Knowing when to shut up - Methods for estimating confidence in DNN predictions

Hey guys, We are excited to announce our March 2019 meetup! This time we'll have a lecture by Ran Shadmi from Nucleai. Ran will share his experience in getting true estimates of model-predictions confidence level.

Abstract:
As machine learning moves from the lab into the real world, reliability is often of paramount importance. One of the clearest examples of critical application is computer-aided diagnostics such as cancer detection in biopsies.

Modern neural networks are very powerful predictive models, but they are often incapable of recognizing when their predictions may be wrong (also known as reject option).

Accurately estimating the level of confidence in the model output is of the utmost importance in clinical settings.

Since we at Nucleai deal with systems that can impact clinical outcome on an everyday basis, I’ve chosen to cover a few related works on this subject, which can also be highly relevant for many other real-life applications of deep-learning models.

About Ran Shadmi:
Ran Shadmi is an algorithm developer and software engineer for image processing, computer vision, machine learning and deep learning applications.

Ran holds a bachelor degree of communication systems engineering from Ben Gurion University and a biomedical engineering master's degree from Tel Aviv University, specializing in medical image processing.

In recent years Ran has focused on applying his technical skills to the development of medical applications. In Zebra Medical, Ran led the development of the Coronary Calcium Scoring algorithm (FDA approved) as well as automatic detection of abnormalities in chest X-ray images.

Ran joined Nucleai in early 2018 as its first employee and led the development of the company’s algorithmic infrastructure.

About Nucleai:
Cancer diagnostics, in many ways, has not changed much in the past century. Trained physicians inspect tissue biopsies under the microscope, using their eyes and years of experience. Even a small cluster of unnoticed cancer cells can lead to a possibly fatal misdiagnosis. In Nucleai, we build AI algorithms that inspect tissue biopsies, reduce diagnostic errors and ultimately save lives.

Nucleai was founded in 2017 after raising 5M$ from Grove Ventures and Vertex Ventures.