Artificial intelligence and deep learning technologies are evolving at an unprecedented pace, and educational pioneers Alexander Amini '17 and Ava Soleimany '16 are revolutionizing how students grasp complex mathematical algorithms behind AI systems that are transforming our everyday experiences.
Their groundbreaking educational offering, 6.S191 (Introduction to Deep Learning), previously kicked off with a surprisingly realistic AI-generated video featuring former President Barack Obama. This academic year, the innovative teaching duo delivered their virtual lectures from what appeared to be MIT's iconic Stata Center — though these sessions were actually meticulously pre-recorded weeks earlier in their home kitchen, specially equipped with professional lighting, a lecture podium, and an advanced green screen setup that digitally projected the familiar Kirsch Auditorium background during Zoom sessions.
"Maintaining student engagement becomes significantly challenging when instructors appear merely as static images on screen," explains Amini. "Our primary objective was to recreate the immersive, interactive environment of a physical classroom experience."
Amini, currently pursuing graduate studies in MIT's prestigious Department of Electrical Engineering and Computer Science (EECS), and Soleimany, a joint graduate student at MIT and Harvard University, have collaboratively developed 6.S191's comprehensive curriculum. They have taught this popular course during MIT's Independent Activities Period (IAP) for four of the past five years. While their lecture materials and hands-on software labs receive annual updates, this year's pandemic-induced virtual format presented unique challenges. Their creative response blended both traditional and cutting-edge solutions, from pre-recording lectures to utilizing a Minecraft-inspired virtual platform that facilitated organic social interaction among participants.
Some astute students discovered the pre-recorded nature of lectures after noticing subtle details like instructors' abrupt wardrobe changes when transitioning from lecture delivery to immediate post-class help sessions. Students who recognized the technical feat praised the instructors in course evaluations, while those unaware reacted with genuine astonishment. "You mean these weren't live streams?" asked PhD candidate Nada Tarkhan, followed by a moment of stunned silence. "The experience genuinely felt like one instructor was presenting the lecture material while simultaneously another was responding to questions in the chat function."
The surging popularity of 6.S191 — both as a credit-bearing MIT course and as a self-paced online learning opportunity — reflects the exponential growth of deep neural networks across applications ranging from language translation to facial recognition systems. Through their series of crystal-clear, captivating presentations, Amini and Soleimany elucidate the technical foundations of deep networks, demonstrating how these sophisticated algorithms identify patterns within massive datasets to generate predictions. They also examine deep learning's diverse applications and guide students in evaluating model predictions for both accuracy and potential bias.
Responding to valuable student feedback, Amini and Soleimany expanded this year's course from one to two weeks, providing learners with additional time to thoroughly comprehend the material and develop their final projects. They also introduced two new lecture modules: one focusing on uncertainty quantification in AI models, and another addressing algorithmic bias and fairness in machine learning systems. The virtual format also enabled them to accommodate an additional 200 students who would have been excluded due to Kirsch Auditorium's 350-person capacity limitation.
To enhance student-instructor and peer-to-peer connectivity, Amini and Soleimany implemented Gather.Town, an innovative virtual platform they discovered at a machine learning conference the previous autumn. Students navigated the digital 6.S191 auditorium using personalized avatars to seek homework assistance or identify potential collaborators for troubleshooting their final project challenges.
Students consistently praised the course for its comprehensive scope and exceptional organization. "I was familiar with buzzwords like reinforcement learning and RNNs, but had never truly understood the practical implementation details, such as creating parameters in TensorFlow and configuring activation functions," explains sophomore Claire Dong. "After completing this course, I emerged with crystal-clear understanding and renewed enthusiasm for the field."
This academic year, 50 student teams presented final projects — double the previous year's participation — covering an even more diverse range of applications, according to Amini and Soleimany. Project topics spanned from cryptocurrency trading algorithms and forest fire prediction systems to sophisticated protein folding simulations within cellular environments.
"The additional week really empowered them to refine their concepts, develop functional prototypes, write the necessary code, and assemble comprehensive presentations," notes Amini.
