ICLR 2024 Papers

With the ever-increasing popularity of pretrained Video-Language Models (VidLMs), there is a pressing need to develop robust evaluation methodologies that delve deeper into their visio-linguistic capabilities. To address this challenge, we present ViLMA (Video Language Model Assessment), a task-agnostic benchmark that places the assessment of fine-grained capabilities of these models on a firm footing. Task-based evaluations, while valuable, fail to capture the complexities and specific temporal aspects of moving images that VidLMs need to process. Through carefully curated counterfactuals, ViLMA offers a controlled evaluation suite that sheds light on the true potential of these models, as well as their performance gaps compared to human-level understanding. ViLMA also includes proficiency tests, which assess basic capabilities deemed essential to solving the main counterfactual tests. We show that current VidLMs’ grounding abilities are no better than those of vision-language models which use static images. This is especially striking once the performance on proficiency tests is factored in. Our benchmark serves as a catalyst for future research on VidLMs, helping to highlight areas that still need to be explored.

This paper addresses the problem of anomaly detection in tabular data, which is usually implemented in an one-class classification setting where the training set only contains normal samples. Inspired by the success of masked image/language modeling in vision and natural language domains, we extend masked modeling methods to address this problem by capturing intrinsic correlations between features in training set. Thus, a sample deviate from such correlations is related to a high possibility of anomaly. To obtain multiple and diverse correlations, we propose a novel masking strategy which generates multiple masks by learning, and design a diversity loss to reduce the similarity of different masks. Extensive experiments show our method achieves state-of-the-art performance. We also discuss the interpretability from the perspective of each individual feature and correlations between features.

We introduce Kosmos-2, a Multimodal Large Language Model (MLLM), enabling new capabilities of perceiving object descriptions (e.g., bounding boxes) and grounding text to the visual world. Specifically, we represent text spans (i.e., referring expressions and noun phrases) as links in Markdown, i.e., [text span](bounding boxes), where object descriptions are sequences of location tokens. To train the model, we construct a large-scale dataset about grounded image-text pairs (GrIT) together with multimodal corpora. In addition to the existing capabilities of MLLMs (e.g., perceiving general modalities, following instructions, and performing in-context learning), Kosmos-2 integrates the grounding capability to downstream applications, while maintaining the conventional capabilities of MLLMs (e.g., perceiving general modalities, following instructions, and performing in-context learning). Kosmos-2 is evaluated on a wide range of tasks, including (i) multimodal grounding, such as referring expression comprehension and phrase grounding, (ii) multimodal referring, such as referring expression generation, (iii) perception-language tasks, and (iv) language understanding and generation. This study sheds a light on the big convergence of language, multimodal perception, and world modeling, which is a key step toward artificial general intelligence. Code can be found in [https://aka.ms/kosmos-2](https://aka.ms/kosmos-2).

This paper introduces VLAP, a novel approach that bridges pretrained vision models and large language models (LLMs) to make frozen LLMs understand the visual world. VLAP transforms the embedding space of pretrained vision models into the LLMs' word embedding space using a single linear layer for efficient and general-purpose visual and language understanding. Specifically, we harness well-established word embeddings to bridge two modality embedding spaces. The visual and text representations are simultaneously assigned to a set of word embeddings within pretrained LLMs by formulating the assigning procedure as an optimal transport problem. We predict the assignment of one modality from the representation of another modality data, enforcing consistent assignments for paired multimodal data. This allows vision and language representations to contain the same information, grounding the frozen LLMs' word embedding space in visual data. Moreover, a robust semantic taxonomy of LLMs can be preserved with visual data since the LLMs interpret and reason linguistic information from correlations between word embeddings. Experimental results show that VLAP achieves substantial improvements over the previous linear transformation-based approaches across a range of vision-language tasks, including image captioning, visual question answering, and cross-modal retrieval. We also demonstrate the learned visual representations hold a semantic taxonomy of LLMs, making visual semantic arithmetic possible.

