More Like this

1. Test Accuracy vs. Generalization Gap: Model Selection in NLP without Accessing Training or Testing Data

   https://doi.org/10.1145/3580305.3599518  

   Yang, Yaoqing; Theisen, Ryan; Hodgkinson, Liam; Gonzalez, Joseph E; Ramchandran, Kannan; Martin, Charles H; Mahoney, Michael W (August 2023, ACM)

   Selecting suitable architecture parameters and training hyperparameters is essential for enhancing machine learning (ML) model performance. Several recent empirical studies conduct large-scale correlational analysis on neural networks (NNs) to search for effective generalization metrics that can guide this type of model selection. Effective metrics are typically expected to correlate strongly with test performance. In this paper, we expand on prior analyses by examining generalization-metric-based model selection with the following objectives: (i) focusing on natural language processing (NLP) tasks, as prior work primarily concentrates on computer vision (CV) tasks; (ii) considering metrics that directly predict test error instead of the generalization gap; (iii) exploring metrics that do not need access to data to compute. From these objectives, we are able to provide the first model selection results on large pretrained Transformers from Huggingface using generalization metrics. Our analyses consider (I) hundreds of Transformers trained in different settings, in which we systematically vary the amount of data, the model size and the optimization hyperparameters, (II) a total of 51 pretrained Transformers from eight families of Huggingface NLP models, including GPT2, BERT, etc., and (III) a total of 28 existing and novel generalization metrics. Despite their niche status, we find that metrics derived from the heavy-tail (HT) perspective are particularly useful in NLP tasks, exhibiting stronger correlations than other, more popular metrics. To further examine these metrics, we extend prior formulations relying on power law (PL) spectral distributions to exponential (EXP) and exponentially-truncated power law (E-TPL) families. 

   more » « less
2. Weak-to-Strong Generalization Through the Data-Centric Lens

   Shin, Changho; Cooper, John; Sala, Frederic (April 2025, ICLR)

   The weak-to-strong generalization phenomenon is the driver for important machine learning applications including highly data-efficient learning and, most recently, performing superalignment. While decades of research have resulted in numerous algorithms that produce strong empirical performance, understanding what aspects of data enable weak-to-strong generalization has been understudied. We propose a simple data-centric mechanism that characterizes weak-to-strong generalization: the overlap density. Intuitively, generalization tracks the number of points that contain overlaps, i.e., both easy patterns (learnable by a weak model) and challenging patterns (only learnable by a stronger model), as with such points, weak predictions can be used to learn challenging patterns by stronger models. We provide a practical overlap detection algorithm to find such points in datasets and leverage them to learn, among multiple sources of data, which to query when seeking to maximize overlap density and thereby enhance weak-to-strong generalization. We present a theoretical result showing that the generalization benefit is a function of the overlap density and a regret bound for our data selection algorithm. Empirically, we validate the mechanism and the overlap detection algorithm on a wide array of settings. 

   more » « less
3. Weak-to-strong generalization through the data-centric lens

   Shin, Changho; Cooper, John; Sala, Frederic (April 2025, ICLR)

   The weak-to-strong generalization phenomenon is the driver for important machine learning applications including highly data-efficient learning and, most recently, performing superalignment. While decades of research have resulted in numerous algorithms that produce strong empirical performance, understanding what aspects of data enable weak-to-strong generalization has been understudied. We propose a simple data-centric mechanism that characterizes weak-to-strong generalization: the overlap density. Intuitively, generalization tracks the number of points that contain overlaps, i.e., both easy patterns (learnable by a weak model) and challenging patterns (only learnable by a stronger model), as with such points, weak predictions can be used to learn challenging patterns by stronger models. We provide a practical overlap detection algorithm to find such points in datasets and leverage them to learn, among multiple sources of data, which to query when seeking to maximize overlap density and thereby enhance weak-to-strong generalization. We present a theoretical result showing that the generalization benefit is a function of the overlap density and a regret bound for our data selection algorithm. Empirically, we validate the mechanism and the overlap detection algorithm on a wide array of settings. 

   more » « less
4. Combining Regional Habitat Selection Models for Large-Scale Prediction: Circumpolar Habitat Selection of Southern Ocean Humpback Whales

   https://doi.org/10.3390/rs13112074  

   Reisinger, Ryan R.; Friedlaender, Ari S.; Zerbini, Alexandre N.; Palacios, Daniel M.; Andrews-Goff, Virginia; Dalla Rosa, Luciano; Double, Mike; Findlay, Ken; Garrigue, Claire; How, Jason; et al (June 2021, Remote Sensing)

   Machine learning algorithms are often used to model and predict animal habitat selection—the relationships between animal occurrences and habitat characteristics. For broadly distributed species, habitat selection often varies among populations and regions; thus, it would seem preferable to fit region- or population-specific models of habitat selection for more accurate inference and prediction, rather than fitting large-scale models using pooled data. However, where the aim is to make range-wide predictions, including areas for which there are no existing data or models of habitat selection, how can regional models best be combined? We propose that ensemble approaches commonly used to combine different algorithms for a single region can be reframed, treating regional habitat selection models as the candidate models. By doing so, we can incorporate regional variation when fitting predictive models of animal habitat selection across large ranges. We test this approach using satellite telemetry data from 168 humpback whales across five geographic regions in the Southern Ocean. Using random forests, we fitted a large-scale model relating humpback whale locations, versus background locations, to 10 environmental covariates, and made a circumpolar prediction of humpback whale habitat selection. We also fitted five regional models, the predictions of which we used as input features for four ensemble approaches: an unweighted ensemble, an ensemble weighted by environmental similarity in each cell, stacked generalization, and a hybrid approach wherein the environmental covariates and regional predictions were used as input features in a new model. We tested the predictive performance of these approaches on an independent validation dataset of humpback whale sightings and whaling catches. These multiregional ensemble approaches resulted in models with higher predictive performance than the circumpolar naive model. These approaches can be used to incorporate regional variation in animal habitat selection when fitting range-wide predictive models using machine learning algorithms. This can yield more accurate predictions across regions or populations of animals that may show variation in habitat selection. 

   more » « less
5. Oracle Inequalities for Model Selection in Offline Reinforcement Learning

   Lee, J.; Tucker, G.; Nachum, O.; Dai, B.; Brunskill, E. (January 2022, Advances in neural information processing systems)

   In offline reinforcement learning (RL), a learner leverages prior logged data to learn a good policy without interacting with the environment. A major challenge in applying such methods in practice is the lack of both theoretically principled and practical tools for model selection and evaluation. To address this, we study the problem of model selection in offline RL with value function approximation. The learner is given a nested sequence of model classes to minimize squared Bellman error and must select among these to achieve a balance between approximation and estimation error of the classes. We propose the first model selection algorithm for offline RL that achieves minimax rate-optimal oracle inequalities up to logarithmic factors. The algorithm, MODBE, takes as input a collection of candidate model classes and a generic base offline RL algorithm. By successively eliminating model classes using a novel one-sided generalization test, MODBE returns a policy with regret scaling with the complexity of the minimally complete model class. In addition to its theoretical guarantees, it is conceptually simple and computationally efficient, amounting to solving a series of square loss regression problems and then comparing relative square loss between classes. We conclude with several numerical simulations showing it is capable of reliably selecting a good model class. 

   more » « less