Cross-Attention Head Position Patterns Can Align with Human Visual Concepts in Text-to-Image Generative Models

Recent text-to-image diffusion models leverage cross-attention layers, which have been effectively utilized to enhance a range of visual generative tasks. However, our understanding of cross-attention layers remains somewhat limited. In this study, we present a method for constructing Head Relevance Vectors (HRVs) that align with useful visual concepts. An HRV for a given visual concept is a vector with a length equal to the total number of cross-attention heads, where each element represents the importance of the corresponding head for the given visual concept. We develop and employ an ordered weakening analysis to demonstrate the effectiveness of HRVs as interpretable features. To demonstrate the utility of HRVs, we propose concept strengthening and concept adjusting methods and apply them to enhance three visual generative tasks. We show that misinterpretations of polysemous words in image generation can be corrected in most cases, five challenging attributes in image editing can be successfully modified, and catastrophic neglect in multi-concept generation can be mitigated. Overall, our work provides an advancement in understanding cross-attention layers and introduces new approaches for fine-controlling these layers at the head level.

A Benchmark Suite for Evaluating Neural Mutual Information Estimators on Unstructured Datasets

Mutual Information (MI) is a fundamental metric for quantifying dependency between two random variables. When we can access only the samples, but not the underlying distribution functions, we can evaluate MI using sample-based estimators. Assessment of such MI estimators, however, has almost always relied on analytical datasets including Gaussian multivariates. Such datasets allow analytical calculations of the true MI values, but they are limited in that they do not reflect the complexities of real-world datasets. This study introduces a comprehensive benchmark suite for evaluating neural MI estimators on unstructured datasets, specifically focusing on images and texts. By leveraging same-class sampling for positive pairing and introducing a binary symmetric channel trick, we show that we can accurately manipulate true MI values of real-world datasets. Using the benchmark suite, we investigate seven challenging scenarios, shedding light on the reliability of neural MI estimators for unstructured datasets.

Towards a Better Evaluation of Out-of-Domain Generalization

The objective of Domain Generalization (DG) is to devise algorithms and models capable of achieving high performance on previously unseen test distributions. In the pursuit of this objective, average measure has been employed as the prevalent measure for evaluating models and comparing algorithms in the existing DG studies. Despite its significance, a comprehensive exploration of the average measure has been lacking and its suitability in approximating the true domain generalization performance has been questionable. In this study, we carefully investigate the limitations inherent in the average measure and propose worst+gap measure as a robust alternative. We establish theoretical grounds of the proposed measure by deriving two theorems starting from two different assumptions. We conduct extensive experimental investigations to compare the proposed worst+gap measure with the conventional average measure. Given the indispensable need to access the true DG performance for studying measures, we modify five existing datasets to come up with SR-CMNIST, C-Cats&Dogs, L-CIFAR10, PACS-corrupted, and VLCS-corrupted datasets. The experiment results unveil an inferior performance of the average measure in approximating the true DG performance and confirm the robustness of the theoretically supported worst+gap measure.

Unveiling Key Aspects of Fine-Tuning in Sentence Embeddings: A Representation Rank Analysis

The latest advancements in unsupervised learning of sentence embeddings predominantly involve employing contrastive learning-based (CL-based) fine-tuning over pre-trained language models. In this study, we analyze the latest sentence embedding methods by adopting representation rank as the primary tool of analysis. We first define Phase 1 and Phase 2 of fine-tuning based on when representation rank peaks. Utilizing these phases, we conduct a thorough analysis and obtain essential findings across key aspects, including alignment and uniformity, linguistic abilities, and correlation between performance and rank. For instance, we find that the dynamics of the key aspects can undergo significant changes as fine-tuning transitions from Phase 1 to Phase 2. Based on these findings, we experiment with a rank reduction (RR) strategy that facilitates rapid and stable fine-tuning of the latest CL-based methods. Through empirical investigations, we showcase the efficacy of RR in enhancing the performance and stability of five state-of-the-art sentence embedding methods.

An Image Grid Can Be Worth a Video: Zero-shot Video Question Answering Using a VLM

Stimulated by the sophisticated reasoning capabilities of recent Large Language Models (LLMs), a variety of strategies for bridging video modality have been devised. A prominent strategy involves Video Language Models (VideoLMs), which train a learnable interface with video data to connect advanced vision encoders with LLMs. Recently, an alternative strategy has surfaced, employing readily available foundation models, such as VideoLMs and LLMs, across multiple stages for modality bridging. In this study, we introduce a simple yet novel strategy where only a single Vision Language Model (VLM) is utilized. Our starting point is the plain insight that a video comprises a series of images, or frames, interwoven with temporal information. The essence of video comprehension lies in adeptly managing the temporal aspects along with the spatial details of each frame. Initially, we transform a video into a single composite image by arranging multiple frames in a grid layout. The resulting single image is termed as an image grid. This format, while maintaining the appearance of a solitary image, effectively retains temporal information within the grid structure. Therefore, the image grid approach enables direct application of a single high-performance VLM without necessitating any video-data training. Our extensive experimental analysis across ten zero-shot video question answering benchmarks, including five open-ended and five multiple-choice benchmarks, reveals that the proposed Image Grid Vision Language Model (IG-VLM) surpasses the existing methods in nine out of ten benchmarks.