Binary Neural networks (BNN) have emerged as an attractive computing paradigm for a wide range of low-power vision tasks. However, state-of-the-art (SOTA) BNNs do not yield any sparsity, and induce a significant number of non-binary operations. On the other hand, activation sparsity can be provided by spiking neural networks (SNN), that too have gained significant traction in recent times. Thanks to this sparsity, SNNs when implemented on neuromorphic hardware, have the potential to be significantly more power-efficient compared to traditional artifical neural networks (ANN). However, SNNs incur multiple time steps to achieve close to SOTA accuracy. Ironically, this increases latency and energy---costs that SNNs were proposed to reduce---and presents itself as a major hurdle in realizing SNNs’ theoretical gains in practice. This raises an intriguing question: *Can we obtain SNN-like sparsity and BNN-like accuracy and enjoy the energy-efficiency benefits of both?* To answer this question, in this paper, we present a training framework for sparse binary activation neural networks (BANN) using a novel variant of the Hoyer regularizer. We estimate the threshold of each BANN layer as the Hoyer extremum of a clipped version of its activation map, where the clipping value is trained using gradient descent with our Hoyer regularizer. This approach shifts the activation values away from the threshold, thereby mitigating the effect of noise that can otherwise degrade the BANN accuracy. Our approach outperforms existing BNNs, SNNs, and adder neural networks (that also avoid energy-expensive multiplication operations similar to BNNs and SNNs) in terms of the accuracy-FLOPs trade-off for complex image recognition tasks. Downstream experiments on object detection further demonstrate the efficacy of our approach. Lastly, we demonstrate the portability of our approach to SNNs with multiple time steps. Codes are publicly available [here](https://github.com/godatta/Ultra-Low-Latency-SNN).

Recent progress in vision language foundation models has shown their ability to understand multimodal data and resolve complicated vision language tasks, including robotics manipulation. We seek a straightforward way of making use of existing vision-language models (VLMs) with simple fine-tuning on robotics data. To this end, we derive a simple and novel vision-language manipulation framework, dubbed RoboFlamingo, built upon the open-source VLMs, OpenFlamingo. Unlike prior works, RoboFlamingo utilizes pre-trained VLMs for single-step vision-language comprehension, models sequential history information with an explicit policy head, and is slightly fine-tuned by imitation learning only on language-conditioned manipulation datasets. Such a decomposition provides RoboFlamingo the flexibility for open-loop control and deployment on low-performance platforms. By exceeding the state-of-the-art performance with a large margin on the tested benchmark, we show RoboFlamingo can be an effective and competitive alternative to adapt VLMs to robot control. Our extensive experimental results also reveal several interesting conclusions regarding the behavior of different pre-trained VLMs on manipulation tasks. We believe RoboFlamingo has the potential to be a cost-effective and easy-to-use solution for robotics manipulation, empowering everyone with the ability to fine-tune their own robotics policy. Our code will be made public upon acceptance.

Contrastive learning, while highly effective for a lot of tasks, shows limited improvement in ordinal regression. We find that the limitation comes from the predefined strong data augmentations employed in contrastive learning. Intuitively, for ordinal regression datasets, the discriminative information (ordinal content information) contained in instances is subtle. The strong augmentations can easily overshadow or diminish this ordinal content information. As a result, when contrastive learning is used to extract common features between weakly and strongly augmented images, the derived features often lack this essential ordinal content, rendering them less useful in training models for ordinal regression. To improve contrastive learning's utility for ordinal regression, we propose a novel augmentation method to replace the predefined strong argumentation based on the principle of minimal change. Our method is designed in a generative manner that can effectively generate images with different styles but contains desired ordinal content information. Extensive experiments validate the effectiveness of our proposed method, which serves as a plug-and-play solution and consistently improves the performance of existing state-of-the-art methods in ordinal regression tasks.

Developing a universal model that can effectively harness heterogeneous resources and respond to a wide range of personalized needs has been a longstanding community aspiration. Our daily choices, especially in domains like fashion and retail, are substantially shaped by multi-modal data, such as pictures and textual descriptions. These modalities not only offer intuitive guidance but also cater to personalized user preferences. However, the predominant personalization approaches mainly focus on ID or text-based recommendation problems, failing to comprehend the information spanning various tasks or modalities. In this paper, our goal is to establish a Unified paradigm for Multi-modal Personalization systems (UniMP), which effectively leverages multi-modal data while eliminating the complexities associated with task- and modality-specific customization. We argue that the advancements in foundational generative modeling have provided the flexibility and effectiveness necessary to achieve the objective. In light of this, we develop a generic and extensible personalization generative framework, that can handle a wide range of personalized needs including item recommendation, product search, preference prediction, explanation generation, and further user-guided image generation. Our methodology enhances the capabilities of foundational language models for personalized tasks by seamlessly ingesting interleaved cross-modal user history information, ensuring a more precise and customized experience for users. To train and evaluate the proposed multi-modal personalized tasks, we also introduce a novel and comprehensive benchmark covering a variety of user requirements. Our experiments on the real-world benchmark showcase the model's potential, outperforming competitive methods specialized for each task.

