This website requires JavaScript.

ModelPred: A Framework for Predicting Trained Model from Training Data

Yingyan ZengJiachen T. WangSi ChenHoang Anh JustRan JinRuoxi Jia
Nov 2021
摘要
In this work, we propose ModelPred, a framework that helps to understand theimpact of changes in training data on a trained model. This is critical forbuilding trust in various stages of a machine learning pipeline: from cleaningpoor-quality samples and tracking important ones to be collected during datapreparation, to calibrating uncertainty of model prediction, to interpretingwhy certain behaviors of a model emerge during deployment. Specifically,ModelPred learns a parameterized function that takes a dataset $S$ as the inputand predicts the model obtained by training on $S$. Our work differs from therecent work of Datamodels [1] as we aim for predicting the trained modelparameters directly instead of the trained model behaviors. We demonstrate thata neural network-based set function class is capable of learning the complexrelationships between the training data and model parameters. We introducenovel global and local regularization techniques to prevent overfitting and werigorously characterize the expressive power of neural networks (NN) inapproximating the end-to-end training process. Through extensive empiricalinvestigations, we show that ModelPred enables a variety of applications thatboost the interpretability and accountability of machine learning (ML), such asdata valuation, data selection, memorization quantification, and modelcalibration.
展开全部
图表提取

暂无人提供速读十问回答

论文十问由沈向洋博士提出,鼓励大家带着这十个问题去阅读论文,用有用的信息构建认知模型。写出自己的十问回答,还有机会在当前页面展示哦。

Q1论文试图解决什么问题?
Q2这是否是一个新的问题?
Q3这篇文章要验证一个什么科学假设?
0
被引用
笔记
问答