strebler

I'm no fan of the A.I. hype of today, but I can comment that deep learning actually makes the "contentious decision" easier to detect. With the "old way", i.e. jam some features into an SVM, there was no accountability / confidence. At least I've never seen a reliable SVM-based confidence score.

But, for example, confidence scores are actually a byproduct of running a modern CNN for image classification. This is very useful for "finding mistakes" (assuming you have 100% accurate training data, and other prerequisites). It even helps fixing issues in the training data, we use it all the time.

That being said, the big dirty secret that nobody's pointing out is: there has been no real "theory-level" advancement in AI in the last 20 years that have borne fruit (at least in computer vision and neural nets). The models are just bigger (and much better) with some slight tweaks (dropout, etc). It's hardware that's made that possible. Academics have done a thorough and wonderful search of the model space and have some great finds. But I think we still need a few solid theoretical leaps before the singularity.

A standard Scanning Electron Microscope can do down to 0.1 nm, so this isn't really letting us see anything we couldn't before. SEMs are regularly used to reverse engineer chips. The main difference is that the one in the article is non-destructive, whereas with a SEM you have to strip the sample (and do imaging) layer by layer.

Japan is sort of the Germany of Asia. Both are very detail-oriented people, well known for craftsmanship, significant amount of global exports, their trains are always on time, etc.

That is absolutely not true. The minimum HOURLY wage in the lowest paid city in any province in the country is above 10 RMB / hour.

https://en.wikipedia.org/wiki/Minimum_wage_in_China

Slow news day?

I understand Prof. Ng's contributions to the field. That being said, AI is so full of hype at this point, I'm not really getting why a placeholder page warrants a whole article, which frankly has a similar level of information in it.

Wake me up when there's a demo to try.

That's a great point. Fundamentally, if you look at something like a CNN, what it's really doing is producing a feature descriptor based on the input image. One can easily use that feature descriptor in a classic SVM, alongside (or instead of) SoftMax.

Well, it's hard to get good advertisers when the ROI isn't there.

It's not something you'll hear unless you're in the know, but there is a pretty high level of animosity towards Pinterest in most retail marketing departments. I love Pinterest, so it was surprising to learn when we started this work, but it boils down to: 1) traffic from Pinterest ads show very poor quality / engagement. 2) Pinterest has not been a good partner - they've actively "snubbed" a number of big labels. There are even a few major retailers who "deleted their Pinterest". 3) Organic Pinterest basically doesn't work, so they won't even put an intern part time onto managing their Pinterest account. Even one of their top retailers, Nordstrom, appears to use a bot to post to their account.

When your ad dollars don't show ROI, people spend elsewhere. And if your main customers aren't engaging your platform, that's definitely trouble too.

Thank you for this! As soon as I saw the problem, I thought "use prolog".

My point was that neither AI group shared any play history at all, even when Kasparov asked specifically for it. If the AI is so good, what do they have to hide?

Lee Sedol won a game too, so perhaps if he had had similar level of play history as it had for him, he could have discovered and exploited its weaknesses.

Yes, I saw that he asked but didn't insist. As in "I'm not participating if the playing field isn't level".

It appears both were somewhat overly confident, so neither really pushed for this.

It does strike me that neither Kasparov nor Lee Sedol insisted to see any play history of their AI opponents.

Not that it would have changed anything, but it's kind of silly to go into any situation one side has the other's entire history, while the other side has nothing. Not exactly fair.

I notice the recalls are up to 2014 models so far - have they addressed it in any way or are they still just cranking out cars with this problem?

I had the same question about cameras dealing with those issues, but this article says "Products are embedded with tracking devices that pair with customers’ phones to charge their accounts". It sounds like this means every product has an RFID / other tag, rather than using computer vision.

Agreed! From my perspective (computer vision), I can think of quite a few Chinese individuals who are true leaders in AI at this point (often working for US companies).

In fact, given both Hinton and Bengio's being Canadian and LeCun being French, I would say US really in the game in terms of producing "AI grandmasters" (assuming that's their definition of grandmaster). Krizhevsky was studying in Canada. DeepMind is European.

From both the research and industrial perspectives, China's education system has well prepared students for the field of deep learning - the rigorous / analytic nature has proven a massive benefit when dealing with such large models and datasets.

You're probably right! It just states "12 particpants over a period of 20 days" on their site - hard to tell what that means.

I doubt they were acquired based on their datasets! The "Frames" dataset states 12 users over 20 days.

But if one of the next 12 users was a Microsoft executive who liked what they saw, well then that's an important data point :) Having Bengio as an advisor definitely helps too.

We have several products, each of which serves different departments within retailers.

