MIT researchers developed a method that helps a user to better understand a machine-learning model’s reasoning, and how that reasoning compares to that of a human. The technique enables a user to rapidly identify and analyze patterns in a model’s behavior.
From jail house calls to body-camera footage, voicemail to video depositions, modern technologies have introduced a new type of evidence to legal professionals: unstructured data..
MIT researchers created a technique that can automatically describe the roles of individual neurons in a neural network with natural language, helping machine learning practitioners better understand how their model will behave in the real world.
Feature-attribution methods are used to determine if a neural network is working correctly when completing a task like image classification. MIT researchers developed a way to evaluate whether these feature-attribution methods are correctly identifying the features of an image that are important to a neural network’s prediction.