arXiv subjects

Nati Srebro

Publications and source records attributed to Nati Srebro.

3 recordsLinked to original sources

arXiv2026

Learning through Internalization

We study internalization processes, by which neural-network-based systems absorb an explicit computational procedure into their own weights, and how they facilitate learning. We investigate how transformers internalize the simulation of semiautomata by internalizing chain-of-thought (CoT) tokens, which classes of semiautomata are harder to internalize, and expose the flip side of internalization, that is, a progressive degradation of out-of-distribution performance. We then provide the first provable analysis of successful internalization: for the task of learning parities, we show that a simplified one-layer transformer provably first learns the target with explicit CoT supervision and then internalizes the autoregressive generation as CoT tokens are progressively removed, learning to directly compute the parity. This task is computationally hard to learn from data without CoT supervision. Finally, we discuss how learning through internalization relates to the \textit{Positive Distribution Shift} phenomenon recently introduced by~\citet{Med+26}.

cs.LG↗

arXiv2022

How catastrophic can catastrophic forgetting be in linear regression?

To better understand catastrophic forgetting, we study fitting an overparameterized linear model to a sequence of tasks with different input distributions. We analyze how much the model forgets the true labels of earlier tasks after training on subsequent tasks, obtaining exact expressions and bounds. We establish connections between continual learning in the linear setting and two other research areas: alternating projections and the Kaczmarz method. In specific settings, we highlight differences between forgetting and convergence to the offline solution as studied in those areas. In particular, when T tasks in d dimensions are presented cyclically for k iterations, we prove an upper bound of T^2 * min{1/sqrt(k), d/k} on the forgetting. This stands in contrast to the convergence to the offline solution, which can be arbitrarily slow according to existing alternating projection results. We further show that the T^2 factor can be lifted when tasks are presented in a random ordering.

cs.LG↗

arXiv2018

Simple Surveys: Response Retrieval Inspired by Recommendation Systems

In the last decade, the use of simple rating and comparison surveys has proliferated on social and digital media platforms to fuel recommendations. These simple surveys and their extrapolation with machine learning algorithms shed light on user preferences over large and growing pools of items, such as movies, songs and ads. Social scientists have a long history of measuring perceptions, preferences and opinions, often over smaller, discrete item sets with exhaustive rating or ranking surveys. This paper introduces simple surveys for social science application. We ran experiments to compare the predictive accuracy of both individual and aggregate comparative assessments using four types of simple surveys: pairwise comparisons and ratings on 2, 5 and continuous point scales in three distinct contexts: perceived Safety of Google Streetview Images, Likeability of Artwork, and Hilarity of Animal GIFs. Across contexts, we find that continuous scale ratings best predict individual assessments but consume the most time and cognitive effort. Binary choice surveys are quick and perform best to predict aggregate assessments, useful for collective decision tasks, but poorly predict personalized preferences, for which they are currently used by Netflix to recommend movies. Pairwise comparisons, by contrast, perform well to predict personal assessments, but poorly predict aggregate assessments despite being widely used to crowdsource ideas and collective preferences. We demonstrate how findings from these surveys can be visualized in a low-dimensional space that reveals distinct respondent interpretations of questions asked in each context. We conclude by reflecting on differences between sparse, incomplete simple surveys and their traditional survey counterparts in terms of efficiency, information elicited and settings in which knowing less about more may be critical for social science.

cs.SI↗