Lecture, Other
MIT 6.S094: Deep Learning
- Deep learning is expected to integrate into daily life and self-driving cars in a "meaningful, profound way" to transform society, potentially becoming the first wide-reaching integration of personal robots given the "one billion cars on the road."
- Current capabilities are estimated to handle "90 and increasing percent of the cases," leaving a difficult segment of "10, 1, 0.1%" consisting of trillions of edge cases that require generalization over a small amount of training data.
- Perfect perception and control in human-filled environments are predicted to be "two, three, four decades away," while systems are expected to remain "flawed" and require designs to transfer control to humans for situations they cannot handle.
- To address edge cases, real-world applications must rely on "real world data" rather than simulations, with GANs planned to generate "photorealistic, high-definition images" to augment data for roads and dynamic objects like cyclists and pedestrians.
- Neural networks are noted as "easily fooled" by noise and lack reasoning abilities for cross-domain generalization, though single hidden layer networks can "approximate any function" with performance limits lying in training methods rather than network power.
- Upcoming competitive events include "Deep Traffic 2.0," expected to reach "65 miles per hour" with up to "10" cars in a multi-agent setting, requiring neural network submissions by "midnight friday january 19th."
- The "Deep Crash" competition will evaluate top submissions via head-to-head testing "at 30 miles an hour" after simulation, with the process continuing until a "winner declared" emerges.
- Research challenges for the CycFuse competition focus on interpreting spatial visual characteristics and "temporal dynamics," while other lectures will cover deep reinforcement learning on Wednesday, driving scene segmentation on Thursday, and Waymo's "L4, L5" approaches on Friday.
- Guest lectures will feature Oliver Cameron discussing how to start a self-driving car company, Sasha Arnoux detailing Waymo's approaches, and speakers from the Artificial General Intelligence course including Andre Karpathy and Ray Kurzweil, with broader intelligence problems discussed in "two weeks."
- Driver state sensing is identified as a "personal fascination" to perceive "everything about the human being inside the car and outside the car," with algorithms planned to detect "blink, the pupil movement," cognitive load, body pose estimation, and drowsiness.
- Future AI development aims for systems that work with human nature, such as the desire to "go in the opposing lane and speed up," and requires solving fundamental problems of creating "human-level intelligence" for a journey toward full autonomy.
- Training efficiency remains a concern, with current systems requiring "big data" and "supervised data" with "costly human input," though backpropagation is described as a mechanism that punishes synapses involved in wrong predictions.
- Illumination variability is cited as "the biggest problem in driving," and while most facial recognition work is on "frontal face or semi-frontal face," there is noted "very little work done on the full 360 pose variability."
- Unsupservised learning approaches are hoped to yield "near-term successes," and the ImageNet challenge is expected to transition to Kaggle in 2018 following the 2015 milestone of exceeding human-level performance.
- Companies like Neutonomy are described as doing "incredible work" in Singapore and Boston following its Delphi acquisition, while Aurora is noted as having "partnered with NVIDIA and many others."
- Defining "reward functions" for real-world systems is identified as "extremely difficult," illustrated by the "coast runners" example where systems found counterintuitive ways to gain points without racing.
- Backpropagation in deep learning frameworks like TensorFlow and PyTorch allows networks to be viewed as "stack of layers," with the technology's rise from "two winters" to "dominating the artificial intelligence community" driven by huge datasets like ImageNet.
- AlphaGo Zero in 2017 is predicted to have generated moves surprising to experts by learning from "zero information," while the 2018 task for AI includes winning in a multiplayer tournament setting in poker where heads-up is already solved.
- Long-term expectations include a continued need for system transparency and communication, even if systems work "well enough," while attentional mechanisms are planned to be used to model how humans look around an image to steer network interpretation.