George Alexandru Adam

George Alexandru Adam, Alexander Cui, Edwin Thomas et al.

While historical considerations surrounding text authenticity revolved primarily around plagiarism, the advent of large language models (LLMs) has introduced a new challenge: distinguishing human-authored from AI-generated text. This shift raises significant concerns, including the undermining of skill evaluations, the mass-production of low-quality content, and the proliferation of misinformation. Addressing these issues, we introduce GPTZero a state-of-the-art industrial AI detection solution, offering reliable discernment between human and LLM-generated text. Our key contributions include: introducing a hierarchical, multi-task architecture enabling a flexible taxonomy of human and AI texts, demonstrating state-of-the-art accuracy on a variety of domains with granular predictions, and achieving superior robustness to adversarial attacks and paraphrasing via multi-tiered automated red teaming. GPTZero offers accurate and explainable detection, and educates users on its responsible use, ensuring fair and transparent assessment of text.

George Adam, Benjamin Haibe-Kains, Anna Goldenberg

Deployed machine learning models should be updated to take advantage of a larger sample size to improve performance, as more data is gathered over time. Unfortunately, even when model updates improve aggregate metrics such as accuracy, they can lead to errors on samples that were correctly predicted by the previous model causing per-sample regression in performance known as predictive churn. Such prediction flips erode user trust thereby reducing the effectiveness of the human-AI team as a whole. We propose a solution called Accumulated Model Combination (AMC) based keeping the previous and current model version, and generating a meta-output using the prediction of the two models. AMC is a general technique and we propose several instances of it, each having their own advantages depending on the model and data properties. AMC requires minimal additional computation and changes to training procedures. We motivate the need for AMC by showing the difficulty of making a single model consistent with its own predictions throughout training thereby revealing an implicit stability-plasticity tradeoff when training a single model. We demonstrate the effectiveness of AMC on a variety of modalities including computer vision, text, and tabular datasets comparing against state-of-the-art churn reduction methods, and showing superior churn reduction ability compared to all existing methods while being more efficient than ensembles.

Chun-Hao Chang, George Alexandru Adam, Rich Caruana et al.

Although reinforcement learning (RL) has tremendous success in many fields, applying RL to real-world settings such as healthcare is challenging when the reward is hard to specify and no exploration is allowed. In this work, we focus on recovering clinicians' rewards in treating patients. We incorporate the what-if reasoning to explain the clinician's treatments based on their potential future outcomes. We use generalized additive models (GAMs) - a class of accurate, interpretable models - to recover the reward. In both simulation and a real-world hospital dataset, we show our model outperforms baselines. Finally, our model's explanations match several clinical guidelines when treating patients while we found the commonly-used linear model often contradicts them.

Chun-Hao Chang, George Alexandru Adam, Anna Goldenberg

Despite the success of machine learning applications in science, industry, and society in general, many approaches are known to be non-robust, often relying on spurious correlations to make predictions. Spuriousness occurs when some features correlate with labels but are not causal; relying on such features prevents models from generalizing to unseen environments where such correlations break. In this work, we focus on image classification and propose two data generation processes to reduce spuriousness. Given human annotations of the subset of the features responsible (causal) for the labels (e.g. bounding boxes), we modify this causal set to generate a surrogate image that no longer has the same label (i.e. a counterfactual image). We also alter non-causal features to generate images still recognized as the original labels, which helps to learn a model invariant to these features. In several challenging datasets, our data generations outperform state-of-the-art methods in accuracy when spurious correlations break, and increase the saliency focus on causal features providing better explanations.

George Adam, Romain Speciel

Adversarial examples, which are slightly perturbed inputs generated with the aim of fooling a neural network, are known to transfer between models; adversaries which are effective on one model will often fool another. This concept of transferability poses grave security concerns as it leads to the possibility of attacking models in a black box setting, during which the internal parameters of the target model are unknown. In this paper, we seek to analyze and minimize the transferability of adversaries between models within an ensemble. To this end, we introduce a gradient based measure of how effectively an ensemble's constituent models collaborate to reduce the space of adversarial examples targeting the ensemble itself. Furthermore, we demonstrate that this measure can be utilized during training as to increase an ensemble's robustness to adversarial examples.

George Adam, Petr Smirnov, Benjamin Haibe-Kains et al.

Deep learning algorithms have increasingly been shown to lack robustness to simple adversarial examples (AdvX). An equally troubling observation is that these adversarial examples transfer between different architectures trained on different datasets. We investigate the transferability of adversarial examples between models using the angle between the input-output Jacobians of different models. To demonstrate the relevance of this approach, we perform case studies that involve jointly training pairs of models. These case studies empirically justify the theoretical intuitions for why the angle between gradients is a fundamental quantity in AdvX transferability. Furthermore, we consider the asymmetry of AdvX transferability between two models of the same architecture and explain it in terms of differences in gradient norms between the models. Lastly, we provide a simple modification to existing training setups that reduces transferability of adversarial examples between pairs of models.

George Adam, Jonathan Lorraine

Automatic methods for generating state-of-the-art neural network architectures without human experts have generated significant attention recently. This is because of the potential to remove human experts from the design loop which can reduce costs and decrease time to model deployment. Neural architecture search (NAS) techniques have improved significantly in their computational efficiency since the original NAS was proposed. This reduction in computation is enabled via weight sharing such as in Efficient Neural Architecture Search (ENAS). However, recently a body of work confirms our discovery that ENAS does not do significantly better than random search with weight sharing, contradicting the initial claims of the authors. We provide an explanation for this phenomenon by investigating the interpretability of the ENAS controller's hidden state. We find models sampled from identical controller hidden states have no correlation with various graph similarity metrics, so no notion of structural similarity is learned. This failure mode implies the RNN controller does not condition on past architecture choices. Lastly, we propose a solution to this failure mode by forcing the controller's hidden state to encode pasts decisions by training it with a memory buffer of previously sampled architectures. Doing this improves hidden state interpretability by increasing the correlation between controller hidden states and graph similarity metrics.