Autopilot AI job posting has some yummy hints

[strangecosmos]

Autopilot Internship/Co-Op (Summer 2019)

Role

As an Autopilot AI Scientist you will perform research and development to advance the state of the art in technologies enabling autonomous driving, research and develop algorithms for complex tasks like full scene understanding, scene prediction, planning and sequential decision making. Devise methods to use to enormous quantities of lightly labelled data in addition to a diverse set of richly labelled data.

Responsibilities

• Individuals in this role are expected to be experts in identified research areas such as computer vision, artificial intelligence, machine learning, and applied mathematics, particularly including areas such as supervised learning, graphical models, reinforcement learning, optimal control:
• Develop state-of-the-art algorithms in one or all of the following areas: full scene understanding, multi-modal data processing, learning for planning and decision making, etc.
• Implement them to run with real time performance in autonomous vehicle production environment.
• Creative ways to use complementary sensor data, and offline processing to create ground truth for training the algorithms.

Backups:

• archive.org
• archive.is

 

[strangecosmos]

“enormous quantities of lightly labelled data” makes me think imitation learning. The light labelling could be the driver’s input to the steering wheel and pedals. The data could be the mid-level representations (or metadata, i.e. 3D bounding boxes and such). Since HW2 Teslas are driving something like 12 million miles a day currently, the amount of lightly labelled data that Tesla could collect is enormous.

Can anyone think of other explanations? Maybe it could be used for supervised learning for perception. Like, if people don’t break for this metallic object the radar detects, it’s probably just a soda can or something else benign.

“reinforcement learning” and “learning for planning and decision making” makes me think Tesla is at least experimenting with using reinforcement learning for the actual driving of the vehicle (not just perception). We’ve known for ages that Tesla is using machine learning for perception, but it’s still unclear what is being used for taking actions. Or what is being developed.

I learned today that Andrej Karpathy interned at DeepMind for a summer and worked on reinforcement learning there. He also apparently worked on reinforcement learning at OpenAI. I thought he only did computer vision. So that’s interesting.

 

[strangecosmos]

Not related to Autopilot, but I found a job posting for a machine learning engineer at the Gigafactory.

Two intriguing tidbits:

Use policy gradient based deep reinforcement learning methods or online optimization planning methods such as model predictive control, to develop, simulate, and implement real time control algorithms that minimize the energy costs, equipment downtime, and setpoint error of systems with highly dimensional continuous state and action spaces.

Cool. Reinforcement learning for robots.

Leverage enormous image data sets to develop deep learning visual inspection architectures in at least one of the following areas: object classification, object detection, semantic segmentation.

Does this mean there are cameras on all the robots?

 

[Bladerskb]

Here's is a AP Internship for Fall 2017 posted in Aug 2017 with the same descriptions so i wouldn't look too much into it.

Autopilot Internship (Fall 2017) | Tesla | Tesla

What's interesting is that I have been unable to find a single Tesla AP engineer that has worked on RL or has worked on RL while at Tesla on their resume.

On the other hand its easy to find for others who we know are actively using it (BMW/Mobileye).

 

[ZeApelido]

Amnon is that you?

 

[croman]

Amnon is that you?

Don't flatter him.

 

[strangecosmos]

Post by Andrej Karpathy from May 2016, about a year before joining Tesla:

Deep Reinforcement Learning: Pong from Pixels

He also tweeted this in February 2018:

Andrej Karpathy on Twitter

 

[Bladerskb]

""Deep Reinforcement Learning Doesn't Work Yet" https://www.alexirpan.com/2018/02/14/rl-hard.html … great read, hits a lot of points I've also come to realize over last ~2 years. 70% is a vast understatement."

I think this puts an end to the;

1. Collect billions of miles of data in a month
2. Run through RL simulator for 10 billion miles in a week
????
Profit... Release Level 5!

Even the work on Go and Starcraft took years of hard work. To think that the greatest under-taking of mankind will be solved in acouple weeks is bizarre. I guess you have to believe that if you want to keep believing whatever Elon keeps saying.