Gradient Sketches for Training Data Attribution and Studying the Loss Landscape

Random projections or sketches of gradients and Hessian vector products play an essential role in applications where one needs to store many such vectors while retaining accurate information about their relative geometry. Two important scenarios are training data attribution (tracing a model's behavior to the training data), where one needs to store a gradient for each training example, and the study of the spectrum of the Hessian (to analyze the training dynamics), where one needs to store multiple Hessian vector products. While sketches that use dense matrices are easy to implement, they are memory bound and cannot be scaled to modern neural networks. Motivated by work on the intrinsic dimension of neural networks, we propose and study a design space for scalable sketching algorithms. We demonstrate the efficacy of our approach in three applications: training data attribution, the analysis of the Hessian spectrum and the computation of the intrinsic dimension when fine-tuning pre-trained language models.

Theoretical and Practical Perspectives on what Influence Functions Do

Influence functions (IF) have been seen as a technique for explaining model predictions through the lens of the training data. Their utility is assumed to be in identifying training examples "responsible" for a prediction so that, for example, correcting a prediction is possible by intervening on those examples (removing or editing them) and retraining the model. However, recent empirical studies have shown that the existing methods of estimating IF predict the leave-one-out-and-retrain effect poorly. In order to understand the mismatch between the theoretical promise and the practical results, we analyse five assumptions made by IF methods which are problematic for modern-scale deep neural networks and which concern convexity, numeric stability, training trajectory and parameter divergence. This allows us to clarify what can be expected theoretically from IF. We show that while most assumptions can be addressed successfully, the parameter divergence poses a clear limitation on the predictive power of IF: influence fades over training time even with deterministic training. We illustrate this theoretical result with BERT and ResNet models. Another conclusion from the theoretical analysis is that IF are still useful for model debugging and correcting even though some of the assumptions made in prior work do not hold: using natural language processing and computer vision tasks, we verify that mis-predictions can be successfully corrected by taking only a few fine-tuning steps on influential examples.

Cross-Lingual Supervision improves Large Language Models Pre-training

The recent rapid progress in pre-training Large Language Models has relied on using self-supervised language modeling objectives like next token prediction or span corruption. On the other hand, Machine Translation Systems are mostly trained using cross-lingual supervision that requires aligned data between source and target languages. We demonstrate that pre-training Large Language Models on a mixture of a self-supervised Language Modeling objective and the supervised Machine Translation objective, therefore including cross-lingual parallel data during pre-training, yields models with better in-context learning abilities. As pre-training is a very resource-intensive process and a grid search on the best mixing ratio between the two objectives is prohibitively expensive, we propose a simple yet effective strategy to learn it during pre-training.

Scaling Up Influence Functions

We address efficient calculation of influence functions for tracking predictions back to the training data. We propose and analyze a new approach to speeding up the inverse Hessian calculation based on Arnoldi iteration. With this improvement, we achieve, to the best of our knowledge, the first successful implementation of influence functions that scales to full-size (language and vision) Transformer models with several hundreds of millions of parameters. We evaluate our approach on image classification and sequence-to-sequence tasks with tens to a hundred of millions of training examples. Our code will be available at https://github.com/google-research/jax-influence.

Distributed Function Minimization in Apache Spark

We report on an open-source implementation for distributed function minimization on top of Apache Spark by using gradient and quasi-Newton methods. We show-case it with an application to Optimal Transport and some scalability tests on classification and regression problems.

Learning to Transport with Neural Networks

We compare several approaches to learn an Optimal Map, represented as a neural network, between probability distributions. The approaches fall into two categories: ``Heuristics'' and approaches with a more sound mathematical justification, motivated by the dual of the Kantorovitch problem. Among the algorithms we consider a novel approach involving dynamic flows and reductions of Optimal Transport to supervised learning.

Optimality of the final model found via Stochastic Gradient Descent

We study convergence properties of Stochastic Gradient Descent (SGD) for convex objectives without assumptions on smoothness or strict convexity. We consider the question of establishing that with high probability the objective evaluated at the candidate minimizer returned by SGD is close to the minimal value of the objective. We compare this result concerning the final candidate minimzer (i.e. the final model parameters learned after all gradient steps) to the online learning techniques of [Zin03] that take a rolling average of the model parameters at the different steps of SGD.