Conversation

W/ humans, quant analyses suggest qual explanations which are criticized to refine quant methods; DNN opacity prevents this meaning++ cycle.

Quote Tweet

Sarah Jamie Lewis

@SarahJamieLewis

· Jan 17, 2017

Replying to @SarahJamieLewis

Just take a look and now imagine that 99.6% certainty is your buying algorithm, or your fraud algorithm or your traffic scaling algorithm.

4:46 PM · Feb 12, 2017Twitter for Mac

Replying to

I suspect better tooling and logic-based methods like this will make them a bit less opaque w/ time:

Quote Tweet

Bharath Ramsundar

@rbhar90

· Feb 12, 2017

The new ReluPlex method sounds like it could become genuinely useful. Proves realistic aircraft avoidance deep net is safe!! twitter.com/goodfellow_ian

Andy Matuschak

@andy_matuschak

Feb 12, 2017

Yep, I think improvement is possible. It’s less clear to me if it’s possible to climb this hill to authentic explanations.

Show replies

Replying to

Quote Tweet

· Feb 11, 2017

Deep learning: statistically generating a whole lot of if/elses but without access to the generated code.