In this paper we motivate the causal mechanisms behind sample selection induced collider bias (selection collider bias) that can cause Large Language Models (LLMs) to learn unconditional dependence between entities that are unconditionally independent in the real world. We show that selection collider bias can become amplified in underspecified learning tasks, and although difficult to overcome, we describe a method to exploit the resulting spurious correlations for determination of when a model may be uncertain about its prediction. We demonstrate an uncertainty metric that matches human uncertainty in tasks with gender pronoun underspecification on an extended version of the Winogender Schemas evaluation set, and we provide an online demo where users can apply our uncertainty metric to their own texts and models.
相關內容
Mixed-effect models are flexible tools for researchers in a myriad of medical fields, but that flexibility comes at the cost of complexity and if users are not careful in how their model is specified, they could be making faulty inferences from their data. We argue that there is significant confusion around appropriate random effects to be included in a model given the study design, with researchers generally being better at specifying the fixed effects of a model which map onto to their research hypotheses. To that end, we present an instructive framework for evaluating the random effects of a model in three different situations: (1) longitudinal designs; (2) factorial repeated measures; and (3) when dealing with multiple sources of variance. We provide worked examples with open-access code and data in an online repository. This framework will be helpful for students and researchers who are new to mixed effect models, and to reviewers who may have to evaluate a novel model as part of their review.
Meta-learning aims to extract useful inductive biases from a set of related datasets. In Bayesian meta-learning, this is typically achieved by constructing a prior distribution over neural network parameters. However, specifying families of computationally viable prior distributions over the high-dimensional neural network parameters is difficult. As a result, existing approaches resort to meta-learning restrictive diagonal Gaussian priors, severely limiting their expressiveness and performance. To circumvent these issues, we approach meta-learning through the lens of functional Bayesian neural network inference, which views the prior as a stochastic process and performs inference in the function space. Specifically, we view the meta-training tasks as samples from the data-generating process and formalize meta-learning as empirically estimating the law of this stochastic process. Our approach can seamlessly acquire and represent complex prior knowledge by meta-learning the score function of the data-generating process marginals instead of parameter space priors. In a comprehensive benchmark, we demonstrate that our method achieves state-of-the-art performance in terms of predictive accuracy and substantial improvements in the quality of uncertainty estimates.
We advance a novel computational model of multi-agent, cooperative joint actions that is grounded in the cognitive framework of active inference. The model assumes that to solve a joint task, such as pressing together a red or blue button, two (or more) agents engage in a process of interactive inference. Each agent maintains probabilistic beliefs about the goal of the joint task (e.g., should we press the red or blue button?) and updates them by observing the other agent's movements, while in turn selecting movements that make his own intentions legible and easy to infer by the other agent (i.e., sensorimotor communication). Over time, the interactive inference aligns both the beliefs and the behavioral strategies of the agents, hence ensuring the success of the joint action. We exemplify the functioning of the model in two simulations. The first simulation illustrates a ''leaderless'' joint action. It shows that when two agents lack a strong preference about their joint task goal, they jointly infer it by observing each other's movements. In turn, this helps the interactive alignment of their beliefs and behavioral strategies. The second simulation illustrates a "leader-follower" joint action. It shows that when one agent ("leader") knows the true joint goal, it uses sensorimotor communication to help the other agent ("follower") infer it, even if doing this requires selecting a more costly individual plan. These simulations illustrate that interactive inference supports successful multi-agent joint actions and reproduces key cognitive and behavioral dynamics of "leaderless" and "leader-follower" joint actions observed in human-human experiments. In sum, interactive inference provides a cognitively inspired, formal framework to realize cooperative joint actions and consensus in multi-agent systems.
To build open-domain chatbots that are able to use diverse communicative skills, we propose a novel framework BotsTalk, where multiple agents grounded to the specific target skills participate in a conversation to automatically annotate multi-skill dialogues. We further present Blended Skill BotsTalk (BSBT), a large-scale multi-skill dialogue dataset comprising 300K conversations. Through extensive experiments, we demonstrate that our dataset can be effective for multi-skill dialogue systems which require an understanding of skill blending as well as skill grounding. Our code and data are available at //github.com/convei-lab/BotsTalk.
Regularized regression models are well studied and, under appropriate conditions, offer fast and statistically interpretable results. However, large data in many applications are heterogeneous in the sense of harboring distributional differences between latent groups. Then, the assumption that the conditional distribution of response Y given features X is the same for all samples may not hold. Furthermore, in scientific applications, the covariance structure of the features may contain important signals and its learning is also affected by latent group structure. We propose a class of mixture models for paired data (X, Y) that couples together the distribution of X (using sparse graphical models) and the conditional Y | X (using sparse regression models). The regression and graphical models are specific to the latent groups and model parameters are estimated jointly (hence the name "regularized joint mixtures"). This allows signals in either or both of the feature distribution and regression model to inform learning of latent structure and provides automatic control of confounding by such structure. Estimation is handled via an expectation-maximization algorithm, whose convergence is established theoretically. We illustrate the key ideas via empirical examples. An R package is available at //github.com/k-perrakis/regjmix.