Selectively Informative Description can Reduce Undesired Embedding Entanglements in Text-to-Image Personalization

In text-to-image personalization, a timely and crucial challenge is the tendency of generated images overfitting to the biases present in the reference images. We initiate our study with a comprehensive categorization of the biases into background, nearby-object, tied-object, substance (in style re-contextualization), and pose biases. These biases manifest in the generated images due to their entanglement into the subject embedding. This undesired embedding entanglement not only results in the reflection of biases from the reference images into the generated images but also notably diminishes the alignment of the generated images with the given generation prompt. To address this challenge, we propose SID~(Selectively Informative Description), a text description strategy that deviates from the prevalent approach of only characterizing the subject's class identification. SID is generated utilizing multimodal GPT-4 and can be seamlessly integrated into optimization-based models. We present comprehensive experimental results along with analyses of cross-attention maps, subject-alignment, non-subject-disentanglement, and text-alignment.

Improving Forward Compatibility in Class Incremental Learning by Increasing Representation Rank and Feature Richness

Class Incremental Learning (CIL) constitutes a pivotal subfield within continual learning, aimed at enabling models to progressively learn new classification tasks while retaining knowledge obtained from prior tasks. Although previous studies have predominantly focused on backward compatible approaches to mitigate catastrophic forgetting, recent investigations have introduced forward compatible methods to enhance performance on novel tasks and complement existing backward compatible methods. In this study, we introduce an effective-Rank based Feature Richness enhancement (RFR) method, designed for improving forward compatibility. Specifically, this method increases the effective rank of representations during the base session, thereby facilitating the incorporation of more informative features pertinent to unseen novel tasks. Consequently, RFR achieves dual objectives in backward and forward compatibility: minimizing feature extractor modifications and enhancing novel task performance, respectively. To validate the efficacy of our approach, we establish a theoretical connection between effective rank and the Shannon entropy of representations. Subsequently, we conduct comprehensive experiments by integrating RFR into eleven well-known CIL methods. Our results demonstrate the effectiveness of our approach in enhancing novel-task performance while mitigating catastrophic forgetting. Furthermore, our method notably improves the average incremental accuracy across all eleven cases examined.

Harmonizing Visual and Textual Embeddings for Zero-Shot Text-to-Image Customization

In a surge of text-to-image (T2I) models and their customization methods that generate new images of a user-provided subject, current works focus on alleviating the costs incurred by a lengthy per-subject optimization. These zero-shot customization methods encode the image of a specified subject into a visual embedding which is then utilized alongside the textual embedding for diffusion guidance. The visual embedding incorporates intrinsic information about the subject, while the textual embedding provides a new, transient context. However, the existing methods often 1) are significantly affected by the input images, eg., generating images with the same pose, and 2) exhibit deterioration in the subject's identity. We first pin down the problem and show that redundant pose information in the visual embedding interferes with the textual embedding containing the desired pose information. To address this issue, we propose orthogonal visual embedding which effectively harmonizes with the given textual embedding. We also adopt the visual-only embedding and inject the subject's clear features utilizing a self-attention swap. Our results demonstrate the effectiveness and robustness of our method, which offers highly flexible zero-shot generation while effectively maintaining the subject's identity.

On-Off Pattern Encoding and Path-Count Encoding as Deep Neural Network Representations

Understanding the encoded representation of Deep Neural Networks (DNNs) has been a fundamental yet challenging objective. In this work, we focus on two possible directions for analyzing representations of DNNs by studying simple image classification tasks. Specifically, we consider \textit{On-Off pattern} and \textit{PathCount} for investigating how information is stored in deep representations. On-off pattern of a neuron is decided as `on' or `off' depending on whether the neuron's activation after ReLU is non-zero or zero. PathCount is the number of paths that transmit non-zero energy from the input to a neuron. We investigate how neurons in the network encodes information by replacing each layer's activation with On-Off pattern or PathCount and evaluating its effect on classification performance. We also examine correlation between representation and PathCount. Finally, we show a possible way to improve an existing DNN interpretation method, Class Activation Map (CAM), by directly utilizing On-Off or PathCount.

Enhancing Contrastive Learning with Efficient Combinatorial Positive Pairing

In the past few years, contrastive learning has played a central role for the success of visual unsupervised representation learning. Around the same time, high-performance non-contrastive learning methods have been developed as well. While most of the works utilize only two views, we carefully review the existing multi-view methods and propose a general multi-view strategy that can improve learning speed and performance of any contrastive or non-contrastive method. We first analyze CMC's full-graph paradigm and empirically show that the learning speed of $K$-views can be increased by $_{K}\mathrm{C}_{2}$ times for small learning rate and early training. Then, we upgrade CMC's full-graph by mixing views created by a crop-only augmentation, adopting small-size views as in SwAV multi-crop, and modifying the negative sampling. The resulting multi-view strategy is called ECPP (Efficient Combinatorial Positive Pairing). We investigate the effectiveness of ECPP by applying it to SimCLR and assessing the linear evaluation performance for CIFAR-10 and ImageNet-100. For each benchmark, we achieve a state-of-the-art performance. In case of ImageNet-100, ECPP boosted SimCLR outperforms supervised learning.