Language models demonstrate both quantitative improvement and new qualitative capabilities with increasing scale. Despite their potentially transformative impact, these new capabilities are as yet poorly characterized. In order to inform future research, prepare for disruptive new model capabilities, and ameliorate socially harmful effects, it is vital that we understand the present and near-future capabilities and limitations of language models. To address this challenge, we introduce the Beyond the Imitation Game benchmark (BIG-bench). BIG-bench currently consists of 204 tasks, contributed by 442 authors across 132 institutions. Task topics are diverse, drawing problems from linguistics, childhood development, math, common-sense reasoning, biology, physics, social bias, software development, and beyond. BIG-bench focuses on tasks that are believed to be beyond the capabilities of current language models. We evaluate the behavior of OpenAI's GPT models, Google-internal dense transformer architectures, and Switch-style sparse transformers on BIG-bench, across model sizes spanning millions to hundreds of billions of parameters. In addition, a team of human expert raters performed all tasks in order to provide a strong baseline. Findings include: model performance and calibration both improve with scale, but are poor in absolute terms (and when compared with rater performance); performance is remarkably similar across model classes, though with benefits from sparsity; tasks that improve gradually and predictably commonly involve a large knowledge or memorization component, whereas tasks that exhibit "breakthrough" behavior at a critical scale often involve multiple steps or components, or brittle metrics; social bias typically increases with scale in settings with ambiguous context, but this can be improved with prompting.
Prior work has proposed a simple strategy for reinforcement learning (RL): label experience with the outcomes achieved in that experience, and then imitate the relabeled experience. These outcome-conditioned imitation learning methods are appealing because of their simplicity, strong performance, and close ties with supervised learning. However, it remains unclear how these methods relate to the standard RL objective, reward maximization. In this paper, we prove that existing outcome-conditioned imitation learning methods do not necessarily improve the policy; rather, in some settings they can decrease the expected reward. Nonetheless, we show that a simple modification results in a method that does guarantee policy improvement, under some assumptions. Our aim is not to develop an entirely new method, but rather to explain how a variant of outcome-conditioned imitation learning can be used to maximize rewards.
Some questions have multiple answers that are not equally correct, i.e. answers are different under different conditions. Conditions are used to distinguish answers as well as to provide additional information to support them. In this paper, we study a more challenging task where answers are constrained by a list of conditions that logically interact, which requires performing logical reasoning over the conditions to determine the correctness of the answers. Even more challenging, we only provide evidences for a subset of the conditions, so some questions may not have deterministic answers. In such cases, models are asked to find probable answers and identify conditions that need to be satisfied to make the answers correct. We propose a new model, TReasoner, for this challenging reasoning task. TReasoner consists of an entailment module, a reasoning module, and a generation module (if the answers are free-form text spans). TReasoner achieves state-of-the-art performance on two benchmark conditional QA datasets, outperforming the previous state-of-the-art by 3-10 points.
In order for AI to be safely deployed in real-world scenarios such as hospitals, schools, and the workplace, they should be able to reason about the physical world by understanding the physical properties and affordances of available objects, how they can be manipulated, and how they interact with other physical objects. This research field of physical commonsense reasoning is fundamentally a multi-sensory task since physical properties are manifested through multiple modalities, two of them being vision and acoustics. Our paper takes a step towards real-world physical commonsense reasoning by contributing PACS: the first audiovisual benchmark annotated for physical commonsense attributes. PACS contains a total of 13,400 question-answer pairs, involving 1,377 unique physical commonsense questions and 1,526 videos. Our dataset provides new opportunities to advance the research field of physical reasoning by bringing audio as a core component of this multimodal problem. Using PACS, we evaluate multiple state-of-the-art models on this new challenging task. While some models show promising results (70% accuracy), they all fall short of human performance (95% accuracy). We conclude the paper by demonstrating the importance of multimodal reasoning and providing possible avenues for future research.
Learning multimodal representations involves discovering correspondences and integrating information from multiple heterogeneous sources of data. While recent research has begun to explore the design of more general-purpose multimodal models (contrary to prior focus on domain and modality-specific architectures), these methods are still largely focused on a small set of modalities in the language, vision, and audio space. In order to accelerate generalization towards diverse and understudied modalities, we investigate methods for high-modality (a large set of diverse modalities) and partially-observable (each task only defined on a small subset of modalities) scenarios. To tackle these challenges, we design a general multimodal model that enables multitask and transfer learning: multitask learning with shared parameters enables stable parameter counts (addressing scalability), and cross-modal transfer learning enables information sharing across modalities and tasks (addressing partial observability). Our resulting model generalizes across text, image, video, audio, time-series, sensors, tables, and set modalities from different research areas, improves the tradeoff between performance and efficiency, transfers to new modalities and tasks, and reveals surprising insights on the nature of information sharing in multitask models. We release our code and benchmarks which we hope will present a unified platform for subsequent theoretical and empirical analysis: https://github.com/pliang279/HighMMT.
