ICML 2024 Papers

This paper proposes a payoff perturbation technique for the Mirror Descent (MD) algorithm in games where the gradient of the payoff functions is monotone in the strategy profile space, potentially containing additive noise. The optimistic family of learning algorithms, exemplified by optimistic MD, successfully achieves *last-iterate* convergence in scenarios devoid of noise, leading the dynamics to a Nash equilibrium. A recent re-emerging trend underscores the promise of the perturbation approach, where payoff functions are perturbed based on the distance from an anchoring, or *slingshot*, strategy. In response, we propose *Adaptively Perturbed MD* (APMD), which adjusts the magnitude of the perturbation by repeatedly updating the slingshot strategy at a predefined interval. This innovation empowers us to find a Nash equilibrium of the underlying game with guaranteed rates. Empirical demonstrations affirm that our algorithm exhibits significantly accelerated convergence.

Are foundation models secure against malicious actors? In this work, we focus on the image input to a vision-language model (VLM). We discover image hijacks, adversarial images that control the behaviour of VLMs at inference time, and introduce the general Behaviour Matching algorithm for training image hijacks. From this, we derive the Prompt Matching method, allowing us to train hijacks matching the behaviour of an arbitrary user-defined text prompt (e.g. 'the Eiffel Tower is now located in Rome') using a generic, off-the-shelf dataset unrelated to our choice of prompt. We use Behaviour matching to craft hijacks for four types of attack: forcing VLMs to generate outputs of the adversary’s choice, leak information from their context window, override their safety training, and believe false statements. We study these attacks against LLaVA, a state-of-the-art VLM based on CLIP and LLaMA-2, and find that all attack types achieve a success rate of over 80%. Moreover, our attacks are automated and require only small image perturbations.

Principal Component Analysis (PCA) aims to find subspaces spanned by the so-called *principal components* that best represent the variance in the dataset. The deflation method is a popular meta-algorithm that sequentially finds individual principal components, starting from the most important ones and working towards the less important ones. However, as deflation proceeds, numerical errors from the imprecise estimation of principal components propagate due to its sequential nature. This paper mathematically characterizes the error propagation of the inexact Hotelling's deflation method. We consider two scenarios: $i)$ when the sub-routine for finding the leading eigenvector is abstract and can represent various algorithms; and $ii)$ when power iteration is used as the sub-routine. In the latter case, the additional directional information from power iteration allows us to obtain a tighter error bound than the sub-routine agnostic case. For both scenarios, we explicitly characterize how the errors progress and affect subsequent principal component estimations.

We seek to enable classic processing of continuous ultra-sparse spatiotemporal data generated by event-based sensors with dense machine learning models. We propose a novel hybrid pipeline composed of asynchronous sensing and synchronous processing that combines several ideas: (1) an embedding based on PointNet models -- the ALERT module -- that can continuously integrate new and dismiss old events thanks to a leakage mechanism, (2) a flexible readout of the embedded data that allows to feed any downstream model with always up-to-date features at any sampling rate, (3) exploiting the input sparsity in a patch-based approach inspired by Vision Transformer to optimize the efficiency of the method. These embeddings are then processed by a transformer model trained for object and gesture recognition. Using this approach, we achieve performances at the state-of-the-art with a lower latency than competitors. We also demonstrate that our asynchronous model can operate at any desired sampling rate.

Recently, multimodal machine learning has enjoyed huge empirical success (e.g. GPT-4). Motivated to develop theoretical justification for this empirical success, Lu (NeurIPS '23, ALT '24) introduces a theory of multimodal learning, and considers possible *separations* between theoretical models of multimodal and unimodal learning. In particular, Lu (ALT '24) shows a computational separation, which is relevant to *worst-case* instances of the learning task. In this paper, we give a stronger *average-case* computational separation, where for "typical" instances of the learning task, unimodal learning is computationally hard, but multimodal learning is easy. We then question how "natural" the average-case separation is. Would it be encountered in practice? To this end, we prove that under basic conditions, any given computational separation between average-case unimodal and multimodal learning tasks implies a corresponding cryptographic key agreement protocol. We suggest to interpret this as evidence that very strong *computational* advantages of multimodal learning may arise *infrequently* in practice, since they exist only for the "pathological" case of inherently cryptographic distributions. However, this does not apply to possible (super-polynomial) *statistical* advantages.

