Gaby Ecanow always enjoyed music, but she never imagined composing her own tracks until enrolling in MIT's 6.S191 (Introduction to Deep Learning). Within just two sessions, this MIT sophomore had crafted an authentic Irish folk melody with assistance from a recurrent neural network. She was already exploring ways to modify the algorithm to produce original Louis the Child-inspired electronic dance music.
"The results were incredible," she recalls. "The output sounded nothing like machine-generated content—it had genuine human-like creativity."
This academic year, 6.S191 commenced as usual, with students filling every seat in Stata Center's Kirsch Auditorium during the Independent Activities Period (IAP). However, the opening lecture delivered an unexpected surprise: a recorded welcome message supposedly from former President Barack Obama. The footage was soon revealed as an AI-generated creation—one of many innovative elements that Alexander Amini '17 and Ava Soleimany '16 incorporate throughout their credit-bearing course to transform complex equations and code into engaging, practical knowledge.
As hundreds of their peers observe, Amini and Soleimany seamlessly alternate at the podium. Their confidence stems from their comprehensive mastery of the material—they personally designed the curriculum and have refined it over three years of instruction. The course explores deep learning's technical foundations and societal implications through lectures and hands-on software laboratories focused on real-world applications. During the final session, students compete for awards by presenting their original research project proposals. In the weeks preceding the course, Amini and Soleimany dedicate countless hours to updating labs, refreshing lecture content, and perfecting their presentations.
As a specialized branch of machine learning, deep learning leverages extensive datasets and algorithms loosely modeled after the brain's information processing mechanisms to make predictions. This course has been instrumental in disseminating machine-learning tools throughout MIT's research laboratories—a deliberate outcome according to Amini, a graduate student in MIT's Department of Electrical Engineering and Computer Science (EECS), and Soleimany, a graduate student at both MIT and Harvard University.
Both instructors apply machine learning in their personal research—Amini in robotics engineering, and Soleimany in developing cancer diagnostic tools—and they were committed to designing a curriculum that would equip students with similar capabilities. Beyond the music-generating AI laboratory, they offer sessions on constructing facial recognition models using convolutional neural networks and creating reinforcement learning bots that play classic Atari games like Pong. Once students grasp these fundamentals, those taking the course for credit proceed to develop their own applications.
This year, 23 teams showcased their projects. Among the award recipients was Carmen Martin, a graduate student in the Harvard-MIT Program in Health Sciences and Technology (HST), who suggested employing graph convolutional networks to forecast coronavirus transmission patterns. She integrated multiple data sources: airline ticket information to track population movements, real-time infection confirmation data, and a rating system evaluating countries' pandemic preparedness and response capabilities.
"The objective is to train the model to predict case trajectories, assisting national governments and the World Health Organization in developing recommendations to limit new infections and preserve lives," she explains.
Another award winner, EECS graduate student Samuel Sledzieski, proposed constructing a model to forecast protein interactions using only amino acid sequences. Predicting protein behavior is crucial for designing drug targets and other clinical applications, and Sledzieski explored whether deep learning could accelerate the identification of viable protein pairs.
"Further development is needed, but I'm amazed by the progress achieved in just three days," he notes. "The accessible TensorFlow and Keras examples enabled me to understand how to independently construct and train these models." He intends to continue this research during his current lab rotation with Bonnie Berger, the Simons Professor of Mathematics in EECS and the Computer Science and Artificial Intelligence Laboratory (CSAIL).
Annually, students also learn about emerging deep-learning applications from course sponsors. David Cox, co-director of the MIT-IBM Watson AI Lab, presented on neuro-symbolic AI, a hybrid approach combining symbolic programs with deep learning's advanced pattern-matching capabilities. Alex Wiltschko, a senior researcher at Google Brain, discussed employing network analysis tools to predict molecular scents. Chuan Li, chief scientific officer at Lambda Labs, explored neural rendering techniques for reconstructing and generating graphical scenes. Animesh Garg, a senior researcher at NVIDIA, outlined strategies for developing robots with more human-like perception and action capabilities.
With 350 students attending the in-person course annually and over a million online lecture viewers, Amini and Soleimany have emerged as influential ambassadors for deep learning education. Interestingly, their partnership began through tennis rather than technology.
Amini competed nationally in tennis during his high school years in Ireland and created an award-winning AI model to help amateur and professional players improve their techniques; Soleimany served twice as captain of the MIT women's tennis team. They connected on the court as undergraduates and discovered their shared enthusiasm for machine learning.
After completing their undergraduate degrees, they decided to challenge themselves by addressing what they perceived as MIT's growing need for a comprehensive deep learning foundation course. Initially launched in 2017 by graduate students Nick Locascio and Harini Suresh, 6.S191 was envisioned by Amini and Soleimany as something more transformative. They developed a series of software laboratories, introduced cutting-edge topics including robust and ethical AI, and incorporated content appealing to diverse students—from computer scientists to aerospace engineers and MBA candidates.
"Alexander and I continuously brainstorm, and these discussions are essential to how 6.S191 and several of our collaborative research projects have evolved," says Soleimany.
They incorporate one such research collaboration into their coursework. During the computer vision laboratory, students examine algorithmic bias and learn to detect and address racial and gender biases in facial recognition tools. This lab is based on an algorithm that Amini and Soleimany developed with their respective advisors, Daniela Rus, CSAIL director, and Sangeeta Bhatia, the John J. and Dorothy Wilson Professor of HST and EECS. This year, they also addressed current robotics topics, including recent developments in Amini's autonomous vehicle research.
However, they have no intention of stopping there. "We're dedicated to making 6.S191 the best it can be with each iteration," Amini states. "That means continuously advancing the course as deep learning technology evolves."
