Program Format

The 5-day class will provide lectures on the mathematics and statistics at the heart of machine learning, plus hands-on training on implementing machine learning tools with the PyTorch software platform, and case studies of the methods applied to specific application areas.

Each day of the MLSS will be arranged as follows (Eastern Time):

• 9:00-10:15am     Lecture 1: Mathematically-light introduction to the focus of the day
• 10:45am-noon     Lecture 2: Mathematically rigorous discussion of the focus of the day
• Afternoons beginning at 1:00pm     Coding sessions and case studies

At the end of the MLSS each student will have a deeper understanding of the fundamental concepts of machine learning and applied computer vision, including context for the rapidly evolving field of artificial intelligence. For those students with sufficient mathematical background, the underlying methodology of machine learning will also be learned. Each student should be able to utilize PyTorch to implement the latest machine learning methods for analysis of images, video, and natural language (text).

Program Details: Location, Registration and Cost

Students (with a valid ID, at Duke or other universities) will pay a course fee of $150; the fee for non-students is $400, payable through the registration site. All fees are non-refundable. Once we reach maximum registration, we will maintain a waitlist, and will contact those on the waitlist as spots become available. We also have a small number of scholarships available for those who would be otherwise unable to join.

There will be no difference in cost for participants who attend in person vs. participants who attend virtually. This will allow maximum flexibility and personal choice for attendance options.

Each participant will receive a personal link for the virtual webinars, which will be held live and provide opportunities for questions and engagement with each lecturer. We strongly encourage live participation, but every participant will also have access to the video recordings to use for their personal reference.

Relevance and Context

Machine learning is a field characterized by development of algorithms that are implemented in software and run on a machine (e.g., computer, mobile device, etc.). Each such algorithm is characterized by a set of parameters, and particular parameter settings yield associated algorithm characteristics. The algorithms have the capacity to learn, based on observed data. By “learn” it is meant that the algorithm can infer (or learn) which algorithm parameter settings are best matched to the data of interest. After algorithm parameters are so learned, the associated model ideally captures the underlying characteristics of the data. The algorithm, with learned parameters, may subsequently be applied to new data, with the goal of making predictions or learning insights. Machine learning methodology is primarily concerned with designing appropriate models/algorithms for datasets and problems of interest, plus the capacity to learn the model parameters given data (with challenges manifested when that data is of a massive scale).

In the context of prediction, one may be interested in developing algorithms that are capable of automatically classifying and interpreting imaging data; for instance, in a healthcare setting, to improve clinical care. In this case, the healthcare data may be radiological images (e.g., x-rays, ultrasound videos, computed tomography volumes, etc.) and/or a history of patient care. In healthcare, the goal is to use machine learning to make improved diagnoses, interpretation (e.g., location of abnormalities within an image) and recommendations for care. Similar concepts are of interest in business, where one may be interested in tailoring advertising and products to individuals or improving image search. In education, machine learning may be used to tailor educational material to the level and interests of each student. Machine learning is increasingly making an impact in almost all areas of personal and professional life.

Recently, with increasing access to massive imaging datasets (e.g., ImageNet), and to significant advances in computing resources, the quality of machine learning performance (e.g., prediction accuracy) has improved markedly. Further, over the last seven years, significant advances have been made in a subfield of machine learning called “deep learning” that have completely changed the landscape and boundaries of computer vision to the point in which computer vision models, in multiple scenarios, have surpassed human ability in image classification and object detection tasks, as well as being able to generate images of such quality that are indistinguishable from real images by most humans.

This class will focus on the areas of machine learning that have made the biggest advances in utility over the last several years, including deep learning. The class will concentrate on methods that allow machine-learning algorithms to train effectively on massive datasets, i.e., “big data”.