MIT's introductory program on deep learning methods with applications to computer vision, natural language processing, biology, and more! Students will gain foundational knowledge of deep learning algorithms and get practical experience in building neural networks in TensorFlow. Program concludes with a project proposal competition with feedback from staff and panel of industry sponsors. Prerequisites assume calculus (i.e. taking derivatives) and linear algebra (i.e. matrix multiplication), we'll try to explain everything else along the way! Experience in Python is helpful but not necessary. Listeners are welcome!
The 2023 in-person edition has completed during MIT IAP 2023 and was held in MIT Room 32-123. The online edition of the course is now live on Friday at 10am ET, every week!
We are expecting very elementary knowledge of linear algebra and calculus. How to multiply matrices, take derivatives and apply the chain rule. Familiarity in Python is a big plus as well. The program will be beginner friendly since we have many registered students from outside of computer science.
If you would like to receive related updates and lecture materials please subscribe to our YouTube channel and sign up for our mailing list.
All materials are available online for free but are copyrighted and licensed under the MIT license. If you are an instructor and would like to use any materials from this program (slides, labs, code), you must add the following reference to each slide:
© Alexander Amini and Ava Amini
MIT Introduction to Deep Learning
IntroToDeepLearning.com
All materials are copyrighted and licensed under the MIT license. If you are an instructor and would like to use any materials from this program (slides, labs, code), you must add the following reference to each slide:
© Alexander Amini and Ava Amini
MIT 6.S191: Introduction to Deep Learning
IntroToDeepLearning.com
If you are an MIT student, postdoc, faculty, or affiliate and would like to become involved with this program please email introtodeeplearning-staff@mit.edu. We are always accepting new applications to join the program staff.
This class would not be possible without our amazing sponsors and has been sponsored by Google, IBM, NVIDIA, Ernst and Young, LambdaLabs, Tencent AI, Microsoft, Amazon, and Onepanel. If you are interested in becoming involved in this program as a sponsor please contact us at introtodeeplearning-staff@mit.edu.
Copyright © MIT 6.S191. banner image
coming soon!
coming soon!
Artificial intelligence has shown incredible promise in safety-critical domains in the past few years, but such advances are far from being realized in the real world due to a fundamental lack of safety and robustness. In addition to being capable of solving a particular problem, models also must be able to diagnose biases and uncertainty, which may lead to erroneous or unworthy decisions. Achieving safe and trustworthy AI requires integrating such safeguards at every stage in the artificial intelligence lifecycle. In this talk, I'll present Themis AI's new learning-based approaches towards dynamic, resilient, and robust decision making of deep learning models. We advance robust decision making in the wild by addressing critical challenges that arise at all stages, stemming from the data used for training, to the models that learn on this data, to the algorithms to reliably adapt to unexpected events during deployment.
Sadhana Lolla is a machine learning scientist at Themis AI and a third-year undergraduate at MIT studying computer science. Her research focuses on the application of deep learning to robotics and designing safe and trustworthy AI. She conducts research at the Distributed Robotics Laboratory with Prof. Daniela Rus and Dr. Alexander Amini. Previously, she has worked on accelerating distributed training at the Keras team at Google, as well as on the Discover Ranking team on recommendation modeling. For her research at MIT, she's been recognized as a Gerstle Undergraduate Innovation Scholar.
We are in the modern era of statistics where overparameterized (very deep and large) learning systems remarkably improve performance in real-world applications ranging from language modeling and signal processing, to computer vision. There exists tremendous empirical evidence for the ability of these increasingly general AI systems in performing pattern matching, interpolation, few-shot and zero-shot generalization across a wide range of tasks. In this space, mixed observations have been reported on emerged behaviors of learning systems, that give rise to fundamental questions such as: is scaling learning models truly necessary for obtaining intelligent behavior? What is the boundary between interpolation and extrapolation in representation learning where models generalize to unseen test data? How to achieve robust decision-making and attain out-of-distribution generalization? And how does memory interplay with reasoning? In this lecture, I will provide some insights on these questions, and motivate how our research on liquid neural networks impact the design of future AI systems in this modern era of statistics.
Dr. Ramin Hasani is a Principal AI and Machine Learning Scientist at the Vanguard Group and a Research Affiliate at MIT CSAIL. Ramin’s research focuses on robust deep learning and decision-making in complex dynamical systems. Previously he was a Postdoctoral Associate at MIT CSAIL, leading research on modeling intelligence and sequential decision-making, with Prof. Daniela Rus. He received his Ph.D. degree with distinction in Computer Science at the Vienna University of Technology (TU Wien), Austria (May 2020). His Ph.D. dissertation and continued research on Liquid Neural Networks got recognized internationally with numerous nominations and awards such as TÜV Austria Dissertation Award nomination in 2020, and HPC Innovation Excellence Award in 2022. Dr. Hasani has also been a frequent TEDx Speaker.
I will talk about Muse, a text-to-image Transformer model that achieves state-of-the-art image generation performance while being significantly more efficient than diffusion or autoregressive models. This is work done by my team at Google Research. Muse is trained on a masked modeling task in discrete token space: given the text embedding extracted from a pre-trained large language model (LLM), Muse is trained to predict randomly masked image tokens. Compared to pixel-space diffusion models, such as Imagen and DALL-E 2, Muse is significantly more efficient due to the use of discrete tokens and requiring fewer sampling iterations; compared to autoregressive models, such as Parti, Muse is more efficient due to the use of parallel decoding. The use of a pre-trained LLM enables fine-grained language understanding, translating to high-fidelity image generation and the understanding of visual concepts such as objects, their spatial relationships, pose, cardinality etc. Our 900M parameter model achieves a new SOTA on CC3M, with an FID score of 6.06. The Muse 3B parameter model achieves an FID of 7.88 on zero-shot COCO evaluation, along with a CLIP score of 0.32. Muse also directly enables a number of image editing applications without the need to fine-tune or invert the model: inpainting, outpainting, and mask-free editing. More results are available here.
Dilip Krishnan is a Research Scientist at Google’s Cambridge office (Massachusetts). He currently works on generative models and representation learning. From August 2013 to November 2014, he was a Postdoctoral Associate with Bill Freeman at MIT’s CSAIL Lab. In June 2013, he received my PhD from New York University, under the supervision of Rob Fergus. Prior to his PhD, he co-founded a startup that developed software for video editing and visual effects, which has over 2 million users.
