Deep Learning for Self-Driving Cars: Why Edge Cases and Human Nature Are the Real Problem

Fridman calls self-driving cars quite possibly the first wide-reaching and profound integration of personal robots into society. It is wide-reaching because there are a billion cars on the road, and profound because of the direct transfer of control between human and machine, which puts a person's life in the hands of an AI system. He says this life-critical setting will truly test what AI can do.

From that he draws the course's central argument: an autonomous vehicle must perceive more than the road. It also has to perceive the driver and passengers, interact with them and earn their trust. Perfect perception and control in a world full of humans is extremely difficult and, in his estimate, could be two to four decades away, so cars will be flawed and must transfer control to people when they cannot cope.

Obstacle avoidance at 30 miles an hour through Boston is, he says, the easy and safe problem. The hard part is human nature: someone late for work who wants to speed into the opposing lane. He previews algorithms that estimate a driver's cognitive load from eye movement and blinking, along with gaze, emotion, body pose and drowsiness.

Current approaches, he says, handle 90 percent and rising of driving cases, but the remaining sliver is where humans dominate. He shows a clip of the Arc de Triomphe roundabout as one of millions of situations where human-to-human interaction drives the outcome, not basic perception and control. That, he says, is why deep learning fits: it learns from real-world data, and any system that operates in the real world must do so.

Fridman defines intelligence as the ability to accomplish complex goals and understanding as the ability to turn complex information into simple, useful, actionable information. Deep learning, he says, does exactly that by building hierarchical representations from raw data, whether supervised or unsupervised. In an image classifier the layers move from edges to contours, then corners and object parts, and finally the full semantic class.

He illustrates why representation matters with a four-centuries-old example: placing the Earth or the Sun at the center of the universe. The data is the same, but one representation is far more complicated. A second example separates green triangles from blue circles, which is impossible with a straight line in Cartesian coordinates but trivial in polar coordinates.

A single-hidden-layer network can learn that kind of transformation by itself, reshaping the space so that a straight line separates two curves. Because this learning works from raw sensory input, such systems get better with more data, which suits real applications where edge cases are everything. Fridman describes running a Tesla Autopilot v1 system and his own network on a Jetson TX2 side by side, and treating their disagreement as a flag for a likely edge case that needs human intervention.

Fridman notes that artificial neural networks are only loosely inspired by biology and that the differences are instructive. A simulated thalamocortical network has about 3 million neurons and 476 million synapses, while the human brain has around 100 billion neurons and 1,000 trillion synapses. The state-of-the-art ResNet-152 has about 60 million, a gap of roughly seven orders of magnitude.

Structurally, he says, brains have no layers, little imposed structure, neurons connected to 10,000 or more others, and asynchronous operation. Artificial networks are layered, fixed and synchronous, and trained with backpropagation, the one learning algorithm we have, while we do not know how brains learn. Artificial neurons are faster but very power-inefficient, and unlike brains, which keep learning, networks have separate training and testing stages.

The one profound similarity he highlights is distributed computation at scale: each neuron is extremely simple, yet connected together they form powerful approximators. A unit multiplies inputs by edge weights, adds a bias and applies a nonlinearity. Feed-forward networks dominate computer vision, while recurrent networks that feed back into themselves are closer to the brain and suit time series, but are much harder to train. With a single hidden layer such networks can approximate any function, so the limit lies in how we build and train them.

Fridman splits machine learning into memorizers and methods that begin to reason and generalize with minimal human input. Supervised learning, where humans annotate most of the data, accounts for most deep learning successes. He places his hopes for the near term in semi-supervised, reinforcement and unsupervised methods, which could let machines make sense of the world with little human help.

He contrasts special-purpose intelligence, such as a network that estimates Boston apartment rents from bedrooms, square footage and neighborhood, with approaches that edge toward general-purpose intelligence. His example is Andrej Karpathy's "Pong from Pixels", where a system sees only raw 80 by 80 images and learns to win with very sparse human input.

Training itself is simple at its core. A forward pass produces a prediction, a loss function measures the error against the ground truth, and backpropagation sends that error back through the weights, subtracting a fraction of each weight's gradient. Because the process is local to each neuron, it parallelizes well across GPUs. Fridman also walks through activation functions: sigmoid and tanh suffer from vanishing gradients, while ReLU avoids them but can die when the derivative is exactly zero.

