Transfer learning involves adapting a pre-trained model to novel downstream tasks. However, we observe that current transfer learning methods often fail to focus on task-relevant features. In this work, we emphasize the importance of refocusing the attention in transfer learning. We introduce Top-Down Attention Steering (TOAST), a novel transfer learning algorithm that keeps the pre-trained backbone frozen, while selecting the task-relevant elements in the output and feeding them back to the model to steer its attention to the task-specific features. By refocusing the attention only, TOAST achieves state-of-the-art results on a number of transfer learning benchmarks, while having a small portion of tunable parameters. Compared to fully fine-tuning, LoRA, and prompt tuning, TOAST substantially improves performance across a range of fine-grained visual classification datasets (e.g., 81.1% -> 86.2% on FGVC). TOAST also outperforms the fully fine-tuned Alpaca model on instruction-following language generation. Code is available at https://github.com/bfshi/TOAST.
The explosive growth of language models and their applications have led to an increased demand for efficient and scalable methods. In this paper, we introduce Flan-MoE, a set of Instruction-Finetuned Sparse Mixture-of-Expert (MoE) models. We show that naively finetuning MoE models on a task-specific dataset (in other words, no instruction-finetuning) often yield worse performance compared to dense models of the same computational complexity. However, our Flan-MoE outperforms dense models under multiple experiment settings: instruction-finetuning only and instruction-finetuning followed by task-specific finetuning. This shows that instruction-finetuning is an essential stage for MoE models. Specifically, our largest model, Flan-MoE-32B, surpasses the performance of Flan-PaLM-62B on four benchmarks, while utilizing only one-third of the FLOPs. The success of Flan-MoE encourages rethinking the design of large-scale, high-performance language models, under the setting of task-agnostic learning.
Diffusion models have been shown to be capable of generating high-quality images, suggesting that they could contain meaningful internal representations. Unfortunately, the feature maps that encode a diffusion model's internal information are spread not only over layers of the network, but also over diffusion timesteps, making it challenging to extract useful descriptors. We propose Diffusion Hyperfeatures, a framework for consolidating multi-scale and multi-timestep feature maps into per-pixel feature descriptors that can be used for downstream tasks. These descriptors can be extracted for both synthetic and real images using the generation and inversion processes. We evaluate the utility of our Diffusion Hyperfeatures on the task of semantic keypoint correspondence: our method achieves superior performance on the SPair-71k real image benchmark. We also demonstrate that our method is flexible and transferable: our feature aggregation network trained on the inversion features of real image pairs can be used on the generation features of synthetic image pairs with unseen objects and compositions. Our code is available at \url{https://diffusion-hyperfeatures.github.io}.
Recent advancements in text-to-image generation with diffusion models have yielded remarkable results synthesizing highly realistic and diverse images. However, these models still encounter difficulties when generating images from prompts that demand spatial or common sense reasoning. We propose to equip diffusion models with enhanced reasoning capabilities by using off-the-shelf pretrained large language models (LLMs) in a novel two-stage generation process. First, we adapt an LLM to be a text-guided layout generator through in-context learning. When provided with an image prompt, an LLM outputs a scene layout in the form of bounding boxes along with corresponding individual descriptions. Second, we steer a diffusion model with a novel controller to generate images conditioned on the layout. Both stages utilize frozen pretrained models without any LLM or diffusion model parameter optimization. We validate the superiority of our design by demonstrating its ability to outperform the base diffusion model in accurately generating images according to prompts that necessitate both language and spatial reasoning. Additionally, our method naturally allows dialog-based scene specification and is able to handle prompts in a language that is not well-supported by the underlying diffusion model.
The ability to judge whether a caption correctly describes an image is a critical part of vision-language understanding. However, state-of-the-art models often misinterpret the correctness of fine-grained details, leading to errors in outputs such as hallucinating objects in generated captions or poor compositional reasoning. In this work, we explore Token-Level Confidence, or TLC, as a simple yet surprisingly effective method to assess caption correctness. Specifically, we fine-tune a vision-language model on image captioning, input an image and proposed caption to the model, and aggregate either algebraic or learned token confidences over words or sequences to estimate image-caption consistency. Compared to sequence-level scores from pretrained models, TLC with algebraic confidence measures achieves a relative improvement in accuracy by 10% on verb understanding in SVO-Probes and outperforms prior state-of-the-art in image and group scores for compositional reasoning in Winoground by a relative 37% and 9%, respectively. When training data are available, a learned confidence estimator provides further improved performance, reducing object hallucination rates in MS COCO Captions by a relative 30% over the original model and setting a new state-of-the-art.
