Episode Summary
Executive Summary: Zach Lipton argues that 2019 marked a shift from scaling supervised deep learning toward confronting its limits: benchmark overfitting, weak out-of-distribution generalization, strategic behavior, and fairness beyond prediction. He highlights papers on benchmark re-evaluation, BERT’s continued dominance, causal and counterfactual approaches, invariant learning, and energy-based models as signs the field is broadening toward causality, robustness, and decision-aware ML.
Main Topics: 2019 as a turning point for ML: Lipton frames the year as one where the easy gains from deep learning began to taper off, pushing researchers to ask harder scientific questions beyond scaling and leaderboard gains. Benchmark generalization and holdout overfitting: Discussion of the CIFAR-10/ImageNet re-test papers that created fresh test sets to evaluate whether leaderboards were overfit and whether model rankings still held under new samples. BERT and the continuing value of scale: BERT is presented as evidence that some supervised/NLP progress still comes from large-scale pretraining and that foundation-style models have become the default base layer for downstream tasks. Causality, strategic behavior, and fairness: Lipton emphasizes that real deployments change behavior: people respond strategically to models, which alters the data-generating process and creates fairness and policy issues not captured by static prediction. Invariant risk minimization and robustness to shift: He discusses IRM as an attempt to learn representations whose optimal predictors remain stable across environments, aiming to improve out-of-domain performance. Energy-based models and unified discriminative/generative behavior: The 'classifier is secretly an energy-based model' paper is highlighted for unifying classification, sampling, calibration, out-of-distribution detection, and robustness in a single framework. Counterfactually augmented data and non-ideal fairness: Lipton describes his group’s work using human-edited counterfactual reviews to expose spurious correlations and argues fairness should be grounded in causal and philosophical accounts, not abstract idealized metrics alone.
Key Arguments: The field is moving from 'easy' supervised-learning wins to deeper questions about causality, robustness, and decision-making in changing environments. Leaderboard performance can still be meaningful, but fresh CIFAR/ImageNet tests show standard benchmarks have been partly exhausted and are brittle under small distribution shifts. BERT represents the maturation of pretraining: the main architectural ideas may be commoditizing, but the model became a standard base for almost all NLP work. Prediction-only framing is insufficient because real systems affect human behavior; once a model is deployed, agents adapt strategically and the data distribution changes. Fairness cannot be solved by a single metric or off-the-shelf classifier tweak; one needs causal, historical, and normative context to decide what justice requires. IRM and related methods matter because they seek representations that generalize across environments instead of exploiting spurious correlations. Counterfactual data augmentation helps models stop relying on superficial cues and improves both in-domain and out-of-domain performance. Energy-based interpretations of classifiers may unify discriminative accuracy with generative sampling, calibration, robustness, and shift detection. ML research is widening into economics, policy, and philosophy because deployment problems demand richer frameworks than train/test accuracy alone.
Data Points: CIFAR fresh test-set performance drop: about 7% - Lipton says models on new CIFAR data performed worse than on the original holdout set. ImageNet fresh test-set performance drop: about 8% - He cites a similar drop when models were re-evaluated on a newly curated ImageNet test set. Counterfactual data accuracy (train original, test revised): ~90% to ~55% - Training on the original sentiment dataset and evaluating on counterfactually revised data caused a large accuracy drop. Counterfactual data accuracy (train revised, test original): ~90% to ~55% - The reverse transfer also dropped sharply, showing the models were exploiting spurious cues. Joint counterfactually augmented training accuracy: ~88% on both datasets - Training on both original and revised data preserved strong performance while reducing reliance on spurious correlations. Crowd workers used for counterfactual augmentation: ~800 - Lipton notes the large manual labor involved in producing the counterfactual review dataset. Year of the IRM paper discussion: 2019 - IRM is discussed as one of the year’s important robustness/generalization contributions. BERT timing: Late 2018 / major impact in 2019 - Lipton notes BERT was first seen in late 2018 but became central in 2019.
Pivotal Quotes: "2012 and just suddenly became really easy to do research." — Zach Lipton: Describing the deep-learning boom and how it changed the pace of research. "The rocket fuel from deep learning is starting, in my eyes, I think it's starting to run out." — Zach Lipton: Summarizing his sense that pure scaling gains are slowing and the field is branching outward. "What we really wanted, or what it really takes to do interesting things in the real world, is something that other notion of generalization to do." — Zach Lipton: On the limitations of IID train/test evaluation and the need for robustness to new environments.
Implications: Listeners should expect ML to become more grounded in causality, robustness, and human behavior. For industry, benchmark wins will matter less than systems that handle distribution shift, strategic adaptation, and real-world policy constraints.