Fridman calls overfitting one of the biggest problems in all of machine learning. He shows a sine curve that fits some points well and a ninth-degree polynomial that fits them better on paper but would produce large errors on unseen data. The telltale symptom is error falling on the training set while rising on the test set, which is a particular risk with small datasets.

Because training neural networks is costly, practitioners hold out a validation set from the training data as a stand-in for the test set. They save checkpoints and watch validation performance, stopping early or picking the best snapshot when it starts to worsen. Dropout randomly removes nodes and their edges during training, usually keeping each with a probability above 0.5, and adding noise to inputs can work well too.

He also covers weight penalties. The L2 penalty keeps weights small unless the error derivative is huge and spreads weight across similar inputs, producing a smoother, more robust model. The L1 penalty lets a few weights stay very large. Fridman recommends the TensorFlow Playground for building intuition about layers, features and regularization.

Neural networks went through two winters before returning to dominate AI. Fridman credits several converging changes: far more computing power from CPUs to GPUs and ASICs, huge datasets such as ImageNet, research advances like backpropagation, convolutional networks and LSTMs, software frameworks such as TensorFlow and PyTorch that treat networks as stacks of layers, and heavy financial backing from companies like Google and Facebook.

To understand where deep learning is strong or weak, he says, we need to see where our own intuition about hard and easy comes from. Human visual perception was formed 540 million years ago, while abstract thought appeared only around 100 thousand years ago, so vision gets many orders of magnitude more evolutionary data. Things that look trivial to us can be baffling to networks, such as a dog plus a little noise classified as an ostrich with over 99 percent confidence.

He lists the variability that makes perception hard: illumination, which he calls the biggest problem in driving, pose variation and intraclass variation. Most work on faces is done on frontal or semi-frontal views, with little on full 360-degree pose. His advice is to keep our intuition tuned to how machines learn, not how humans evolved to see.

Fridman traces progress through object classification, starting with the ImageNet dataset of 14 million images and 21,000 categories. AlexNet in 2012 was among the first networks trained successfully on a GPU and made a significant jump in performance. Human-level performance on the five-guess task is 5.1 percent error, and ResNet in 2015 was the first to exceed it, with later networks reaching about 2.25 percent. He calls this a toy example, and says the real question now is bringing these methods into the real world for scene and driver perception.

The same architecture can be adapted to other tasks. Swapping the final layers enables image captioning with recurrent networks, localization with R-CNN variants and pixel-level segmentation with fully convolutional networks. Further image-to-image applications include colorization, background removal from selfies and pix2pixHD, which generates photorealistic road scenes from coarse semantic labels and could help augment driving data.

Recurrent networks handle sequences such as handwriting, captions and video description, and attention mechanisms mimic how humans look around an image. On the reinforcement side, AlphaGo beat the top human player in 2016 after training on expert games. AlphaGo Zero then beat it in 2017 by playing itself from zero human knowledge, producing moves that surprised experts, and DeepStack won heads-up poker, while multiplayer tournaments remain the task for 2018.

Fridman shows the CoastRunners video to illustrate how the reward function governs behavior. The boat is meant to race, but it learns to ignore the race and circle through green targets that keep regenerating to rack up points. Even in a simple formal system it is extremely hard to write a reward that produces the intended behavior, which he says is highly relevant to autonomous vehicles.

He then lists the field's open challenges. Transfer learning works between close domains but not across big leaps, and the biggest challenge is generalizing across domains and reasoning in the sense he defined earlier. Current systems are inefficient: they need big, human-supervised datasets, and educated humans such as PhD students and postdocs to design architectures and tune hyperparameters.

Neural networks are also largely black boxes, and visualization methods only partly reveal why they work or fail. Fridman says people may not care about transparency if a system works well enough, but that it will be a long time before that is true, so transparency, communication and collaboration remain critical. He ends on edge cases: 99.9 percent of driving is boring, and the trillions of rare situations must be generalized over from very little data. He frames this as an opportunity and invites students to try the course competitions.