Hacker News Reader

Comment by manux | Hacker News Reader

She makes a number of fair points, but:

> Can you really explain eight self-driving car trade secrets in 45 minutes?

Maybe I'm fooling myself, but I think I can do that in 45 seconds.

  - hey what's your learning rate scheduling for module X?
  - we use an EMA with coefficient 1e-4
  - oh great, and for Y, do you do constant strides?
  - no we use backtracking to figure out some sampling heuristic

etc.

View on HN

None of which would be useful for a judge or jury. Heck, I'm generally pretty decent technically compared to the rest of the population and what you wrote here is barely relevant to me.

Sure what I just wrote is technobabble, I don't have much self-driving car expertise, but it was just to make a point. From my experience with large scale engineering projects, especially ones with ML involved, the tiniest details can make the stupidest difference.

The 45 minutes she's referring to, is the 45 minutes that Waymo spent talking about it in court. Not 45 minutes that Levandowski spent talking with Uber employees.

If all of the models line up a ton, an expert witness could explain that these the chances of reaching the same tuning on all these choices via independent work is practically 0.

How long do you think it would take to explain that to 10 random strangers?

Not even ten random strangers. Ten strangers that are almost certainly not technical, since lawyers always try to dismiss smart people from juries?

As someone who has been on two juries, I object! Jury selection is complicated, and there are limitations to how many potential jurors can be struck. There are also multiple strategies that legal teams may adopt. In my last trial, the plaintiff made a very emotional case, but the defense presented a very logical case. There were two jurors who I would say were logically minded, and we were able to influence the rest of the jury, who were tending to stick on what I thought were irrelevant points. A lawyer friend of mine said that she has seen that one or two logically minded jurors can often steer the jury, and if she is able to put together a logical case, she wants to get some logical people on the jury. (However, being a public defender, she is not always able to put together a nice logical case. She has to do the best she can with what she has to work with.)

Aren't you proving cbhl's point? When she can't put together a logical case and has to rely on "this person has kids please don't send them to prison for this crime they obviously committed," she's going to be trying to keep off the most logical and articulate people at any costs.

> When she can't put together a logical case and has to rely on "this person has kids please don't send them to prison for this crime they obviously committed," she's going to be trying to keep off the most logical and articulate people at any costs.

Nah, people who support nullification based on perception of the specific case and offender circumstances are not generally less logical and articulate than others.

You obviously want to eliminate people who either see the law as inherently just or see it as important to uphold the law even when the result would be unjust, as well as people inclined to see punishing your client as just despite the circumstances. But there's no reason, in general, to think that logical thinking or articulateness are, in general, going to be good correlated for those moral inclinations.

chbl wrote: "Ten strangers that are almost certainly not technical, since lawyers always try to dismiss smart people from juries"

The point is that smart people are not always dismissed. Different legal strategies work better with different jury selections.

> lawyers always try to dismiss smart people from juries?

Lawyers don't “always try to dismiss smart people from juries”, or even try to identify smart people.

They do try to remove people that they expect will be more likely than average to be unfavorably inclined to their client or specific arguments, or that they expect to be a wildcard because they will be unable to restrict themselves to the courtroom presentation.

Neither of those, in general, is particularly correlated with intelligence.

Parameters are probably the lowest hanging fruit in any machine learning optimisation, being a trade secret is a big stretch.

Also use LBFG so you never worry about learning rate.

"Also use LBFG so you never worry about learning rate."

Literally never happens in practice.

Hyperparameter optimization for neural networks can be insanely expensive though[1], to the point where there are even startups (e.g: SigOpt) that do this. [1] https://twitter.com/beenwrekt/status/961263599674150912

the tweet you link to is architectural search, or figuring out what the neural network should look like (along with its parameters).

If self-driving car ML is "data scientists replacing a driver with computers", architecture search is computers replacing data scientists. $1.2M is roughly 3-4 data scientists' annual compensations at a company like Google or Uber.

Finding hyperparameters for an existing architecture is much easier because you would typically deal with a much lower cardinality in hyperparameter space. SigOpt is helping with that. Google has an internal solution that is able to take advantage of parameters within their job scheduling infrastructure (they actually have a bunch of them, the one I'm talking about is called hyperband [0]). CERN, as far as I know, was using hyperopt [1]. Avant uses an open-source version of SigOpt that we wrote in 2016 called loop [2].

[0] - https://arxiv.org/pdf/1603.06560.pdf [1] - http://hyperopt.github.io/hyperopt [2] - https://github.com/avantoss/loop

Before this gets inundated with people making comments about Googler data scientists making $400k, these are folks with PhDs in math or statistics, very accomplished academic literature, leaders in their fields. These are not people messing around with Haskell on the weekends and leaving an insurance firm's IT department to go to Google.

They also don't make $400k – an engineer that costs $400k doesn't get that in take home pay.

There are enough non-executive engineers at Google making half a million or more base/cash comp that you could fill a room with them. That's not the point.

Entirely not true.

Yea, and my point was that this can be expensive regardless of the semantics of whether or not architectural degrees of freedom are hyperparameters.

no one uses LBFG in deep learning or deep reinforcement learning, which is what self driving cars use

So, I am sure this will get downvoted, but thi is not my field - but as a curious person.... can you please explain in-long-hand-but-still-like-I-am-5 what your comments means?

I dont understand it, but I want to.

coursera has a great course on self-driving cars. the only prerequisite is a college course in linear algebra and some python knowledge. If you don't have either of the two the course is still very much approachable - you'll just have to learn two (or three) things at the same time, so expect to double (or triple) the time you sink into it.

It's worth it

Reply on news.ycombinator.com