Vision and Language (VL) models have demonstrated remarkable zero-shot performance in a variety of tasks. However, recent studies have shown that even the best VL models struggle to capture aspects of scene understanding, such as object attributes, relationships, and action states. In contrast, obtaining structured annotations, e.g., scene graphs (SGs) that could improve these models is time-consuming, costly, and tedious, and thus cannot be used on a large scale. Here we ask, can small datasets containing SG annotations provide sufficient information for enhancing structured understanding of VL models? We show that it is indeed possible to improve VL models using such data by utilizing a specialized model architecture and a new training paradigm. Our approach captures structure-related information for both the visual and textual encoders by directly supervising both components when learning from SG labels. We use scene graph supervision to generate fine-grained captions based on various graph augmentations highlighting different compositional aspects of the scene, and to predict SG information using an open vocabulary approach by adding special ``Adaptive SG tokens'' to the visual encoder. Moreover, we design a new adaptation technique tailored specifically to the SG tokens that allows better learning of the graph prediction task while still maintaining zero-shot capabilities. Our model shows strong performance improvements on the Winoground and VL-checklist datasets with only a mild degradation in zero-shot performance.
Image editing using diffusion models has witnessed extremely fast-paced growth recently. There are various ways in which previous works enable controlling and editing images. Some works use high-level conditioning such as text, while others use low-level conditioning. Nevertheless, most of them lack fine-grained control over the properties of the different objects present in the image, i.e. object-level image editing. In this work, we consider an image as a composition of multiple objects, each defined by various properties. Out of these properties, we identify structure and appearance as the most intuitive to understand and useful for editing purposes. We propose Structure-and-Appearance Paired Diffusion model (PAIR-Diffusion), which is trained using structure and appearance information explicitly extracted from the images. The proposed model enables users to inject a reference image's appearance into the input image at both the object and global levels. Additionally, PAIR-Diffusion allows editing the structure while maintaining the style of individual components of the image unchanged. We extensively evaluate our method on LSUN datasets and the CelebA-HQ face dataset, and we demonstrate fine-grained control over both structure and appearance at the object level. We also applied the method to Stable Diffusion to edit any real image at the object level.
Current attention algorithms (e.g., self-attention) are stimulus-driven and highlight all the salient objects in an image. However, intelligent agents like humans often guide their attention based on the high-level task at hand, focusing only on task-related objects. This ability of task-guided top-down attention provides task-adaptive representation and helps the model generalize to various tasks. In this paper, we consider top-down attention from a classic Analysis-by-Synthesis (AbS) perspective of vision. Prior work indicates a functional equivalence between visual attention and sparse reconstruction; we show that an AbS visual system that optimizes a similar sparse reconstruction objective modulated by a goal-directed top-down signal naturally simulates top-down attention. We further propose Analysis-by-Synthesis Vision Transformer (AbSViT), which is a top-down modulated ViT model that variationally approximates AbS, and achieves controllable top-down attention. For real-world applications, AbSViT consistently improves over baselines on Vision-Language tasks such as VQA and zero-shot retrieval where language guides the top-down attention. AbSViT can also serve as a general backbone, improving performance on classification, semantic segmentation, and model robustness.
Given the enormous number of instructional videos available online, learning a diverse array of multi-step task models from videos is an appealing goal. We introduce a new pre-trained video model, VideoTaskformer, focused on representing the semantics and structure of instructional videos. We pre-train VideoTaskformer using a simple and effective objective: predicting weakly supervised textual labels for steps that are randomly masked out from an instructional video (masked step modeling). Compared to prior work which learns step representations locally, our approach involves learning them globally, leveraging video of the entire surrounding task as context. From these learned representations, we can verify if an unseen video correctly executes a given task, as well as forecast which steps are likely to be taken after a given step. We introduce two new benchmarks for detecting mistakes in instructional videos, to verify if there is an anomalous step and if steps are executed in the right order. We also introduce a long-term forecasting benchmark, where the goal is to predict long-range future steps from a given step. Our method outperforms previous baselines on these tasks, and we believe the tasks will be a valuable way for the community to measure the quality of step representations. Additionally, we evaluate VideoTaskformer on 3 existing benchmarks -- procedural activity recognition, step classification, and step forecasting -- and demonstrate on each that our method outperforms existing baselines and achieves new state-of-the-art performance.
The field of natural language processing (NLP) has made significant strides in recent years, particularly in the development of large-scale vision-language models (VLMs). These models aim to bridge the gap between text and visual information, enabling a more comprehensive understanding of multimedia data. However, as these models become larger and more complex, they also become more challenging to train and deploy. One approach to addressing this challenge is the use of sparsely-gated mixture-of-experts (MoE) techniques, which divide the model into smaller, specialized sub-models that can jointly solve a task. In this paper, we explore the effectiveness of MoE in scaling vision-language models, demonstrating its potential to achieve state-of-the-art performance on a range of benchmarks over dense models of equivalent computational cost. Our research offers valuable insights into stabilizing the training of MoE models, understanding the impact of MoE on model interpretability, and balancing the trade-offs between compute performance when scaling VLMs. We hope our work will inspire further research into the use of MoE for scaling large-scale vision-language models and other multimodal machine learning applications.