With the advancements in Large Language Models (LLMs), Vision-Language Models (VLMs) have reached a new level of sophistication, showing notable competence in executing intricate cognition and reasoning tasks. However, existing evaluation benchmarks, primarily relying on rigid, hand-crafted datasets to measure task-specific performance, face significant limitations in assessing the alignment of these increasingly anthropomorphic models with human intelligence. In this work, we address the limitations via Auto-Bench, which delves into exploring LLMs as proficient aligners, measuring the alignment between VLMs and human intelligence and value through automatic data curation and assessment. Specifically, for data curation, Auto-Bench utilizes LLMs (e.g., GPT-4) to automatically generate a vast set of question-answer-reasoning triplets via prompting on visual symbolic representations (e.g., captions, object locations, instance relationships, and etc. The curated data closely matches human intent, owing to the extensive world knowledge embedded in LLMs. Through this pipeline, a total of 28.5K human-verified and 3,504K unfiltered question-answer-reasoning triplets have been curated, covering 4 primary abilities and 16 sub-abilities. We subsequently engage LLMs like GPT-3.5 to serve as judges, implementing the quantitative and qualitative automated assessments to facilitate a comprehensive evaluation of VLMs. Our validation results reveal that LLMs are proficient in both evaluation data curation and model assessment, achieving an average agreement rate of 85%. We envision Auto-Bench as a flexible, scalable, and comprehensive benchmark for evaluating the evolving sophisticated VLMs.

Current vision-language generative models rely on expansive corpora of $\textit{paired}$ image-text data to attain optimal performance and generalization capabilities. However, automatically collecting such data (e.g. via large-scale web scraping) leads to low quality and poor image-text correlation, while human annotation is more accurate but requires significant manual effort and expense. We introduce $\textbf{ITIT}$ ($\textbf{I}$n$\textbf{T}$egrating $\textbf{I}$mage $\textbf{T}$ext): an innovative training paradigm grounded in the concept of cycle consistency which allows vision-language training on $\textit{unpaired}$ image and text data. ITIT is comprised of a joint image-text encoder with disjoint image and text decoders that enable bidirectional image-to-text and text-to-image generation in a single framework. During training, ITIT leverages a small set of paired image-text data to ensure its output matches the input reasonably well in both directions. Simultaneously, the model is also trained on much larger datasets containing only images or texts. This is achieved by enforcing cycle consistency between the original unpaired samples and the cycle-generated counterparts. For instance, it generates a caption for a given input image and then uses the caption to create an output image, and enforces similarity between the input and output images. Our experiments show that ITIT with unpaired datasets exhibits similar scaling behavior as using high-quality paired data. We demonstrate image generation and captioning performance on par with state-of-the-art text-to-image and image-to-text models with orders of magnitude fewer (only 3M) paired image-text data. Code will be released at https://github.com/LTH14/itit.

Proximal operators are ubiquitous in inverse problems, commonly appearing as part of algorithmic strategies to regularize problems that are otherwise ill-posed. Modern deep learning models have been brought to bear for these tasks too, as in the framework of plug-and-play or deep unrolling, where they loosely resemble proximal operators. Yet, something essential is lost in employing these purely data-driven approaches: there is no guarantee that a general deep network represents the proximal operator of any function, nor is there any characterization of the function for which the network might provide some approximate proximal. This not only makes guaranteeing convergence of iterative schemes challenging but, more fundamentally, complicates the analysis of what has been learned by these networks about their training data. Herein we provide a framework to develop *learned proximal networks* (LPN), prove that they provide exact proximal operators for a data-driven nonconvex regularizer, and show how a new training strategy, dubbed *proximal matching*, provably promotes the recovery of the log-prior of the true data distribution. Such LPN provide general, unsupervised, expressive proximal operators that can be used for general inverse problems with convergence guarantees. We illustrate our results in a series of cases of increasing complexity, demonstrating that these models not only result in state-of-the-art performance, but provide a window into the resulting priors learned from data.

