Owkin releases GrAIdient: open source framework for interpretable deep learning on Mac GPUs

By exposing the layers of a neural network to developers, GrAIdient provides a unique way to design deep learning models with greater understanding, control and reproducibility – helping researchers to transition away from black box models. GrAIdient is currently used by Owkin to develop AI models to improve the treatment and diagnosis of diseases.

Previously, training or running models on Macs was difficult, with engineers instead typically relying on cloud solutions. With the arrival of the powerful new Mac M1 processor chip last year, the potential for Macs to be used to train and run models has increased.

GrAIdient was designed with speed, interpretability and reproducibility in mind, helping engineers to capture and understand the insights developed by models. It gives machine learning engineers and data scientists direct access to the layers of a neural network and to the backpropagation implementation, the foundation of the learning process of deep learning models. This access ensures model reproducibility by avoiding ‘under the hood’ assumptions that are hidden from engineers.

The framework provides a foundation for making ML models more interpretable, a crucial consideration for the medical field, in which personalized treatment relies upon a deep understanding of the mechanisms of disease. Through GrAIdient, model interpretability is pursued through a technique called maximal activation, which consists of computing the input that minimizes or maximizes the output. This technique reverses the standard way ML models operate, forcing them to express the typical inputs that lead to the output, ultimately providing more context to the correlation between inputs and outputs, and thus making the model more explainable.

The launch of GrAIdient follows the launch of Owkin’s open science push in November, through which Substra, the influential AI software behind the pharmaceutical industry’s largest ever collaborative AI project, and two further AI innovations were open sourced.

Lionel Guillou, VP Technology Development and Data for Diagnostics at Owkin, said:

The open source release of GrAIdient is an important step in transitioning from black box to white box AI. Our goal is to facilitate the implementation of model interpretability techniques, as ensuring that users can understand and interpret the outputs of machine learning models is crucial. This is especially true in medicine, in which medical professionals must understand and trust the results of models in order to confidently make important treatment decisions.

‍

View GrAIdient on GitHub