The exact things we do depends entirely on which department(s) are licensing it. Basically, anywhere there's a product image (from their own inventory to mobile to social) and we can provide some kind of help, we do. Every department needs totally different things, so it varies quite a bit...but it's all leveraging our core automated detection + recognition + search APIs.

We're a computer vision company, we do a lot of product detection + recognition + search, primarily for retailers, but we've also got revenue in other verticals with large volumes of imagery. My co-founder and I both did our thesis' on computer vision.

In our space, the recent AI / ML advances have made things possible that were simply not realistic before.

That being said, the hype around Deep Learning is getting pretty bad. Several of our competitors have gone out of business (even though they were using the magic of Deep Learning). For example, JustVisual went under a couple of months ago ($20M+ raised) and Slyce ($50M+ raised) is apparently being sold for pennies on the dollar later this month.

Yes, Deep Learning has made some very fundamental advances, but that doesn't mean it's going to make money just as magically!

Those are great links! But the 2008 Hinton paper would not be considered deep learning, it is classic neural nets. It makes no mention of CNNs or GPUs, which is what really got this all going back in 2012 with ImageNet / Krizhevsky.

The ImageNet paper is from 2012, not 2010. That's when the computer vision community really went "wow". IIRC, almost every entry in ImageNet 2013 was using CNNs.

Yep, I checked out at: "...rise of deep learning since 2013, more or less when Google’s X Lab developed a machine learning algorithm able to autonomously browse YouTube to identify the videos that contained cats".

First, this started in 2012. Second, it wasn't Google - it was when Krizhevsky et al published their seminal work. Realistically, Google was slow to adopt to GPUs at the time, which I understand even contributed to Prof Ng's departure. It was Baidu who launched the first large scale deep-learning based image search, well ahead of Google.

Google has certainly caught up, but nobody can say they started it (and be taken seriously).

I love this, I have been (passively) wanting to find something like this for ages (i.e. bash script, with minimal deps, to generate a nice gallery in HTML). Thank you!

One idea / suggestion I have would be to use Isotope for layout: http://isotope.metafizzy.co/ Then you'd basically be able to generate a mini Pinterest...

Yes I read the paper and I know what you mean, the media loves to jump to conclusions. Of course, they're also the reason I even found out about it. I'm super excited to read what comes from the future observations.

I don't think so. Dark matter is needed to explain the contents / behavior of galaxies we can see. This article is talking about galaxies we can't see.

On that note, here's some interesting / related recent news in terms of a galaxy we have a hard time seeing: https://www.washingtonpost.com/news/speaking-of-science/wp/2...

I'm not sure, but that sounds like it might work. The real problem is: good luck getting all hospitals in the world to agree on a rotated schedule of which antibiotics to use / not use and when.

I have heard Dr. Andrew Ng point out the "well known" fact (in the computer vision community) that nobody, including Dr. David Lowe himself, really understands why SIFT works either. So at least for those of us in the computer vision industry, this isn't maybe such a surprise.

Other studies show significantly decreased mobility and sperm quality as well (some in-vitro with microscope).

Anecdotally, I'm quite confident this was a factor with our 2nd child. With our 1st, I didn't own a cell phone and he was conceived literally the first try.

I bought a smart phone when he was 1, which I carried in my pocket. A few years later, we wanted another child - tried for 10 months with no results! Frustrating. I randomly found some studies covering this exact subject in mid July. I started putting my phone into my backpack. August, our 2nd was on his way.

For anyone who's having troubles, I would strongly advise to take the phone out of your pocket. It might even be best for everyone to do so - who knows the long term effects.

I had the same problem - first clear your cookies for WSJ.com, then google the headline. Worked for me.

We did it before Deep Learning existed. If we had to do it again today from scratch, I think we'd use an approach similar to yours. My comment about VGG was based on experience with VGG vs AlexNet in other projects.

Landsat7 is trivial to do this, it has a thermal infrared sensor which is definitely not present on the RapidEye constellation.

This is neat. We've deployed a 98+% accuracy satellite & aerial cloud detection solution for many years, so I would have the following suggestions:

-Why use AlexNet and not VGG (or Googlenet)?

-Make sure to train on clouds vs desert. There are a lot of instances where their spectral signatures are very close, depending on the satellite.

-Make sure to train on clouds vs snow. They are even more close.

-Dark clouds. This might not show up much unless you're working with the satellite vendor, but there are cloud formations where shadows of clouds project onto other clouds. Very difficult to deal with and NN may be well suited to it.

I would say since it's absolutely possible to get higher accuracy using older methods on this exact satellite constellation, there is definitely room for improvement. Just switching to VGG might even do the trick. But this is a great first step!