric bottom-up top-down network structure that can integrate standard bottom-up networks with a symmetric top-down counterpart, allowing each network to guide and influence the other. The same top-down network is being used for both learning, via back-propagating feedback signals, and at the same time also for top-down attention, by guiding the bottom-up network to perform a selected task. We show that our method achieves competitive performance on a standard multi-task learning benchmark. Yet, we rely on standard single-task architectures and optimizers, without any task-specific parameters. Additionally, our learning algorithm addresses in a new way some neuroscience issues that arise in biological modeling of learning in the brain.

Dense and Aligned Captions (DAC) Promote Compositional Reasoning in VL Models

Vision and Language (VL) models offer an effective method for aligning representation spaces of images and text, leading to numerous applications such as cross-modal retrieval, visual question answering, captioning, and more. However, the aligned image-text spaces learned by all the popular VL models are still suffering from the so-called `object bias' - their representations behave as `bags of nouns', mostly ignoring or downsizing the attributes, relations, and states of objects described/appearing in texts/images. Although some great attempts at fixing these `compositional reasoning' issues were proposed in the recent literature, the problem is still far from being solved. In this paper, we uncover two factors limiting the VL models' compositional reasoning performance. These two factors are properties of the paired VL dataset used for finetuning and pre-training the VL model: (i) the caption quality, or in other words `image-alignment', of the texts; and (ii) the `density' of the captions in the sense of mentioning all the details appearing on the image. We propose a fine-tuning approach for automatically treating these factors leveraging a standard VL dataset (CC3M). Applied to CLIP, we demonstrate its significant compositional reasoning performance increase of up to $\sim27\%$ over the base model, up to $\sim20\%$ over the strongest baseline, and by $6.7\%$ on average.

Teaching Structured Vision&Language Concepts to Vision&Language Models

Vision and Language (VL) models have demonstrated remarkable zero-shot performance in a variety of tasks. However, some aspects of complex language understanding still remain a challenge. We introduce the collective notion of Structured Vision&Language Concepts (SVLC) which includes object attributes, relations, and states which are present in the text and visible in the image. Recent studies have shown that even the best VL models struggle with SVLC. A possible way of fixing this issue is by collecting dedicated datasets for teaching each SVLC type, yet this might be expensive and time-consuming. Instead, we propose a more elegant data-driven approach for enhancing VL models' understanding of SVLCs that makes more effective use of existing VL pre-training datasets and does not require any additional data. While automatic understanding of image structure still remains largely unsolved, language structure is much better modeled and understood, allowing for its effective utilization in teaching VL models. In this paper, we propose various techniques based on language structure understanding that can be used to manipulate the textual part of off-the-shelf paired VL datasets. VL models trained with the updated data exhibit a significant improvement of up to 15% in their SVLC understanding with only a mild degradation in their zero-shot capabilities both when training from scratch or fine-tuning a pre-trained model.

A model for full local image interpretation

We describe a computational model of humans' ability to provide a detailed interpretation of components in a scene. Humans can identify in an image meaningful components almost everywhere, and identifying these components is an essential part of the visual process, and of understanding the surrounding scene and its potential meaning to the viewer. Detailed interpretation is beyond the scope of current models of visual recognition. Our model suggests that this is a fundamental limitation, related to the fact that existing models rely on feed-forward but limited top-down processing. In our model, a first recognition stage leads to the initial activation of class candidates, which is incomplete and with limited accuracy. This stage then triggers the application of class-specific interpretation and validation processes, which recover richer and more accurate interpretation of the visible scene. We discuss implications of the model for visual interpretation by humans and by computer vision models.

