Jason Poulos

Long Phan, Alice Gatti, Ziwen Han et al. · amazon-science, apple-ml

Benchmarks are important tools for tracking the rapid advancements in large language model (LLM) capabilities. However, benchmarks are not keeping pace in difficulty: LLMs now achieve over 90\% accuracy on popular benchmarks like MMLU, limiting informed measurement of state-of-the-art LLM capabilities. In response, we introduce Humanity's Last Exam (HLE), a multi-modal benchmark at the frontier of human knowledge, designed to be the final closed-ended academic benchmark of its kind with broad subject coverage. HLE consists of 2,500 questions across dozens of subjects, including mathematics, humanities, and the natural sciences. HLE is developed globally by subject-matter experts and consists of multiple-choice and short-answer questions suitable for automated grading. Each question has a known solution that is unambiguous and easily verifiable, but cannot be quickly answered via internet retrieval. State-of-the-art LLMs demonstrate low accuracy and calibration on HLE, highlighting a significant gap between current LLM capabilities and the expert human frontier on closed-ended academic questions. To inform research and policymaking upon a clear understanding of model capabilities, we publicly release HLE at https://lastexam.ai.

Zhenhua Wang, Olanrewaju Akande, Jason Poulos et al.

Multiple imputation (MI) is a popular approach for dealing with missing data arising from non-response in sample surveys. Multiple imputation by chained equations (MICE) is one of the most widely used MI algorithms for multivariate data, but it lacks theoretical foundation and is computationally intensive. Recently, missing data imputation methods based on deep learning models have been developed with encouraging results in small studies. However, there has been limited research on evaluating their performance in realistic settings compared to MICE, particularly in big surveys. We conduct extensive simulation studies based on a subsample of the American Community Survey to compare the repeated sampling properties of four machine learning based MI methods: MICE with classification trees, MICE with random forests, generative adversarial imputation networks, and multiple imputation using denoising autoencoders. We find the deep learning imputation methods are superior to MICE in terms of computational time. However, with the default choice of hyperparameters in the common software packages, MICE with classification trees consistently outperforms, often by a large margin, the deep learning imputation methods in terms of bias, mean squared error, and coverage under a range of realistic settings.

David Rios Insua, Roi Naveiro, Victor Gallego et al.

Adversarial Machine Learning (AML) is emerging as a major field aimed at protecting machine learning (ML) systems against security threats: in certain scenarios there may be adversaries that actively manipulate input data to fool learning systems. This creates a new class of security vulnerabilities that ML systems may face, and a new desirable property called adversarial robustness essential to trust operations based on ML outputs. Most work in AML is built upon a game-theoretic modelling of the conflict between a learning system and an adversary, ready to manipulate input data. This assumes that each agent knows their opponent's interests and uncertainty judgments, facilitating inferences based on Nash equilibria. However, such common knowledge assumption is not realistic in the security scenarios typical of AML. After reviewing such game-theoretic approaches, we discuss the benefits that Bayesian perspectives provide when defending ML-based systems. We demonstrate how the Bayesian approach allows us to explicitly model our uncertainty about the opponent's beliefs and interests, relaxing unrealistic assumptions, and providing more robust inferences. We illustrate this approach in supervised learning settings, and identify relevant future research problems.

Jason Poulos, Rafael Valle

The paper approaches the task of handwritten text recognition (HTR) with attentional encoder-decoder networks trained on sequences of characters, rather than words. We experiment on lines of text from popular handwriting datasets and compare different activation functions for the attention mechanism used for aligning image pixels and target characters. We find that softmax attention focuses heavily on individual characters, while sigmoid attention focuses on multiple characters at each step of the decoding. When the sequence alignment is one-to-one, softmax attention is able to learn a more precise alignment at each step of the decoding, whereas the alignment generated by sigmoid attention is much less precise. When a linear function is used to obtain attention weights, the model predicts a character by looking at the entire sequence of characters and performs poorly because it lacks a precise alignment between the source and target. Future research may explore HTR in natural scene images, since the model is capable of transcribing handwritten text without the need for producing segmentations or bounding boxes of text in images.

Jason Poulos, Shuxi Zeng

This paper proposes a method for estimating the effect of a policy intervention on an outcome over time. We train recurrent neural networks (RNNs) on the history of control unit outcomes to learn a useful representation for predicting future outcomes. The learned representation of control units is then applied to the treated units for predicting counterfactual outcomes. RNNs are specifically structured to exploit temporal dependencies in panel data, and are able to learn negative and nonlinear interactions between control unit outcomes. We apply the method to the problem of estimating the long-run impact of U.S. homestead policy on public school spending.

Jason Poulos, Rafael Valle

Missing data imputation can help improve the performance of prediction models in situations where missing data hide useful information. This paper compares methods for imputing missing categorical data for supervised classification tasks. We experiment on two machine learning benchmark datasets with missing categorical data, comparing classifiers trained on non-imputed (i.e., one-hot encoded) or imputed data with different levels of additional missing-data perturbation. We show imputation methods can increase predictive accuracy in the presence of missing-data perturbation, which can actually improve prediction accuracy by regularizing the classifier. We achieve the state-of-the-art on the Adult dataset with missing-data perturbation and k-nearest-neighbors (k-NN) imputation.