Introduction to Machine Learning on Biomedical Data
The course will begin with a very brief overview of the mathematical foundations of ML, specifically linear regression, logistic regression, and multilayer perceptrons. Applying these models to the well-studied dataset of hand written digits, MNIST, we’ll gain first-hand experience with model selection, training and validation. We will then introduce several abstractions from the ML4CVD codebase (TensorMaps, Model Factories, and Tensorization) which accelerate and simplify the process of preparing data for ML, building models, and evaluating them. Then we will examine several real world biomedical datasets of various sizes, structures and quality. Specifically, from the Allen Brain Atlas we will load high-dimensional brain MRI data linked with gene expression microarrays, from Qure.ai we will use the CQ500 dataset of 500 CT scans containing 193K slices linked with medical reports from 3 senior radiologists, and lastly from the Erowid website raw natural language text testimonials of drug experiences linked with basic demographics. These data will present many new challenges, such as noisy labels, small sample sizes, confounding by indication, and batch effects. These issues are less prevalent in typical ML data, like MNIST or imagenet, but very common with biomedical data. After visually and statistically exploring the datasets, we will setup several ML problems using them as raw data. Lastly, we will explore model interpretation by visualizing saliency maps, class activation maps, and training adversarial examples. The course will conclude with a general discussion on framing biological questions as machine learning problems and the opportunity for participants to brainstorm how ML might apply to their own datasets.
February 14: Registration closes at 5:00PM
February 19: Attendees notified of acceptance or wait list status
Friday February 28, 2020
10:00 AM - 1:00 PM
75 Ames Street
Cambridge, MA 02142
Please note: When registering for this workshop your initial status is considered PENDING until notified of acceptance.
Please use your Broad or partner institute email address.
Please select the statement that best describes you.
I am a Broad Employee
I am a Broad Affiliate
I am an Associate Member
I am a member of an Associate Member's lab
Never submit passwords through Google Forms.
This form was created inside of Broad Institute of MIT and Harvard.