NeurIPS 2023 论文列表

Joint speech-language training is challenging due to the large demand for training data and GPU consumption, as well as the modality gap between speech and language. We present ComSL, a speech-language model built atop a composite architecture of public pre-trained speech-only and language-only models and optimized data-efficiently for spoken language tasks. Particularly, we propose to incorporate cross-modality learning into transfer learning and conduct them simultaneously for downstream tasks in a multi-task learning manner. Our approach has demonstrated effectiveness in end-to-end speech-to-text translation tasks, achieving a new state-of-the-art average BLEU score of 31.5 on the multilingual speech to English text translation task for 21 languages, as measured on the public CoVoST2 evaluation set.

Learning feature interactions can be the key for multivariate predictive modeling. ReLU-activated neural networks create piecewise linear prediction models, and other nonlinear activation functions lead to models with only high-order feature interactions. Recent methods incorporate candidate polynomial terms of fixed orders into deep learning, which is subject to the issue of combinatorial explosion, or learn the orders that are difficult to adapt to different regions of the feature space. We propose a Polyhedron Attention Module (PAM) to create piecewise polynomial models where the input space is split into polyhedrons which define the different pieces and on each piece the hyperplanes that define the polyhedron boundary multiply to form the interactive terms, resulting in interactions of adaptive order to each piece. PAM is interpretable to identify important interactions in predicting a target. Theoretic analysis shows that PAM has stronger expression capability than ReLU-activated networks. Extensive experimental results demonstrate the superior classification performance of PAM on massive datasets of the click-through rate prediction and PAM can learn meaningful interaction effects in a medical problem.

We present Prompt Diffusion, a framework for enabling in-context learning in diffusion-based generative models. Given a pair of task-specific example images, such as depth from/to image and scribble from/to image, and a text guidance, our model automatically understands the underlying task and performs the same task on a new query image following the text guidance. To achieve this, we propose a vision-language prompt that can model a wide range of vision-language tasks and a diffusion model that takes it as input. The diffusion model is trained jointly on six different tasks using these prompts. The resulting Prompt Diffusion model becomes the first diffusion-based vision-language foundation model capable of in-context learning. It demonstrates high-quality in-context generation for the trained tasks and effectively generalizes to new, unseen vision tasks using their respective prompts. Our model also shows compelling text-guided image editing results. Our framework aims to facilitate research into in-context learning for computer vision. We share our code and pre-trained models at https://github.com/Zhendong-Wang/Prompt-Diffusion.

Modern neural language models that are widely used in various NLP tasks risk memorizing sensitive information from their training data. Understanding this memorization is important in real world applications and also from a learning-theoretical perspective. An open question in previous studies of language model memorization is how to filter out ``common'' memorization. In fact, most memorization criteria strongly correlate with the number of occurrences in the training set, capturing memorized familiar phrases, public knowledge, templated texts, or other repeated data. We formulate a notion of counterfactual memorization which characterizes how a model's predictions change if a particular document is omitted during training. We identify and study counterfactually-memorized training examples in standard text datasets. We estimate the influence of each memorized training example on the validation set and on generated texts, showing how this can provide direct evidence of the source of memorization at test time.

The tremendous success of large models trained on extensive datasets demonstrates that scale is a key ingredient in achieving superior results. Therefore, the reflection on the rationality of designing knowledge distillation (KD) approaches for limited-capacity architectures solely based on small-scale datasets is now deemed imperative. In this paper, we identify the small data pitfall that presents in previous KD methods, which results in the underestimation of the power of vanilla KD framework on large-scale datasets such as ImageNet-1K. Specifically, we show that employing stronger data augmentation techniques and using larger datasets can directly decrease the gap between vanilla KD and other meticulously designed KD variants. This highlights the necessity of designing and evaluating KD approaches in the context of practical scenarios, casting off the limitations of small-scale datasets. Our investigation of the vanilla KD and its variants in more complex schemes, including stronger training strategies and different model capacities, demonstrates that vanilla KD is elegantly simple but astonishingly effective in large-scale scenarios. Without bells and whistles, we obtain state-of-the-art ResNet-50, ViT-S, and ConvNeXtV2-T models for ImageNet, which achieve 83.1%, 84.3%, and 85.0% top-1 accuracy, respectively. PyTorch code and checkpoints can be found at https://github.com/Hao840/vanillaKD.