Medical vision-language pre-training methods mainly leverage the correspondence between paired medical images and radiological reports. Although multi-view spatial images and temporal sequences of image-report pairs are available in off-the-shelf multi-modal medical datasets, most existing methods have not thoroughly tapped into such extensive supervision signals. In this paper, we introduce the Med-ST framework for fine-grained spatial and temporal modeling to exploit information from multiple spatial views of chest radiographs and temporal historical records. For spatial modeling, Med-ST employs the *Mixture of View Expert (MoVE)* architecture to integrate different visual features from both frontal and lateral views. To achieve a more comprehensive alignment, Med-ST not only establishes the global alignment between whole images and texts but also introduces modality-weighted local alignment between text tokens and spatial regions of images. For temporal modeling, we propose a novel cross-modal bidirectional cycle consistency objective by forward mapping classification (FMC) and reverse mapping regression (RMR). By perceiving temporal information from simple to complex, Med-ST can learn temporal semantics. Experimental results across four distinct tasks demonstrate the effectiveness of Med-ST, especially in temporal classification tasks. Our code and model are available at https://github.com/SVT-Yang/MedST.

The multi-plane representation has been highlighted for its fast training and inference across static and dynamic neural radiance fields. This approach constructs relevant features via projection onto learnable grids and interpolating adjacent vertices. However, it has limitations in capturing low-frequency details and tends to overuse parameters for low-frequency features due to its bias toward fine details, despite its multi-resolution concept. This phenomenon leads to instability and inefficiency when training poses are sparse. In this work, we propose a method that synergistically integrates multi-plane representation with a coordinate-based MLP network known for strong bias toward low-frequency signals. The coordinate-based network is responsible for capturing low-frequency details, while the multi-plane representation focuses on capturing fine-grained details. We demonstrate that using residual connections between them seamlessly preserves their own inherent properties. Additionally, the proposed progressive training scheme accelerates the disentanglement of these two features. We demonstrate empirically that our proposed method not only outperforms baseline models for both static and dynamic NeRFs with sparse inputs, but also achieves comparable results with fewer parameters.

We propose the attention-inspired numerical solver (AttNS), a concise method that helps the generalization and robustness issues faced by the AI-Hybrid numerical solver in solving differential equations due to limited data. AttNS is inspired by the effectiveness of attention modules in Residual Neural Networks (ResNet) in enhancing model generalization and robustness for conventional deep learning tasks. Drawing from the dynamical system perspective of ResNet, We seamlessly incorporate attention mechanisms into the design of numerical methods tailored for the characteristics of solving differential equations. Our results on benchmarks, ranging from high-dimensional problems to chaotic systems, showcase AttNS consistently enhancing various numerical solvers without any intricate model crafting. Finally, we analyze AttNS experimentally and theoretically, demonstrating its ability to achieve strong generalization and robustness while ensuring the convergence of the solver. This includes requiring less data compared to other advanced methods to achieve comparable generalization errors and better prevention of numerical explosion issues when solving differential equations.

Large Language Models (LLMs) excel in diverse areas, yet struggle with complex scientific reasoning, especially in the field of chemistry. Different from the simple chemistry tasks (e.g., molecule classification) addressed in previous studies, complex chemistry problems require not only vast knowledge and precise calculation, but also compositional reasoning about rich dynamic interactions of different concepts (e.g., temperature changes). Our study shows that even advanced LLMs, like GPT-4, can fail easily in different ways. Interestingly, the errors often stem not from a lack of domain knowledge within the LLMs, but rather from the absence of an effective reasoning *structure* that guides the LLMs to elicit the right knowledge, incorporate the knowledge in step-by-step reasoning, and iteratively refine results for further improved quality. On this basis, we introduce StructChem, a simple yet effective prompting strategy that offers the desired guidance and substantially boosts the LLMs' chemical reasoning capability. Testing across four chemistry areas---quantum chemistry, mechanics, physical chemistry, and kinetics---StructChem substantially enhances GPT-4's performance, with up to 30% peak improvement. Our analysis also underscores the unique difficulties of precise grounded reasoning in science with LLMs, highlighting a need for more research in this area.

