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

Course Overview and Logistics

• Course 6S094 introduces deep learning methods using self-driving cars as the primary case study.
• Instructors include Lex Friedman, with teaching assistants Dan Brown, William Angio, Spencer Dodd, and Benedict Jenick.
• Two main simulation projects are utilized: "Deep Traffic" (a reinforcement learning game) and "Deep Tesla" (end-to-end autonomous driving).
• "Deep Traffic" is a top-down, seven-lane browser game where agents are controlled by neural networks trained via JavaScript (ComNetJS).
• "Deep Tesla" utilizes forward-facing video data from Tesla vehicles to train convolutional neural networks to directly predict steering angles from single images.
• To receive course credit, students must design a neural network that drives a virtual car at or above 65 mph; winning requires exceeding this speed.
• Course materials, including PDF lecture slides and tutorials, are hosted at deepcars.mit.edu; full video lectures will be posted a few days after recording.

Defining Intelligence and Driving Complexity

• Driving is characterized not as a formal task (like chess with defined rules) but as a "mundane" task involving undefined goals and under-actuated environments, similar to natural language or human conversation.
• Autonomous driving systems rely on a sensor suite including radar, lidar, visible light cameras, stereo vision, GPS, IMU, and CAN network data for vehicle telemetry.
• Under-studied research areas include audio input (road surface sounds, tire screeching) and internal-facing sensors to monitor driver state (drowsiness, emotional state).
• The architecture of an autonomous vehicle involves four core modules: localization/mapping, scene understanding (object classification), movement planning, and driver state monitoring for handover scenarios.
• Historical milestones include the Stanford Stanley team winning the 2005 DARPA Grand Challenge and CMU's Boss winning the 2007 DARPA Urban Challenge.
• Current industry players include Google (Waymo), Tesla (Autopilot 2), Uber, and local firms like Mutonomy testing in Boston.

Neural Network Fundamentals and Learning Paradigms

• Artificial neurons are simplified computational units inspired by human biology, acting as linear classifiers that map weighted inputs to outputs via activation functions.
• Neural networks exhibit universality, where a single hidden layer can approximate any continuous function, implying a theoretically perfect driving network exists given sufficient capacity.
• Three learning paradigms are distinguished: Supervised (learning from ground truth data), Unsupervised (no output labels), and Reinforcement Learning (learning from sparse rewards/punishments based on final outcomes).
• Deep reinforcement learning was demonstrated via an AI that learned to play Pong from raw pixels, achieving human-level performance without pre-programmed game rules.
• A significant challenge for deep learning is data inefficiency; humans learn complex tasks from single examples, whereas neural networks require massive datasets (e.g., 200,000 game simulations).
• Hyperparameter tuning is complex, involving decisions on network architecture, activation functions, learning rates, loss functions, and optimizers.

Challenges in Deep Learning and Safety

• Neural networks are vulnerable to adversarial attacks where minor pixel distortions can cause high-confidence misclassifications (e.g., classifying a tortoise as a rifle).
• AI systems can find "local optima" in reward functions that satisfy the mathematical goal while violating human intent, such as an agent in Coast Runners looping to collect points instead of finishing the race.
• The margin of error for self-driving cars is extremely low (one fatality per 100 million miles), requiring robustness against weather, lighting variations, and unpredictable human behavior.
• Historical AI winters (1970s, late 1980s) resulted from over-hyping capabilities (e.g., 1958 claims of self-reproducing electronic computers) beyond current technological limits.
• Security risks include LiDAR spoofing, where replay attacks can trick vehicles into perceiving non-existent obstacles or vehicles.

Technological Landscape and Future Directions

• Recent breakthroughs are attributed to increased compute power (GPUs, ASICs), large-scale digitized data (ImageNet), and algorithmic innovations (backpropagation, CNNs, LSTMs).
• Deep learning acts as "representation learning," automatically discovering hierarchical features (edges -> corners -> objects) rather than relying on manually engineered features.
• Convolutional Neural Networks (CNNs) have achieved human-level performance on ImageNet classification (error rates below 4% in 2015) despite challenges in illumination, pose, and occlusion.
• Recurrent Neural Networks (RNNs) enable sequence processing for natural language generation, image captioning, and video analysis.
• The "next five years" of research focus on mobile deployment, unsupervised learning, multimodal data integration, and commercial viability.
• Common deep learning libraries include TensorFlow (Google), Torch (Facebook), PyTorch (implied via Torch context), MXNet (Amazon), and browser-based options like ComNetJS and Keras.js.