Although pairwise causal relations have been extensively studied in observational longitudinal analyses across many disciplines, incorporating knowledge of causal pairs into deep learning models for temporal event sequences remains largely unexplored. In this paper, we propose a novel approach for enhancing the performance of transformer-based models in multivariate event sequences by injecting pairwise qualitative causal knowledge such as `event Z amplifies future occurrences of event Y'. We establish a new framework for causal inference in temporal event sequences using a transformer architecture, providing a theoretical justification for our approach, and show how to obtain unbiased estimates of the proposed measure. Experimental results demonstrate that our approach outperforms several state-of-the-art models in terms of prediction accuracy by effectively leveraging knowledge about causal pairs. We also consider a unique application where we extract knowledge around sequences of societal events by generating them from a large language model, and demonstrate how a causal knowledge graph can help with event prediction in such sequences. Overall, our framework offers a practical means of improving the performance of transformer-based models in multivariate event sequences by explicitly exploiting pairwise causal information.

Quantifying uncertainty is important for actionable predictions in real-world applications. A crucial part of predictive uncertainty quantification is the estimation of epistemic uncertainty, which is defined as an integral of the product between a divergence function and the posterior. Current methods such as Deep Ensembles or MC dropout underperform at estimating the epistemic uncertainty, since they primarily consider the posterior when sampling models. We suggest Quantification of Uncertainty with Adversarial Models (QUAM) to better estimate the epistemic uncertainty. QUAM identifies regions where the whole product under the integral is large, not just the posterior. Consequently, QUAM has lower approximation error of the epistemic uncertainty compared to previous methods. Models for which the product is large correspond to adversarial models (not adversarial examples!). Adversarial models have both a high posterior as well as a high divergence between their predictions and that of a reference model. Our experiments show that QUAM excels in capturing epistemic uncertainty for deep learning models and outperforms previous methods on challenging tasks in the vision domain.

Despite widespread use of LLMs as conversational agents, evaluations of performance fail to capture a crucial aspect of communication: interpreting language in context---incorporating its pragmatics. Humans interpret language using beliefs and prior knowledge about the world. For example, we intuitively understand the response "I wore gloves" to the question "Did you leave fingerprints?" as meaning "No". To investigate whether LLMs have the ability to make this type of inference, known as an implicature, we design a simple task and evaluate four categories of widely used state-of-the-art models. We find that, despite only evaluating on utterances that require a binary inference (yes or no), models in three of these categories perform close to random. However, LLMs instruction-tuned at the example-level perform significantly better. These results suggest that certain fine-tuning strategies are far better at inducing pragmatic understanding in models. We present our findings as the starting point for further research into evaluating how LLMs interpret language in context and to drive the development of more pragmatic and useful models of human discourse.

Language models are increasingly being deployed for general problem solving across a wide range of tasks, but are still confined to token-level, left-to-right decision-making processes during inference. This means they can fall short in tasks that require exploration, strategic lookahead, or where initial decisions play a pivotal role. To surmount these challenges, we introduce a new framework for language model inference, Tree of Thoughts (ToT), which generalizes over the popular Chain of Thought approach to prompting language models, and enables exploration over coherent units of text (thoughts) that serve as intermediate steps toward problem solving. ToT allows LMs to perform deliberate decision making by considering multiple different reasoning paths and self-evaluating choices to decide the next course of action, as well as looking ahead or backtracking when necessary to make global choices. Our experiments show that ToT significantly enhances language models’ problem-solving abilities on three novel tasks requiring non-trivial planning or search: Game of 24, Creative Writing, and Mini Crosswords. For instance, in Game of 24, while GPT-4 with chain-of-thought prompting only solved 4\% of tasks, our method achieved a success rate of 74\%. Code repo with all prompts: https://github.com/princeton-nlp/tree-of-thought-llm.

Recently, a new class of non-convex optimization problems motivated by the statistical problem of learning an acyclic directed graphical model from data has attracted significant interest. While existing work uses standard first-order optimization schemes to solve this problem, proving the global optimality of such approaches has proven elusive. The difficulty lies in the fact that unlike other non-convex problems in the literature, this problem is not "benign", and possesses multiple spurious solutions that standard approaches can easily get trapped in. In this paper, we prove that a simple path-following optimization scheme globally converges to the global minimum of the population loss in the bivariate setting.