Image interpretation by iterative bottom-up top-down processing

Scene understanding requires the extraction and representation of scene components together with their properties and inter-relations. We describe a model in which meaningful scene structures are extracted from the image by an iterative process, combining bottom-up (BU) and top-down (TD) networks, interacting through a symmetric bi-directional communication between them (counter-streams structure). The model constructs a scene representation by the iterative use of three components. The first model component is a BU stream that extracts selected scene elements, properties and relations. The second component (cognitive augmentation) augments the extracted visual representation based on relevant non-visual stored representations. It also provides input to the third component, the TD stream, in the form of a TD instruction, instructing the model what task to perform next. The TD stream then guides the BU visual stream to perform the selected task in the next cycle. During this process, the visual representations extracted from the image can be combined with relevant non-visual representations, so that the final scene representation is based on both visual information extracted from the scene and relevant stored knowledge of the world. We describe how a sequence of TD-instructions is used to extract from the scene structures of interest, including an algorithm to automatically select the next TD-instruction in the sequence. The extraction process is shown to have favorable properties in terms of combinatorial generalization, generalizing well to novel scene structures and new combinations of objects, properties and relations not seen during training. Finally, we compare the model with relevant aspects of the human vision, and suggest directions for using the BU-TD scheme for integrating visual and cognitive components in the process of scene understanding.

Detector-Free Weakly Supervised Grounding by Separation

Nowadays, there is an abundance of data involving images and surrounding free-form text weakly corresponding to those images. Weakly Supervised phrase-Grounding (WSG) deals with the task of using this data to learn to localize (or to ground) arbitrary text phrases in images without any additional annotations. However, most recent SotA methods for WSG assume the existence of a pre-trained object detector, relying on it to produce the ROIs for localization. In this work, we focus on the task of Detector-Free WSG (DF-WSG) to solve WSG without relying on a pre-trained detector. We directly learn everything from the images and associated free-form text pairs, thus potentially gaining an advantage on the categories unsupported by the detector. The key idea behind our proposed Grounding by Separation (GbS) method is synthesizing `text to image-regions' associations by random alpha-blending of arbitrary image pairs and using the corresponding texts of the pair as conditions to recover the alpha map from the blended image via a segmentation network. At test time, this allows using the query phrase as a condition for a non-blended query image, thus interpreting the test image as a composition of a region corresponding to the phrase and the complement region. Using this approach we demonstrate a significant accuracy improvement, of up to $8.5\%$ over previous DF-WSG SotA, for a range of benchmarks including Flickr30K, Visual Genome, and ReferIt, as well as a significant complementary improvement (above $7\%$) over the detector-based approaches for WSG.

What can human minimal videos tell us about dynamic recognition models?

In human vision objects and their parts can be visually recognized from purely spatial or purely temporal information but the mechanisms integrating space and time are poorly understood. Here we show that human visual recognition of objects and actions can be achieved by efficiently combining spatial and motion cues in configurations where each source on its own is insufficient for recognition. This analysis is obtained by identifying minimal videos: these are short and tiny video clips in which objects, parts, and actions can be reliably recognized, but any reduction in either space or time makes them unrecognizable. State-of-the-art deep networks for dynamic visual recognition cannot replicate human behavior in these configurations. This gap between humans and machines points to critical mechanisms in human dynamic vision that are lacking in current models.

What takes the brain so long: Object recognition at the level of minimal images develops for up to seconds of presentation time

Rich empirical evidence has shown that visual object recognition in the brain is fast and effortless, with relevant brain signals reported to start as early as 80 ms. Here we study the time trajectory of the recognition process at the level of minimal recognizable images (termed MIRC). These are images that can be recognized reliably, but in which a minute change of the image (reduction by either size or resolution) has a drastic effect on recognition. Subjects were assigned to one of nine exposure conditions: 200, 500, 1000, 2000 ms with or without masking, as well as unlimited time. The subjects were not limited in time to respond after presentation. The results show that in the masked conditions, recognition rates develop gradually over an extended period, e.g. average of 18% for 200 ms exposure and 45% for 500 ms, increasing significantly with longer exposure even above 2 secs. When presented for unlimited time (until response), MIRC recognition rates were equivalent to the rates of full-object images presented for 50 ms followed by masking. What takes the brain so long to recognize such images? We discuss why processes involving eye-movements, perceptual decision-making and pattern completion are unlikely explanations. Alternatively, we hypothesize that MIRC recognition requires an extended top-down process complementing the feed-forward phase.