Deep Learning for AI and Computer Vision

Deep learning innovations are driving exciting breakthroughs in the field of computer vision. Robots and drones not only “see”, but respond and learn from their environment. Autonomous cars avoid collisions by extracting meaning from patterns in the visual signals surrounding the vehicle. 

This course covers the latest developments in vision AI, with a sharp focus on advanced deep learning methods, specifically convolutional neural networks, that enable smart vision systems to recognize, reason, interpret and react to images with improved precision. 

COVID-19 Updates: MIT Professional Education fully expects to hold courses on-campus during the Summer of 2023 (only a handful will be delivered in the live virtual or hybrid modality). In the event there is a change in MIT's COVID-19 policies and a course cannot be held on-campus, we will deliver courses via live virtual format. Find the latest information here.

Participants will explore the latest developments in neural network research and deep learning models that are enabling highly accurate and intelligent computer vision systems capable of understanding and learning from images. We will start from fundamental topics in image modeling, including image formation, feature extraction, and multiview geometry, then move on to the latest applications in object detection, 3D scene understanding, vision and language, image synthesis, and vision for embodied agents. By the end, participants will:

• Be familiar with fundamental concepts and applications in computer vision
• Grasp the principles of state-of-the art deep neural networks 
• Understand low-level image processing methods such as filtering and edge detection
• Gain knowledge of high-level vision tasks such as object recognition, scene recognition, face detection and human motion categorization 
• Develop practical skills necessary to build highly-accurate, advanced computer vision applications

Program Outline

This course meets 9:00 am - 5:00 pm each day.

Day One:
9:00am: Introduction to computer vision (Torralba)
10:00am: Cameras and image formation (Torralba)
11:00am: Coffee break
11:15am: Introduction to machine learning (Isola)
12:15pm: Lunch
1:30pm: The problem of generalization (Isola)
2:45pm: Coffee break
3:00pm: Lab on Pytorch
5:00pm : Adjourn

Day Two:
9:00am: Neural networks (Isola)
10:00am: Filters and CNNs (Torralba)
11:00am: Coffee break
11:15am: Stochastic gradient descent (Torralba)
12:15pm: Lunch break
1:30pm: Temporal processing and RNNs (Isola)
2:45pm: Coffee break
3:00pm: Lab on using modern computing infrastructure
5:00pm: Adjourn

Day Three:
9:00am: Multiview geometry (Torralba)
10:00am: 3D deep learning (Torralba)
11:00am: Coffee break
11:15am: Scene understanding part 1 (Isola)
12:15pm: Lunch break 
1:30pm: Scene understanding part 1 (Isola)
2:45pm: Coffee break
3:00pm: Lab on scene understanding
5:00pm: Adjourn

Day Four:
9:00am: People understanding (Torralba)
10:00am: Vision and language (Torralba)
11:00am: Coffee break
11:15am: Image synthesis and generative models (Isola)
12:15pm: Lunch break
1:30pm: AR/VR and graphics applications (Isola)
2:45pm: Coffee break
3:00pm: Lab on generative adversarial networks
5:00pm: Adjourn

Day Five:
9:00am: Vision for embodied agents (Isola)
10:00am: Modern computer vision in industry: self-driving, medical imaging, and social networks
(Torralba)
11:00am: Coffee break
11:15am: Datasets, bias, and adaptation, robustness, and security (Torralba)
12:15pm: Lunch break 
1:30pm: Deepfakes and their antidotes (Isola)
2:45pm: Coffee break
3:00pm: Lab on your own work (bring your project and we will help you to get started)
4:55pm: closing remarks
5:00pm: Adjourn

Links & Resources

• MIT CSAIL researchers discuss frontiers of generative AI, MIT News, April 12, 2023
• Artificial intelligence that understands object relationships: A new machine-learning model could enable robots to understand interactions in the world in the way humans do. MIT News, November 29, 2021
• Machine-learning system tackles speech and object recognition, all at once: Model learns to pick out objects within an image, using spoken description. MIT News, September 18, 2018
• Q&A: Phillip Isola on the art and science of generative models, MIT News, May 30, 2019

This course is designed for professionals from industries such as healthcare, government, security, and automotive manufacturing, who are looking to solve computer vision problems with deep learning. Participants who will benefit from the curriculum include:

• Engineers looking to design and build advanced computer vision applications 

• Data scientists who want to explore cutting-edge developments in neural network research and deep learning models, and acquire strategies for applying them to their work

• Managers and business executives seeking to capitalize on the latest advancements in deep learning for computer vision and AI 

• Government and military officials who need to understand how computer vision applications are deployed in pertinent technologies, such as drones and other defense systems

Requirements

Laptops with which you have administrative privileges along with Python installed are encouraged but not required for this course (all coding will be done in a browser).

Participants should have experience in programming with Python, as well as experience with linear algebra, calculus, statistics, and probability.