How can we make predictions for nodes in a heterogeneous graph when an entire type of node (e.g. user) has no labels (perhaps due to privacy issues) at all? Although heterogeneous graph neural networks (HGNNs) have shown superior performance as powerful representation learning techniques, there is no direct way to learn using labels rooted at different node types. Domain adaptation (DA) targets this setting, however, existing DA can not be applied directly to HGNNs. In heterogeneous graphs, the source and target domains have different modalities, thus HGNNs provide different feature extractors to them, while most of DA assumes source and target domains share a common feature extractor. In this work, we address the issue of zero-shot domain adaptation in HGNNs. We first theoretically induce a relationship between source and target domain features extracted from HGNNs, then propose a novel domain adaptation method, Knowledge Transfer Networks for HGNNs (HGNN-KTN). HGNN-KTN learns the relationship between source and target features, then maps the target distributions into the source domain. HGNN-KTN outperforms state-of-the-art baselines, showing up to 73.3% higher in MRR on 18 different domain adaptation tasks running on real-world benchmark graphs.
The ability for a human to understand an Artificial Intelligence (AI) model's decision-making process is critical in enabling stakeholders to visualize model behavior, perform model debugging, promote trust in AI models, and assist in collaborative human-AI decision-making. As a result, the research fields of interpretable and explainable AI have gained traction within AI communities as well as interdisciplinary scientists seeking to apply AI in their subject areas. In this paper, we focus on advancing the state-of-the-art in interpreting multimodal models - a class of machine learning methods that tackle core challenges in representing and capturing interactions between heterogeneous data sources such as images, text, audio, and time-series data. Multimodal models have proliferated numerous real-world applications across healthcare, robotics, multimedia, affective computing, and human-computer interaction. By performing model disentanglement into unimodal contributions (UC) and multimodal interactions (MI), our proposed approach, DIME, enables accurate and fine-grained analysis of multimodal models while maintaining generality across arbitrary modalities, model architectures, and tasks. Through a comprehensive suite of experiments on both synthetic and real-world multimodal tasks, we show that DIME generates accurate disentangled explanations, helps users of multimodal models gain a deeper understanding of model behavior, and presents a step towards debugging and improving these models for real-world deployment. Code for our experiments can be found at https://github.com/lvyiwei1/DIME.
We argue that a form of the valuable information provided by the auxiliary information is its implied data clustering information. For instance, considering hashtags as auxiliary information, we can hypothesize that an Instagram image will be semantically more similar with the same hashtags. With this intuition, we present a two-stage weakly-supervised contrastive learning approach. The first stage is to cluster data according to its auxiliary information. The second stage is to learn similar representations within the same cluster and dissimilar representations for data from different clusters. Our empirical experiments suggest the following three contributions. First, compared to conventional self-supervised representations, the auxiliary-information-infused representations bring the performance closer to the supervised representations, which use direct downstream labels as supervision signals. Second, our approach performs the best in most cases, when comparing our approach with other baseline representation learning methods that also leverage auxiliary data information. Third, we show that our approach also works well with unsupervised constructed clusters (e.g., no auxiliary information), resulting in a strong unsupervised representation learning approach.
Conditional contrastive learning frameworks consider the conditional sampling procedure that constructs positive or negative data pairs conditioned on specific variables. Fair contrastive learning constructs negative pairs, for example, from the same gender (conditioning on sensitive information), which in turn reduces undesirable information from the learned representations; weakly supervised contrastive learning constructs positive pairs with similar annotative attributes (conditioning on auxiliary information), which in turn are incorporated into the representations. Although conditional contrastive learning enables many applications, the conditional sampling procedure can be challenging if we cannot obtain sufficient data pairs for some values of the conditioning variable. This paper presents Conditional Contrastive Learning with Kernel (CCL-K) that converts existing conditional contrastive objectives into alternative forms that mitigate the insufficient data problem. Instead of sampling data according to the value of the conditioning variable, CCL-K uses the Kernel Conditional Embedding Operator that samples data from all available data and assigns weights to each sampled data given the kernel similarity between the values of the conditioning variable. We conduct experiments using weakly supervised, fair, and hard negatives contrastive learning, showing CCL-K outperforms state-of-the-art baselines.