Course Overview
The Mit Deep Learning course (6.S898), offered by the Department of Electrical Engineering and Computer Science (EECS), provides a comprehensive exploration into the fundamentals and advanced applications of deep learning. This Fall 2023 syllabus outlines a rigorous program designed for advanced undergraduate and graduate students seeking to master the theoretical underpinnings and practical implementations of deep learning methodologies.
This MIT Deep Learning course delves into a wide array of crucial topics, starting with the foundational architectures of neural networks. Students will gain in-depth knowledge of Multilayer Perceptrons (MLPs), Convolutional Neural Networks (CNNs), Recurrent Neural Networks (RNNs), Graph Neural Networks, and Transformers. The curriculum extends beyond architecture, exploring the geometric and invariant properties inherent in deep learning models, crucial for understanding how these models learn and generalize.
A significant portion of the course is dedicated to the mechanics of training deep networks, covering backpropagation, automatic differentiation, and stochastic gradient descent (SGD). Furthermore, the MIT Deep Learning syllabus addresses the critical aspects of learning theory and generalization, particularly in high-dimensional spaces, which are typical in modern deep learning applications. The course culminates in exploring diverse applications of deep learning across cutting-edge fields including computer vision, natural language processing (NLP), and robotics, demonstrating the versatility and power of these techniques.
Prerequisites for this demanding yet rewarding MIT Deep Learning course include a solid foundation in:
- (6.3900 [6.036] or 6.C01 or 6.3720 [6.401]): Foundational computer science or machine learning courses.
- (6.3700 [6.041] or 6.3800 [6.008] or 18.05): Probability and statistics.
- (18.C06 or 18.06): Linear algebra.
This 3-0-9 unit MIT Deep Learning course is designed to be challenging and deeply informative, making it suitable for both advanced undergraduates and graduate students eager to specialize in artificial intelligence. Non-MIT students interested in accessing course resources like Piazza or Canvas should contact Anthea Li ([email protected]) for manual enrollment details. For non-MIT students from other institutions, cross-registration information is available via cross-registration.
Meet Your Instructors: Leading Deep Learning Experts at MIT
The MIT Deep Learning course is led by a team of distinguished researchers and educators, each bringing unique expertise and perspectives to the field:
Phillip Isola
Phillip Isola is a renowned expert in computer vision and deep learning. His research focuses on understanding visual perception through the lens of computation, with significant contributions to image synthesis and image-to-image translation using deep neural networks. Students of the MIT Deep Learning course will benefit from his profound insights into neural network architectures and representation learning.
- Email: phillipi at mit dot edu
- Office Hours: Mondays, 2:00 PM – 3:00 PM, Room 34-302
Sara Beery
Sara Beery specializes in applying MIT Deep Learning techniques to tackle real-world problems, particularly in the domain of environmental sustainability and conservation. Her work emphasizes the importance of robust and generalizable deep learning models for complex, unstructured data. Her lectures in the MIT Deep Learning course will cover crucial aspects like transfer learning and out-of-distribution generalization.
- Email: beery at mit dot edu
- Office Hours: Tuesdays, 9:00 AM – 10:00 AM, Room 36-153
Jeremy Bernstein
Jeremy Bernstein brings a strong theoretical foundation to the MIT Deep Learning course. His expertise lies in the mathematical principles underpinning deep learning, including optimization, generalization theory, and scaling laws. His lectures will provide students with a deeper understanding of the theoretical landscape of MIT Deep Learning.
- Email: jbernstein at mit dot edu
- Office Hours: Tuesdays, 4:00 PM – 5:00 PM, Room 34-302
Dedicated Teaching Assistants: Support for Your Deep Learning Journey
The instructors are supported by a team of highly qualified Teaching Assistants (TAs), who are integral to the MIT Deep Learning experience, offering guidance and support throughout the course:
Anthea Li
Anthea Li is a TA for MIT Deep Learning.