Growing concerns over negligent or malicious uses of AI have increased the appetite for tools that help manage the risks of the technology. In 2018, licenses with behaviorial-use clauses (commonly referred to as Responsible AI Licenses) were proposed to give developers a framework for releasing AI assets while specifying their users to mitigate negative applications. As of the end of 2023, on the order of 40,000 software and model repositories have adopted responsible AI licenses licenses. Notable models licensed with behavioral use clauses include BLOOM (language) and LLaMA2 (language), Stable Diffusion (image), and GRID (robotics). This paper explores why and how these licenses have been adopted, and why and how they have been adapted to fit particular use cases. We use a mixed-methods methodology of qualitative interviews, clustering of license clauses, and quantitative analysis of license adoption. Based on this evidence we take the position that responsible AI licenses need standardization to avoid confusing users or diluting their impact. At the same time, customization of behavioral restrictions is also appropriate in some contexts (e.g., medical domains). We advocate for “standardized customization” that can meet users’ needs and can be supported via tooling.

Artificial intelligence (AI) has seen a tremendous surge in capabilities thanks to the use of foundation models trained on internet-scale data. On the flip side, the uncurated nature of internet-scale data also poses significant privacy and legal risks, as they often contain personal information or copyrighted material that should not be trained on without permission. In this work, we propose as a mitigation measure a recipe to train foundation vision models via self-supervised learning with differential privacy (DP) guarantee. We identify masked autoencoders as a suitable learning algorithm that aligns well with DP-SGD, and train *ViP*---a **Vi**sion transformer with differential **P**rivacy---under a strict privacy budget of $\epsilon=8$ on the LAION400M dataset. We evaluate the quality of representation learned by ViP using standard downstream vision tasks; in particular, ViP achieves a (non-private) linear probing accuracy of 55.7% on ImageNet, comparable to that of end-to-end trained AlexNet (trained and evaluated on ImageNet). Our result suggests that scaling to internet-scale data can be practical for private learning. Code and DP pre-trained models are available at https://github.com/facebookresearch/ViP-MAE.

This paper proposes the use of "multicalibration": to yield interpretable and reliable confidence scores for outputs generated by large language models (LLMs). Multicalibration asks for calibration not just marginally, but simultaneously across various intersecting groupings of the data. We show how to form groupings for prompt/completion pairs that are correlated with the probability of correctness via two techniques: clustering within an embedding space, and "self-annotation" - querying the LLM by asking it various yes-or-no questions about the prompt. We also develop novel variants of multicalibration algorithms that offer performance improvements by reducing their tendency to overfit. Through systematic benchmarking across various question answering datasets and LLMs, we show how our techniques can yield confidence scores that provide substantial improvements in fine-grained measures of both calibration and accuracy compared to existing methods.

Pre-trained Vision-Language Models (VLMs) are able to understand visual concepts, describe and decompose complex tasks into sub-tasks, and provide feedback on task completion. In this paper, we aim to leverage these capabilities to support the training of reinforcement learning (RL) agents. In principle, VLMs are well suited for this purpose, as they can naturally analyze image-based observations and provide feedback (reward) on learning progress. However, inference in VLMs is computationally expensive, so querying them frequently to compute rewards would significantly slowdown the training of an RL agent. To address this challenge, we propose a framework named Code as Reward (VLM-CaR). VLM-CaR produces dense reward functions from VLMs through code generation, thereby significantly reducing the computational burden of querying the VLM directly. We show that the dense rewards generated through our approach are very accurate across a diverse set of discrete and continuous environments, and can be more effective in training RL policies than the original sparse environment rewards.

Modern language models can contain billions of parameters, raising the question of whether they can generalize beyond the training data or simply parrot their training corpora. We provide the first non-vacuous generalization bounds for pretrained large language models (LLMs), indicating that language models are capable of discovering regularities that generalize to unseen data. In particular, we derive a compression bound that is valid for the unbounded log-likelihood loss using prediction smoothing, and we extend the bound to handle subsampling, making bound computation 900 times faster on massive datasets. To achieve the extreme level of compression required for non-vacuous bounds, we devise SubLoRA, a simple low-dimensional nonlinear parameterization that leads to non-vacuous generalization bounds for very large models with up to 849 million parameters. Finally, we use our bounds to understand LLM generalization and find that larger models have better generalization bounds and are more compressible than smaller models.

