tachim

This is false. People were thinking about and researching autonomous driving seriously since the 80s: https://en.wikipedia.org/wiki/History_of_autonomous_cars.

Thanks. Do you guys do any meta-learning based on information from optimization runs on the many instances that SigOpt runs on? E.g. if Alice uses Sigopt for a neural net architecture and Bob does too, is the information from Alice's run used to improve Bob's?

How does SigOpt compare to GPs?

Conferences certainly dominate journals, but most people publish their work on arxiv first, anyway.

Surprising that feed.ly hasn't been mentioned. It's like gmail for feeds, and it has all the arxiv categories prepopulated. My workflow is as follows: (i) check feedly every day, see ~20-30 new articles, (ii) skim all the abstracts in 5-10 minutes, (iii) mark 0-2 to read later in the day, (iv) mark rest as read, and repeat.

Differential GPS can hit ~10cm level accuracy.

> "They rely on optics to 'see' lane markings, road surface maps, and surrounding infrastructure to orient themselves. Optical systems work well in fair weather conditions, but it is challenging, even impossible, for them to work when snow covers the markings and surfaces or precipitation obscures points of reference."

Humans drivers are a pretty glaring counterexample.

"What we know is that the vehicle was on a divided highway with Autopilot engaged when a tractor trailer drove across the highway perpendicular to the Model S. Neither Autopilot nor the driver noticed the white side of the tractor trailer against a brightly lit sky, so the brake was not applied."

That's really disingenuous. There's no way to know for certain what the driver did or didn't see, but there are a lot more salient features for them to go by other than the white trailer, like the fact that the cabin has driven across the road or that the wheels of the trailer were square in the driver's lane. A much more likely explanation is that the autopilot set expectations too high, the driver was not paying attention (like in all the youtube videos of the system in action), and the autopilot failed and crashed into the truck.

Codalab is designed to solve exactly this problem: https://worksheets.codalab.org/worksheets/0x818930127c4d47de....

How does the design process for these reactors work? How much design is done in simulation vs experimentally and how high-fidelity are the simulations?

If the play is so standard and automatic then where does skill come in? Or just skill just correlate highly with how many situations one has observed before and potentially memorized?

0.1% accuracy increments correspond to 10 images in the testing set; they should be reporting standard error bars with those numbers.

Totally agree with this. The narrative shows up in more discussions than just startup vs bigco; it's also a very common decisionmaking framework when choosing first jobs out of college, when choosing projects to work on while at the company, etc.

It's pretty bad overall for net societal productivity. The main reasons this framework shows up so much is that (i) it's hard to really think through what kind of unique value one offers and to figure out how to leverage it, and (ii) it's tough psychologically to do something that doesn't optimize for the typical value function (prestige + income).

Check out https://www.codalab.org/ too. It's a more complete toolkit for reproducible computational research.

This isn't quite right. Moment methods (that rely on tensor decompositions) have a few problems:

(i) they have convergence bounds, but in practice need more data than we have available

(ii) they don't do as well as EM usually, but using them to initialize parameters for EM sometimes does better than EM with random initialization schemes

(iii) it turns out variational methods can also be embarrassingly parallelized without losing much accuracy in practice

(iv) right now moment methods don't work for arbitrary graphical models

What's your email address?

Testelf is fantastic to use.

Homomorphic encryption is unfortunately still very far from being fast enough to be used in this context.

You can click on "Create Blog" at the bottom of the "Write" menu at the top right of your home page.

Awesome! This looks super useful, I'll definitely give it a shot with my own apps.

The point of most encryption algorithms isn't to compress the data or remove redundancy, at all. In fact, most widely used algorithms are block ciphers that don't change the size of the data.

It's not the actual connection itself that matters. If your phone ever transmits a packet, it can be detected and associated with your phone's radio id.

Self-driving cars have much more lax sensing requirements, and the (on-road) environments they navigate around are much simpler than unstructured home environments.

The processors are basically free compared to the cost of the arms. For example, the barrett WAM arm costs about $250k.

They have many advantages, but afaik nobody's come out with an OLED screen that doesn't burn in after 10 min of the same color.

16 + High School

'near pittsburgh' doesn't work. I like the idea, though.