"The projects were absolutely brilliant," adds Soleimany. "The quality, sophistication, and organization of their ideas and presentations exceeded our expectations."
Four outstanding projects received special recognition awards.
The first award-winning project proposed a sophisticated system for classifying brain signals to distinguish between right and left hand movement intentions. Before transferring to MIT from Miami-Dade Community College, Nelson Hidalgo had researched brain-computer interfaces designed to help individuals with paralysis regain limb control. For his final project, Hidalgo, now a sophomore in EECS, utilized EEG brain wave recordings to construct a model capable of differentiating signals associated with intended right-hand versus left-hand movements.
His innovative neural network architecture featured parallel convolutional and recurrent neural networks working in tandem to extract both sequential and spatial patterns from the neurological data. The resulting model demonstrated superior performance compared to existing methods for predicting the brain's movement intentions, potentially making this technology more accessible for daily patient use.
A second praiseworthy project explored AI applications in sustainable forestry management. Tree planting initiatives have become increasingly popular among corporations seeking carbon emission offsets, but accurately quantifying the actual carbon dioxide absorption by these trees remains an imprecise science. Peter McHale, a master's candidate at the MIT Sloan School of Management, proposed that his recently launched startup, Gaia AI, could deploy specialized drones to capture detailed imagery of forest canopies from multiple perspectives.
These high-resolution aerial images could enable forest managers to more accurately estimate tree growth rates and calculate carbon sequestration capacity, McHale explains. The visual data could also provide valuable insights regarding optimal tree species for specific climate conditions and environments. "Drones can collect measurements more cost-effectively and precisely than human surveyors," he notes.
During Gaia AI's initial development phase, McHale plans to focus on marketing premium drone-gathered sensor data to timber companies seeking affordable, accurate surveying solutions, as well as organizations providing third-party verification for carbon offset initiatives. In phase two, McHale envisions leveraging both the collected data and generated profits to combat climate change through large-scale drone-assisted reforestation efforts.
A third exceptional project investigated cutting-edge approaches to encoding intelligent behavior in robotic systems. As a participant in the SuperUROP research program in Professor Sangbae Kim's laboratory, Savva Morozov works with the mini cheetah robot and is particularly interested in developing methods that enable robots to acquire meta-learning capabilities.
For his project presentation, Morozov, a junior in the Department of Aeronautics and Astronautics, presented a compelling scenario: a mini cheetah-like robot encounters difficulty climbing a debris pile. The robot identifies a potentially useful wooden plank that could be grasped with its robotic arm and repurposed as a ramp. However, the machine lacks both the creative imagination and skill repertoire necessary to construct such a tool to reach its objective. Morozov explained how various meta-learning approaches could potentially solve this complex problem.
A fourth distinguished project proposed leveraging deep learning to streamline the analysis of building street-view images for modeling urban energy consumption patterns. An algorithm developed by MIT's Sustainable Design Lab and graduate student Jakub Szczesniak can estimate window-to-wall ratios for buildings based on details extracted from photographs, but processing these images requires substantial preliminary manual work.
PhD candidate Nada Tarkhan from the School of Architecture and Planning proposed integrating an image-processing convolutional neural network into the workflow to accelerate and enhance the reliability of this analysis. "We hope this technology can help us gather more precise data about urban building characteristics — including façade features, construction materials, and window-to-wall proportions," she explains. "Our ultimate objective is to improve our understanding of building performance patterns across entire cities."
Based on overwhelmingly positive student feedback, Amini and Soleimany plan to maintain the enhanced focus on uncertainty quantification and algorithmic fairness while expanding the course into emerging AI domains. "We're thrilled when students report that our course inspired them to pursue additional AI/ML coursework after completing 6.S191," shares Soleimany. "We're committed to continuous innovation to ensure our course remains at the forefront of AI education."
The course received generous support from Ernst & Young, Google, the MIT-IBM Watson AI Lab, and NVIDIA.