A central notion in practical and theoretical machine learning is that of a *weak learner*, classifiers that achieve better-than-random performance (on any given distribution over data), even by a small margin. Such weak learners form the practical basis for canonical machine learning methods such as boosting. In this work, we illustrate that prompt-based large language models can operate effectively as said weak learners. Specifically, we illustrate the use of a large language model (LLM) as a weak learner in a boosting algorithm applied to tabular data. We show that by providing (properly sampled according to the distribution of interest) text descriptions of tabular data samples, LLMs can produce a summary of the samples that serves as a template for classification, and achieves the aim of acting as a weak learner on this task. We incorporate these models into a boosting approach, which in many settings can leverage the knowledge within the LLM to outperform traditional tree-based boosting. The model outperforms both few-shot learning and occasionally even more involved fine-tuning procedures, particularly for some tasks involving small numbers of data points. The results illustrate the potential for prompt-based LLMs to function not just as few-shot learners themselves, but as components of larger machine learning models.

Machine learning based decision-support tools in criminal justice systems are subjects of intense discussions and academic research. There are important open questions about the utility and fairness of such tools. Academic researchers often rely on a few small datasets that are not sufficient to empirically study various real-world aspects of these questions. In this paper, we contribute WCLD, a curated large dataset of 1.5 million criminal cases from circuit courts in the U.S. state of Wisconsin. We used reliable public data from 1970 to 2020 to curate attributes like prior criminal counts and recidivism outcomes. The dataset contains large number of samples from five racial groups, in addition to information like sex and age (at judgment and first offense). Other attributes in this dataset include neighborhood characteristics obtained from census data, detailed types of offense, charge severity, case decisions, sentence lengths, year of filing etc. We also provide pseudo-identifiers for judge, county and zipcode. The dataset will not only enable researchers to more rigorously study algorithmic fairness in the context of criminal justice, but also relate algorithmic challenges with various systemic issues. We also discuss in detail the process of constructing the dataset and provide a datasheet. The WCLD dataset is available at https://clezdata.github.io/wcld/.

Recent work has shown that language models' (LMs) prompt-based learning capabilities make them well suited for automating data labeling in domains where manual annotation is expensive. The challenge is that while writing an initial prompt is cheap, improving a prompt is costly---practitioners often require significant labeled data in order to evaluate the impact of prompt modifications. Our work asks whether it is possible to improve prompt-based learning _without_ additional labeled data. We approach this problem by attempting to modify the predictions of a prompt, rather than the prompt itself. Our intuition is that accurate predictions should also be consistent: samples which are similar under some feature representation should receive the same prompt prediction. We propose Embroid, a method which computes multiple representations of a dataset under different embedding functions, and uses the consistency between the LM predictions for neighboring samples to identify mispredictions. Embroid then uses these neighborhoods to create additional predictions for each sample, and combines these predictions with a simple latent variable graphical model in order to generate a final corrected prediction. In addition to providing a theoretical analysis of Embroid, we conduct a rigorous empirical evaluation across six different LMs and up to 95 different tasks. We find that (1) Embroid substantially improves performance over original prompts (e.g., by an average of 7.3 points on GPT-JT), (2) also realizes improvements for more sophisticated prompting strategies (e.g., chain-of-thought), and (3) can be specialized to domains like law through the embedding functions.

Large language models (LLMs) such as ChatGPT have seen widespread adoption due to their ability to follow user instructions well. Developing these LLMs involves a complex yet poorly understood workflow requiring training with human feedback. Replicating and understanding this instruction-following process faces three major challenges: the high cost of data collection, the lack of trustworthy evaluation, and the absence of reference method implementations. We address these bottlenecks with AlpacaFarm, a simulator that enables research and development for learning from feedback at a low cost. First, we design LLM based simulator for human feedback that is 45x cheaper than crowdworkers and displays high agreement with humans. Second, we identify an evaluation dataset representative of real-world instructions and propose an automatic evaluation procedure. Third, we contribute reference implementations for several methods (PPO, best-of-n, expert iteration, among others) that learn from pairwise feedback. Finally, as an end-to-end validation of AlpacaFarm, we train and evaluate eleven models on 10k pairs of human feedback and show that rankings of models trained in AlpacaFarm match rankings of models trained on human data. As a demonstration of the research possible in AlpacaFarm, we find that methods that use a reward model can substantially improve over supervised fine-tuning and that our reference PPO implementation leads to a +10% win-rate improvement against Davinci003.

Certifiable, adaptive uncertainty estimates for unknown quantities are an essential ingredient of sequential decision-making algorithms. Standard approaches rely on problem-dependent concentration results and are limited to a specific combination of parameterization, noise family, and estimator. In this paper, we revisit the likelihood-based inference principle and propose to use \emph{likelihood ratios} to construct \emph{any-time valid} confidence sequences without requiring specialized treatment in each application scenario. Our method is especially suitable for problems with well-specified likelihoods, and the resulting sets always maintain the prescribed coverage in a model-agnostic manner. The size of the sets depends on a choice of estimator sequence in the likelihood ratio. We discuss how to provably choose the best sequence of estimators and shed light on connections to online convex optimization with algorithms such as Follow-the-Regularized-Leader. To counteract the initially large bias of the estimators, we propose a reweighting scheme that also opens up deployment in non-parametric settings such as RKHS function classes. We provide a \emph{non-asymptotic} analysis of the likelihood ratio confidence sets size for generalized linear models, using insights from convex duality and online learning. We showcase the practical strength of our method on generalized linear bandit problems, survival analysis, and bandits with various additive noise distributions.

