Apr 1, Sun
Unbelievably productive day, not in terms of my research output but of my input.
Teach at our study group: MLAPP(4.1 & 4.3)
Richard Murray's video on safety of autonomous systems
- Not a particularly interesting talk scientificly, but it does include lots of interesting history of how airplane designs uses redundancy to keep safe.
Amnon Shashua CVPR 2016 keynote: Autonomous Driving, Computer Vision and Machine Learning
- This is so great! In fact it's the best talk ever on self-driving tech. Several notable points:
- Deep learning is great because it can use context. All previous approaches are local feature extractions, and that can't give you anything like semantic segmentation.
- We need 3D bounding box of cars not 2D.
- Sensing and planning are intertwined. Sensing can be easily modularized into smaller tasks, classification / 2d->3d / bounding box etc. But to learn control, we need a function of Space(sensing x policy) -> Space(control). And actually there is also temporal dimension. I'm not entirely sure about this, but policy seems to be referring to high level abstractions (follow, pass, switch-lane...).
- Planning is negotiation -> use reinforcement learning to learn control. Q-learning is a very bad model here, because you can't assume others to be Markovian!
- Mapping is considered by him as a redundancy to sensing...
Apr 2, Mon
On a Formal Model of Safe and Scalable Self-driving Cars
As I was hopelessly captivated (lol) by the clarity and insights glittering in Mobileye's CEO Amnon Shashua's 2016 CVPR talk, the next logical stop would be to follow up their recent paper featuring a safety model for self-driving cars: "Responsibility-Sensitive Safety (RSS)". It's a fairly long paper (couldn't even finish it in the morning), but they formulate with math the following common-sense RSS principles:
- Keep a safe distance from the car in front of you, so that if it will brake abruptly you will be able to stop in time;
- Keep a safe distance from cars on your side, and when performing lateral manoeuvres and cutting-in to another car’s trajectory, you must leave the other car enough space to respond;
- You should respect “right-of-way” rules, but “right-of-way” is given not taken;
- Be cautious of occluded areas, for example, a little kid might be occluded behind a parked car.
Statistical Mechanics of Phase Transitions (Chapter 1-4 & 8-9)
- A very relaxing read into renormalization theory needed by my current research. This book is super super concise but clear. Highly recommended!
Apr 3-5, Tue-Thu
OK, very much in unproductive state right now, I did manage to work on some code development though.
I did read the cool book Other Minds: The Octopus and the Evolution of Intelligent Life. Besides a fun read, it struck me on what we can now learn from the octopus neural systems simply because it's so much different from our centralized ones.
Apr 6, Fri
I continued reading Other Minds, Chapter 1-2.
Apr 7, Sat
I redid my taxes. That is a big deal... So it's a great excuse for not doing reading.