Visually-conditioned language models (VLMs) have seen growing adoption in applications such as visual dialogue, scene understanding, and robotic task planning; adoption that has fueled a wealth of new models such as LLaVa, InstructBLIP, and PaLI-3. Despite the volume of new releases, key design decisions around image preprocessing, architecture, and optimization are under-explored, making it challenging to understand what factors account for model performance – a challenge further complicated by the lack of objective, consistent evaluations. To address these gaps, we first compile a suite of standardized evaluations spanning visual question answering, object localization, and challenge sets that probe properties such as hallucination; evaluations that provide fine-grained insight VLM capabilities. Second, we rigorously investigate VLMs along key design axes, including pretrained visual representations and training from base vs. instruct-tuned language models, amongst others. We couple our analysis with three resource contributions: (1) a unified framework for evaluating VLMs, (2) optimized, flexible training code, and (3) checkpoints for all models, including a family of VLMs at the 7-13B scale that strictly outperform InstructBLIP and LLaVa v1.5, the state-of-the-art in open VLMs.

Reconstructions of visual perception from brain activity have improved tremendously, but the practical utility of such methods has been limited. This is because such models are trained independently per subject where each subject requires dozens of hours of expensive fMRI training data to attain high-quality results. The present work showcases high-quality reconstructions using only 1 hour of fMRI training data. We pretrain our model across 7 subjects and then fine-tune on minimal data from a new subject. Our novel functional alignment procedure linearly maps all brain data to a shared-subject latent space, followed by a shared non-linear mapping to CLIP image space. We then map from CLIP space to pixel space by fine-tuning Stable Diffusion XL to accept CLIP latents as inputs instead of text. This approach improves out-of-subject generalization with limited training data and also attains state-of-the-art image retrieval and reconstruction metrics compared to single-subject approaches. MindEye2 demonstrates how accurate reconstructions of perception are possible from a single visit to the MRI facility. All code is available on Github: https://github.com/MedARC-AI/MindEyeV2

Understanding the internal representations of large language models (LLMs) can help explain models' behavior and verify their alignment with human values. Given the capabilities of LLMs in generating human-understandable text, we propose leveraging the model itself to explain its internal representations in natural language. We introduce a framework called Patchscopes and show how it can be used to answer a wide range of questions about an LLM's computation. We show that many prior interpretability methods based on projecting representations into the vocabulary space and intervening on the LLM computation can be viewed as instances of this framework. Moreover, several of their shortcomings such as failure in inspecting early layers or lack of expressivity can be mitigated by Patchscopes. Beyond unifying prior inspection techniques, Patchscopes also opens up *new* possibilities such as using a more capable model to explain the representations of a smaller model, and multihop reasoning error correction.

We introduce SPARse fine-grained Contrastive alignment (SPARC), a simple method for pretraining more fine-grained multimodal representations from image-text pairs. Given that multiple image patches often correspond to single words, we propose to learn a grouping of image patches for every token in the caption. To achieve this, we use a sparse similarity metric between image patches and language tokens and compute for each token a language-grouped vision embedding as the weighted average of patches. The token and language-grouped vision embeddings are then contrasted through a fine-grained sequence-wise loss that only depends on individual samples and does not require other batch samples as negatives, i.e., more detailed information is encoded in a computationally inexpensive way. SPARC combines this fine-grained loss with a contrastive loss between global image and text embeddings to learn representations that simultaneously encode global and local information. We thoroughly evaluate SPARC and show improved performance over competing approaches both on image-level tasks relying on coarse-grained information, e.g. classification, as well as region-level tasks relying on fine-grained information, e.g., retrieval, object detection, segmentation while also improving model faithfulness and captioning in foundational vision-language models.

Pretrained vision language models (VLMs) present an opportunity to caption unlabeled 3D objects at scale. The leading approach to summarize VLM descriptions from different views of an object (Luo et al., 2023) relies on a language model (GPT4) to produce the final output. This text-based aggregation is susceptible to hallucinations as it merges potentially contradictory descriptions. We propose an alternative algorithm to marginalize over factors such as the viewpoint that affect the VLM's response. Instead of merging text-only responses, we utilize the VLM's joint image-text likelihoods. We show our probabilistic aggregation is not only more reliable and efficient, but sets the SoTA on inferring object types with respect to human-verified labels. The aggregated annotations are also useful for conditional inference; they improve downstream predictions (e.g., of object material) when the object’s type is specified as an auxiliary text-based input. Such auxiliary inputs allow ablating the contribution of visual reasoning over visionless reasoning in an unsupervised setting. With these supervised and unsupervised evaluations, we show how a VLM-based pipeline can be leveraged to produce reliable annotations for 764K objects from the Objaverse dataset.