Large-scale pre-trained models have achieved remarkable success in many applications, but how to leverage them to improve the prediction reliability of downstream models is undesirably under-explored. Moreover, modern neural networks have been found to be poorly calibrated and make overconfident predictions regardless of inherent sample difficulty and data uncertainty. To address this issue, we propose to utilize large-scale pre-trained models to guide downstream model training with sample difficulty-aware entropy regularization. Pre-trained models that have been exposed to large-scale datasets and do not overfit the downstream training classes enable us to measure each training sample’s difficulty via feature-space Gaussian modeling and relative Mahalanobis distance computation. Importantly, by adaptively penalizing overconfident prediction based on the sample difficulty, we simultaneously improve accuracy and uncertainty calibration across challenging benchmarks (e.g., +0.55% ACC and −3.7% ECE on ImageNet1k using ResNet34), consistently surpassing competitive baselines for reliable prediction. The improved uncertainty estimate further improves selective classification (abstaining from erroneous predictions) and out-of-distribution detection.

Structured state space sequence (S4) models have recently achieved state-of-the-art performance on long-range sequence modeling tasks. These models also have fast inference speeds and parallelisable training, making them potentially useful in many reinforcement learning settings. We propose a modification to a variant of S4 that enables us to initialise and reset the hidden state in parallel, allowing us to tackle reinforcement learning tasks. We show that our modified architecture runs asymptotically faster than Transformers in sequence length and performs better than RNN's on a simple memory-based task. We evaluate our modified architecture on a set of partially-observable environments and find that, in practice, our model outperforms RNN's while also running over five times faster. Then, by leveraging the model’s ability to handle long-range sequences, we achieve strong performance on a challenging meta-learning task in which the agent is given a randomly-sampled continuous control environment, combined with a randomly-sampled linear projection of the environment's observations and actions. Furthermore, we show the resulting model can adapt to out-of-distribution held-out tasks. Overall, the results presented in this paper show that structured state space models are fast and performant for in-context reinforcement learning tasks. We provide code at https://github.com/luchris429/s5rl.

Concept Bottleneck Models (CBMs) tackle the opacity of neural architectures by constructing and explaining their predictions using a set of high-level concepts. A special property of these models is that they permit concept interventions, wherein users can correct mispredicted concepts and thus improve the model's performance. Recent work, however, has shown that intervention efficacy can be highly dependent on the order in which concepts are intervened on and on the model's architecture and training hyperparameters. We argue that this is rooted in a CBM's lack of train-time incentives for the model to be appropriately receptive to concept interventions. To address this, we propose Intervention-aware Concept Embedding models (IntCEMs), a novel CBM-based architecture and training paradigm that improves a model's receptiveness to test-time interventions. Our model learns a concept intervention policy in an end-to-end fashion from where it can sample meaningful intervention trajectories at train-time. This conditions IntCEMs to effectively select and receive concept interventions when deployed at test-time. Our experiments show that IntCEMs significantly outperform state-of-the-art concept-interpretable models when provided with test-time concept interventions, demonstrating the effectiveness of our approach.

Instruction tuning is an effective technique to align large language models (LLMs) with human intent. In this work, we investigate how an adversary can exploit instruction tuning by injecting specific instruction-following examples into the training data that intentionally changes the model's behavior. For example, an adversary can achieve content injection by injecting training examples that mention target content and eliciting such behavior from downstream models. To achieve this goal, we propose \textit{AutoPoison}, an automated data poisoning pipeline. It naturally and coherently incorporates versatile attack goals into poisoned data with the help of an oracle LLM. We showcase two example attacks: content injection and over-refusal attacks, each aiming to induce a specific exploitable behavior. We quantify and benchmark the strength and the stealthiness of our data poisoning scheme. Our results show that AutoPoison allows an adversary to change a model's behavior by poisoning only a small fraction of data while maintaining a high level of stealthiness in the poisoned examples. We hope our work sheds light on how data quality affects the behavior of instruction-tuned models and raises awareness of the importance of data quality for responsible deployments of LLMs.