Coordinate network or implicit neural representation (INR) is a fast-emerging method for encoding natural signals (such as images and videos) with the benefits of a compact neural representation. While numerous methods have been proposed to increase the encoding capabilities of an INR, an often overlooked aspect is the inference efficiency, usually measured in multiply-accumulate (MAC) count. This is particularly critical in use cases where inference bandwidth is greatly limited by hardware constraints. To this end, we propose the Activation-Sharing Multi-Resolution (ASMR) coordinate network that combines multi-resolution coordinate decomposition with hierarchical modulations. Specifically, an ASMR model enables the sharing of activations across grids of the data. This largely decouples its inference cost from its depth which is directly correlated to its reconstruction capability, and renders a near $O(1)$ inference complexity irrespective of the number of layers. Experiments show that ASMR can reduce the MAC of a vanilla SIREN model by up to 500$\times$ while achieving an even higher reconstruction quality than its SIREN baseline.

One fascinating aspect of pre-trained vision-language models (VLMs) learning under language supervision is their impressive zero-shot generalization capability. However, this ability is hindered by distribution shifts between the training and testing data. Previous test time adaptation (TTA) methods for VLMs in zero-shot classification rely on minimizing the entropy of model outputs, tending to be stuck in incorrect model predictions. In this work, we propose TTA with feedback to rectify the model output and prevent the model from becoming blindly confident. Specifically, a CLIP model is adopted as the reward model during TTA and provides feedback for the VLM. Given a single test sample, the VLM is forced to maximize the CLIP reward between the input and sampled results from the VLM output distribution. The proposed \textit{reinforcement learning with CLIP feedback~(RLCF)} framework is highly flexible and universal. Beyond the classification task, with task-specific sampling strategies and a proper reward baseline choice, RLCF can be easily extended to not only discrimination tasks like retrieval but also generalization tasks like image captioning, improving the zero-shot generalization capacity of VLMs. According to the characteristics of these VL tasks, we build different fully TTA pipelines with RLCF to improve the zero-shot generalization ability of various VLMs. Extensive experiments along with promising empirical results demonstrate the effectiveness of RLCF. The code is available at https://github.com/mzhaoshuai/RLCF.

Recent works have showcased the ability of large-scale language models (LLMs) to embody diverse personas in their responses, exemplified by prompts like ‘_You are Yoda. Explain the Theory of Relativity._’ While this ability allows personalization of LLMs and enables human behavior simulation, its effect on LLMs’ capabilities remains unclear. To fill this gap, we present the first extensive study of the unintended side-effects of persona assignment on the ability of LLMs to perform _basic reasoning tasks_. Our study covers 24 reasoning datasets (spanning mathematics, law, medicine, morals, and more), 4 LLMs (2 versions of ChatGPT-3.5, GPT-4-Turbo, and Llama-2-70b-chat), and 19 diverse personas (e.g., ‘an Asian person’) spanning 5 socio-demographic groups: race, gender, religion, disability, and political affiliation. Our experiments unveil that LLMs harbor deep rooted bias against various socio-demographics underneath a veneer of fairness. While they overtly reject stereotypes when explicitly asked (‘_Are Black people less skilled at mathematics?_’), they manifest stereotypical and often erroneous presumptions when prompted to answer questions while adopting a persona. These can be observed as abstentions in the model’s response, e.g., ‘_As a Black person, I am unable to answer this question as it requires math knowledge_’, and generally result in a substantial drop in performance on reasoning tasks. Our experiments with ChatGPT-3.5 show that this bias is _ubiquitous_—80% of our personas demonstrate bias; it is _significant_—some datasets show performance drops of 70%+; and can be especially _harmful for certain groups_—some personas suffer statistically significant drops on 80%+ of the datasets. Overall, all four LLMs exhibit persona-induced bias to varying extents, with GPT-4-Turbo showing the least but still a problematic amount of bias (evident in 42% of the personas). Further analysis shows that these persona-induced errors can be hard-to-discern as they do not always manifest as explicit abstentions, and can also be hard-to-avoid—we find de-biasing prompts to have minimal to no effect. Our findings serve as a cautionary tale that the practice of assigning personas to LLMs—a trend on the rise—can surface their deep-rooted biases and have unforeseeable and detrimental side-effects.

We design a new family of hybrid CNN-ViT neural networks, named FasterViT, with a focus on high image throughput for computer vision (CV) applications. FasterViT combines the benefits of fast local representation learning in CNNs and global modeling properties in ViT. Our newly introduced Hierarchical Attention (HAT) approach decomposes global self-attention with quadratic complexity into a multi-level attention with reduced computational costs. We benefit from efficient window-based self-attention. Each window has access to dedicated carrier tokens that participate in local and global representation learning. At a high level, global self-attentions enable the efficient cross-window communication at lower costs. FasterViT achieves a SOTA Pareto-front in terms of accuracy and image throughput. We have extensively validated its effectiveness on various CV tasks including classification, object detection and segmentation. We also show that HAT can be used as a plug-and-play module for existing networks and enhance them. We further demonstrate significantly faster and more accurate performance than competitive counterparts for images with high resolution. Code is available at https://github.com/NVlabs/FasterViT.

