dbecker

I've had a few DS jobs. And I've been a hiring manager at two of those jobs. I can confirm that it's roughly the wild west still.

Adversarial noise isn't specific to deep learning models. Most model working on high-dimensional input will confidently misclassify noise with high confidence if you construct the noise right.

The original "adversarial pixel" paper demonstrates this with logistic regression.

https://arxiv.org/pdf/1412.6572v3.pdf

This is primarily a fun toy problem.

It uses a Raspberry Pi and ~50 lines of code. So I don't think anyone should expect it to do something that's impossible with other approaches.

I don't have great links for this, but for something less technical, you might look at the blog posts from Kaggle competition winners. Here are a couple examples

http://benanne.github.io/2015/03/17/plankton.html http://blog.kaggle.com/2016/04/13/diagnosing-heart-diseases-...

or check out what's available in a deep learning library like keras http://keras.io/preprocessing/image/

Sorry I don't have better academic references.

I've seen many examples where networks are trained on thousands or tens of thousands.

The most common example is digit recognition with the MNIST dataset. This is a common problem given to beginners, and even many beginners to CNN's achieve human-level accuracy. That dataset is in the tens of thousands.

there should be a push towards adapting CNNs to calculate/predict how the object might look under different conditions

Data augmentation like rotations, horizontal flipping and random cropping are a widespread practice.

For a faster GPU, that allows me to iterate more quickly, I'd be more than happy to pay an extra few bucks an hour.

Keras is significantly higher-level (in the sense that Python is a higher-level language than C). I find it much easier to write in Keras, and think I'm less likely to make mistakes.

There are a variety of other wrappers for Tensorflow (skflow, tf-slim) that may someday be released as part of the Tensorflow project. But Keras is far more mature than those right now.

I wonder where Salesforce is going to integrate MetaMind's technology. From afar, the overlapping use cases aren't obvious.

I recently used indico.io for a project, and I'm a huge fan of what they offer.

From my brief interactions, they seem like awesome people too.

I briefly taught at a bootcamp, but no longer have any affiliation with them, and I don't have any incentive to be dishonest one way or another.

They aren't perfect. There are cheaper ways to learn. Instructor quality is uneven, and the tendency to hire their best students as instructors means that many instructors never see real-world applications.

But the cynicism about their intentions is misplaced. The people who run bootcamps are generally passionate about changing students' lives.

The students I know all said the bootcamp was a great educational experience.

Rather than pump up the employment numbers with fluff, our bootcamp was careful not to inflate numbers (for instance, not counting the people we hired internally towards the placement rate).

That's just my experience with one bootcamp, and I'm not saying it would have been right for you.

But that level of cynicism is totally off-base.

When Casadevall (and others) argue for more ethics and epistemology in the first year of a PhD program, they should be transparent that they would take away time and attention for much of the material currently taught in the first year of PhD programs.

I'm not arguing that it's not a worthwhile tradeoff. I just wish we were more transparent about these tradeoffs when discussing what we add to a curriculum.

I'm grateful for everything Continuum has done for the Python community. They seem to have such an amazing team, and I'm glad to hear this.

I recently switched from a company using HipChat to a company using Slack, and the differences seem mostly cosmetic and immaterial. Are the differences between these various chat programs big enough to affect user experience or productivity? What distinguishes one from the next?

I'm trying to figure out if I can keep my current car until autonomous vehicles become broadly available. Anyone here have an informed opinion on when that will happen?

You are taking that quote wildly out of context. That quote refers to the use of water as part of the drilling process, not the contamination of water-sources by fracked wells.

People in the industry seem to universally agree that there are some sloppy operators.

I've been trying to follow the Keras development, and didn't even realize there is a mailing list. Could you post a link to it.

I'm surprised to hear this. I think most techies hate how crowded the .com space is (especially with domain squatters).

I'm thrilled to see a dramatic increase in namespace.

Most supposedly anti-consumerism arguments I hear imply we should devote more attention to our purchases.

But, if you want an identity that isn't based on consumerism, it seems we should devote less attention to your purchases rather than more.

Honestly, I don't care about toilet paper. If I can hit a button and move on with the rest of my life, that seems better than spending time shopping for it.

This blog post by Joel Spolsky makes a case for teaching low-level stuff that I find very compelling.

http://www.joelonsoftware.com/articles/fog0000000319.html

Someone who hasn't thought about sorting algorithms is likely to implement a "Shlemiel the painter" solution without even realizing that their problem is fundamentally sorting related.

There are applications of job search in both micro and macro (I happen to have written a PhD dissertation involving a micro approach to job search).

The treatment in the book we are discussing is decidedly macro, and looks generally similar to the treatment in a previous Sargent book titled "Recursive Macroeconomic Theory."

Can you point to any evidence to support these opinions/claims? (I don't mean that to be snarky, but rather that I would find it more interesting and compelling)

It's very rare for anyone to commission studies like this in the social sciences. Rudolph is a professor at Korea University, and professors do research without extra compensation because it is part of how they get tenure.

Zuckerberg's bio on Wikipedia says that he is the son of a dentist and a psychiatrist. I'm sure there are plenty of Americans with wealthier or more connected families, and none of them have succeeded with their startup the way he has.

I've used both Domino and AWS.

AWS isn't hard to use, but Domino is much easier and more streamlined for analytics.

My recollection is also that AWS rounds charges up to the hour, whereas domino does not.

I didn't know spinning up analysis in the cloud could be so easy and frictionless before trying domino.

So, if I start with two devices holding lists, and I place the union of those lists on each device, what is that called?

Do you have a reference for the technical definition of sync that this fails to meet?

I find it hard to believe you were among the first to use this phrase for social media and/or advertising services.

It seems in fairly widespread use even for that use case.

The fact that the price of something is high does not mean the elasticity is low.

Even if we agree that the total cost (including time) of acquiring and reading a book is high, that does not indicate that demand is inelastic.

If you include the value of your time, the library is a more costly option to acquire a book than Amazon for many affluent readers. And most libraries have a small fraction of the books one might want.

The fact that people are choosing an efficient way to get the books they want is hardly evidence of inelastic demand.

Indeed. My original comment was meant as a joke, not as a serious mathematical claim.