- Email: yichenl at mit dot edu
- Office Hours: Wednesdays, 2:00 PM – 3:00 PM, Room 24-308
Thien Le
Thien Le is a TA for MIT Deep Learning.
- Email: thienle at mit dot edu
- Office Hours: Thursdays, 3:00 PM – 4:00 PM, Room 34-302
Saachi Jain
Saachi Jain is a TA for MIT Deep Learning.
- Email: saachij at mit dot edu
- Office Hours: Wednesdays, 10:00 AM – 11:00 AM, Room 24-319
Veevee Cai
Veevee Cai is a TA for MIT Deep Learning.
- Email: cail at mit dot edu
- Office Hours: Fridays, 11:00 AM – 12:00 PM, Room 24-319
Pratyusha Sharma
Pratyusha Sharma is a TA for MIT Deep Learning.
- Email: pratyusha at mit dot edu
- Office Hours: Wednesdays, 9:00 AM – 10:00 AM, Room 24-323
Jocelyn Shen
Jocelyn Shen is a TA for MIT Deep Learning.
- Email: joceshen at mit dot edu
- Office Hours: Tuesdays, 11:30 AM – 12:30 PM, Room 24-323
Course Logistics and Resources
Essential Information for Students
- Class Meetings: Tuesdays and Thursdays, 1:00 PM – 2:30 PM in Room 2-190.
- Online Platforms: The MIT Deep Learning course utilizes Piazza and Canvas for announcements and communication.
- Piazza: For all course-related questions and discussions.
- Canvas: For important announcements and homework assignments.
- Gradescope: For homework submission and grade viewing.
- Personal Inquiries: For logistical or personal questions, students are encouraged to contact the instructors and TAs via private Piazza posts, which is the preferred method, or via email.
Grading and Coursework
Assessment Breakdown
The MIT Deep Learning course employs a balanced grading scheme to evaluate student learning:
- Problem Sets (65%): Problem sets are designed to reinforce the theoretical concepts and practical skills taught in the lectures. They are a crucial component of the MIT Deep Learning curriculum, providing hands-on experience with deep learning techniques.
- Final Project (35%): The final project is a research-oriented endeavor, allowing students to delve deeper into a specific area of deep learning that interests them. Students will conduct experiments, perform in-depth analysis, and present their findings in a blog post format. This project emphasizes clarity, insight, novelty, and the depth of experimental analysis within the realm of MIT Deep Learning.
Final Project Details
The final project for MIT Deep Learning is designed to be a significant learning experience, encouraging students to explore and contribute to the field. Key aspects of the final project include:
- Research Focus: Students choose a research question within deep learning, allowing for personalized exploration and specialization.
- Experimental and Analytical Work: Projects require running experiments and conducting thorough analysis to validate or explore research hypotheses.
- Blog Post Write-up: The final output is a blog post, requiring students to communicate complex technical information in an accessible and engaging manner. This includes explaining background material, detailing investigations, and clearly presenting results.
- Visualizations: Students are encouraged to enhance their blog posts with plots, animations, and interactive graphics to improve clarity and engagement, mirroring best practices in communicating MIT Deep Learning research.
- Inspiration: Examples of exemplary research blog posts are provided for inspiration: [1] [2] [3] [4].
Grading of the final project will assess not only the novelty and depth of research but also the clarity of presentation and the insights demonstrated, reflecting the high standards of MIT Deep Learning education. Detailed guidelines for the final project will be provided later in the semester.
Class Schedule: A Deep Dive into Deep Learning Topics
Fall 2023 Course Outline
The MIT Deep Learning course schedule is meticulously planned to cover a wide spectrum of topics, progressing from foundational concepts to advanced research areas. Note that the schedule is subject to adjustments.
| Week | Date | Topics | Speaker | Course Materials | Assignments |
|---|---|---|---|---|---|
| Week 1 | Thu 9/7 | Course overview, introduction to deep neural networks and their basic building blocks | Sara Beery | slides notation for this course notes optional reading: Neural nets as distribution transformers “` |