Multimodal VAEs have recently gained significant attention as generative models for weakly-supervised learning with multiple heterogeneous modalities. In parallel, VAE-based methods have been explored as probabilistic approaches for clustering tasks. At the intersection of these two research directions, we propose a novel multimodal VAE model in which the latent space is extended to learn data clusters, leveraging shared information across modalities. Our experiments show that our proposed model improves generative performance over existing multimodal VAEs, particularly for unconditional generation. Furthermore, we propose a post-hoc procedure to automatically select the number of true clusters thus mitigating critical limitations of previous clustering frameworks. Notably, our method favorably compares to alternative clustering approaches, in weakly-supervised settings. Finally, we integrate recent advancements in diffusion models into the proposed method to improve generative quality for real-world images.

In deep learning, test-time adaptation has gained attention as a method for model fine-tuning without the need for labeled data. A prime exemplification is the recently proposed test-time prompt tuning for large-scale vision-language models such as CLIP. Unfortunately, these prompts have been mainly developed to improve accuracy, overlooking the importance of calibration, which is a crucial aspect for quantifying prediction uncertainty. However, traditional calibration methods rely on substantial amounts of labeled data, making them impractical for test-time scenarios. To this end, this paper explores calibration during test-time prompt tuning by leveraging the inherent properties of CLIP. Through a series of observations, we find that the prompt choice significantly affects the calibration in CLIP, where the prompts leading to higher text feature dispersion result in better-calibrated predictions. Introducing the Average Text Feature Dispersion (ATFD), we establish its relationship with calibration error and present a novel method, Calibrated Test-time Prompt Tuning (C-TPT), for optimizing prompts during test-time with enhanced calibration. Through extensive experiments on different CLIP architectures and datasets, we show that C-TPT can effectively improve the calibration of test-time prompt tuning without needing labeled data. The code is publicly accessible at https://github.com/hee-suk-yoon/C-TPT.

Our understanding of the visual world is centered around various concept axes, characterizing different aspects of visual entities. While different concept axes can be easily specified by language, e.g., color, the exact visual nuances along each axis often exceed the limitations of linguistic articulations, e.g., a particular style of painting. In this work, our goal is to learn a language-informed visual concept representation, by simply distilling large pre-trained vision-language models. Specifically, we train a set of concept encoders to encode the information pertinent to a set of language-informed concept axes, with an objective of reproducing the input image through a pre-trained Text-to-Image (T2I) model. To encourage better disentanglement of different concept encoders, we anchor the concept embeddings to a set of text embeddings obtained from a pre-trained Visual Question Answering (VQA) model. At inference time, the model extracts concept embeddings along various axes from new test images, which can be remixed to generate images with novel compositions of visual concepts. With a lightweight test-time finetuning procedure, it can also generalize to novel concepts unseen at training. Project page at https://cs.stanford.edu/~yzzhang/projects/concept-axes.

Foundational models with billions of parameters which have been trained on large corpus of data have demonstrated non-trivial skills in a variety of domains. However, due to their monolithic structure, it is challenging and expensive to augment them or impart new skills. On the other hand, due to their adaptation abilities,several new instances of these models are being trained towards new domains and tasks. In this work, we study the problem of efficient and practical composition of existing foundation models with more specific models to enable newer capabilities. To this end, we propose CALM—Composition to Augment Language Models—which introduces cross-attention between models to compose their representations and enable new capabilities. Salient features of CALM are: (i) Scales up LLMs on new tasks by ‘re-using’ existing LLMs along with a few additional parameters and data, (ii) Existing model weights are kept intact, and hence preserves existing capabilities, and (iii) Applies to diverse domains and settings. We illustrate that augmenting PaLM2-S with a smaller model trained on low-resource languages results in an absolute improvement of up to 13% on tasks like translation into English and arithmetic reasoning for low-resource languages. Similarly,when PaLM2-S is augmented with a code-specific model, we see a relative improvement of 40% over the base model for code generation and explanation tasks—on-par with fully fine-tuned counterparts.