Protecting large language models from privacy leakage is becoming increasingly crucial with their wide adoption in real-world products. Yet applying differential privacy (DP), a canonical notion with provable privacy guarantees for machine learning models, to those models remains challenging due to the trade-off between model utility and privacy loss. Utilizing the fact that sensitive information in language data tends to be sparse, Shi et al. (2021) formalized a DP notion extension called Selective Differential Privacy (SDP) to protect only the sensitive tokens defined by a policy function. However, their algorithm only works for RNN-based models. In this paper, we develop a novel framework, Just Fine-tune Twice (JFT), that achieves SDP for state-of-the-art large transformer-based models. Our method is easy to implement: it first fine-tunes the model with redacted in-domain data, and then fine-tunes it again with the original in-domain data using a private training mechanism. Furthermore, we study the scenario of imperfect implementation of policy functions that misses sensitive tokens and develop systematic methods to handle it. Experiments show that our method achieves strong utility compared to previous baselines. We also analyze the SDP privacy guarantee empirically with the canary insertion attack.
Understanding the impact of the most effective policies or treatments on a response variable of interest is desirable in many empirical works in economics, statistics and other disciplines. Due to the widespread winner's curse phenomenon, conventional statistical inference assuming that the top policies are chosen independent of the random sample may lead to overly optimistic evaluations of the best policies. In recent years, given the increased availability of large datasets, such an issue can be further complicated when researchers include many covariates to estimate the policy or treatment effects in an attempt to control for potential confounders. In this manuscript, to simultaneously address the above-mentioned issues, we propose a resampling-based procedure that not only lifts the winner's curse in evaluating the best policies observed in a random sample, but also is robust to the presence of many covariates. The proposed inference procedure yields accurate point estimates and valid frequentist confidence intervals that achieve the exact nominal level as the sample size goes to infinity for multiple best policy effect sizes. We illustrate the finite-sample performance of our approach through Monte Carlo experiments and two empirical studies, evaluating the most effective policies in charitable giving and the most beneficial group of workers in the National Supported Work program.
A fundamental task in science is to design experiments that yield valuable insights about the system under study. Mathematically, these insights can be represented as a utility or risk function that shapes the value of conducting each experiment. We present PDBAL, a targeted active learning method that adaptively designs experiments to maximize scientific utility. PDBAL takes a user-specified risk function and combines it with a probabilistic model of the experimental outcomes to choose designs that rapidly converge on a high-utility model. We prove theoretical bounds on the label complexity of PDBAL and provide fast closed-form solutions for designing experiments with common exponential family likelihoods. In simulation studies, PDBAL consistently outperforms standard untargeted approaches that focus on maximizing expected information gain over the design space. Finally, we demonstrate the scientific potential of PDBAL through a study on a large cancer drug screen dataset where PDBAL quickly recovers the most efficacious drugs with a small fraction of the total number of experiments.
Building scalable models to learn from diverse, multimodal data remains an open challenge. For vision-language data, the dominant approaches are based on contrastive learning objectives that train a separate encoder for each modality. While effective, contrastive learning approaches introduce sampling bias depending on the data augmentations used, which can degrade performance on downstream tasks. Moreover, these methods are limited to paired image-text data, and cannot leverage widely-available unpaired data. In this paper, we investigate whether a large multimodal model trained purely via masked token prediction, without using modality-specific encoders or contrastive learning, can learn transferable representations for downstream tasks. We propose a simple and scalable network architecture, the Multimodal Masked Autoencoder (M3AE), which learns a unified encoder for both vision and language data via masked token prediction. We provide an empirical study of M3AE trained on a large-scale image-text dataset, and find that M3AE is able to learn generalizable representations that transfer well to downstream tasks. Surprisingly, we find that M3AE benefits from a higher text mask ratio (50-90%), in contrast to BERT whose standard masking ratio is 15%, due to the joint training of two data modalities. We also provide qualitative analysis showing that the learned representation incorporates meaningful information from both image and language. Lastly, we demonstrate the scalability of M3AE with larger model size and training time, and its flexibility to train on both paired image-text data as well as unpaired data.
Since hardware resources are limited, the objective of training deep learning models is typically to maximize accuracy subject to the time and memory constraints of training and inference. We study the impact of model size in this setting, focusing on Transformer models for NLP tasks that are limited by compute: self-supervised pretraining and high-resource machine translation. We first show that even though smaller Transformer models execute faster per iteration, wider and deeper models converge in significantly fewer steps. Moreover, this acceleration in convergence typically outpaces the additional computational overhead of using larger models. Therefore, the most compute-efficient training strategy is to counterintuitively train extremely large models but stop after a small number of iterations. This leads to an apparent trade-off between the training efficiency of large Transformer models and the inference efficiency of small Transformer models. However, we show that large models are more robust to compression techniques such as quantization and pruning than small models. Consequently, one can get the best of both worlds: heavily compressed, large models achieve higher accuracy than lightly compressed, small models.
小貼士
登錄享
相關主題
注冊地址： 北京市海淀區羊坊店路18號2幢3層301-191