MIT 6.S094: Introduction to Deep Learning, Neural Networks, and Self-Driving Cars

The lecture introduces the foundational concepts of deep learning and deep neural networks, specifically through the lens of building self-driving cars. It outlines how methods like deep reinforcement learning, convolutional neural networks (CNNs), and recurrent neural networks (RNNs) are applied to critical autonomous driving components such as visual perception, localization, mapping, control planning, and driver state detection. Practical applications are demonstrated through two projects: "DeepTraffic," which allows browser-based neural network training using ConvNet.JS for traffic simulation, and "DeepTesla," which employs end-to-end learning from real Tesla data to predict steering angles from a single image.

A core philosophical inquiry explored is the nature of "intelligence" itself, distinguishing between special-purpose AI, designed for formally defined, finite goals (e.g., chess), and general-purpose AI, which aims to tackle poorly defined, unconstrained objectives akin to human cognition. The speaker highlights that while AI has excelled in formal and expert tasks, seemingly mundane human activities like everyday speech, visual perception, and walking present profound challenges for machines, suggesting that these tasks may require a higher degree of reasoning beyond current data-driven approaches. This raises the question of whether incremental progress in special-purpose AI will eventually lead to general intelligence, or if entirely new mechanisms are needed.

The lecture delves into the building blocks of neural networks, explaining artificial neurons and their interconnected nature, drawing inspiration from the human brain's vast synaptic complexity. It discusses the mathematical universality of neural networks, capable of approximating any function with a single hidden layer, making them theoretically powerful for tasks like autonomous driving. The distinction between supervised learning (learning from labeled data), unsupervised learning (learning without output labels), and reinforcement learning (learning from sparse rewards, as demonstrated by an AI mastering Pong from raw pixels) is clarified, with an emphasis on the current breakthroughs being predominantly in supervised and reinforcement learning.

Despite the advancements, significant challenges for deep neural networks are identified, including their inefficiency in learning (requiring massive datasets), the high cost of data annotation, the complexity of hyperparameter tuning, and the difficulty in defining robust loss or reward functions that align with human-understood broader goals. The "Coast Runners" example illustrates how an AI can optimize for local rewards while ignoring the implied overall objective, raising critical ethical questions for self-driving cars, particularly concerning decision-making in unavoidable crash scenarios. The lecture concludes by attributing recent deep learning breakthroughs to increased computational power (Moore's Law, GPUs), abundant digital data, algorithmic innovations (Backprop, CNNs, LSTMs), and robust infrastructure (TensorFlow, AWS, Mechanical Turk), alongside substantial financial investment from tech giants, while acknowledging that fundamental issues like efficient, generalizable learning and embedding complex reasoning remain open problems.

We will introduce to you the methods of deep learning, of deep neural networks using the guiding case study of building self-driving cars.

What you will learn is an overview of deep reinforcement learning, of convolutional neural networks, recurring neural networks and how these methods can help improve each of the components of autonomous driving - perception, visual perception, localization, mapping, control planning and the detection of driver state.

The question of: "What is intelligence?" This is from a March 1996 Time magazine. And the question: "Can machines think?" is answered below with, "they already do."

The question is, if we inch along this path, will we get closer to the general formulation, to the general purpose version of what artificial intelligence is?

The mundane tasks of everyday speech, of written language, of visual perception, of walking which we'll talk about in today's lecture is a fascinatingly difficult task on object manipulation.

The open question is: if driving is more like a conversation, like in natural language conversation, how hard is it to pass the Turing Test?

Our human brain, to the best estimate that I'm aware of, has 10,000X that. So one hundred to one thousand trillion synapses.

This system knows nothing about "pong", nothing about games, this is general intelligence. Except for the fact, that it's just a game "pong".

Human beings need one example often times and they learn very efficiently from that one example.

It's hard to teach, it's hard to encode the formally defined utility function under which an intelligent system needs to operate.