atpaino

I like the analogy to de-noising autoencoders; that's a good way of thinking about this.

> In the limit, you might even have a neural network generate increasingly harder types of grammatical corruptions, with the goal of "fooling" the corrector network.

Very interesting. I wonder how many constraints would need to be added to the corruptor model to ensure the corrupted sentence retains the same meaning as the original. Somewhat related to that, I've thought that a more basic curriculum learning setup could be deployed quite effectively here, and am hoping to try that out soon.

Yep, it definitely has room to improve. The work thus far has primarily been a proof-of-concept for the methodology used to generate training samples (i.e. starting with grammatically correct text and introducing errors). Next step is to try to include more high quality data, after which I may try out comment data from HN, etc. I think it would be interesting to see what the effect of somewhat noisier data like that would have on the model.

Thanks for the links! I'll have to try out some of these. The data is definitely the limiting factor at this point.

Nice post. However, I was kind of surprised to see the author list interpretability as one of the drawbacks of random forests:

> Another point that some might find a concern is that random forest models are black boxes that are very hard to interpret.

I generally agree that random forests are more difficult to interpret than linear models such as logistic regression, but I think they're still far more interpretable than more comparable non-linear models such as neural networks or SVMs with non-linear kernels. At the end of the day, a random forest is just a bunch of decision trees, each of which are very straightforward for humans to understand. Additionally, there are a number of straightforward methods available for assessing the importance of each individual feature in a random forest in aggregate ([1], section 15.4). Neural networks, on the other hand, result in models that are too cryptic to be meaningfully inspected by a human, and have more complex variable importance measures [2].

The relative interpretability of these non-linear models played a large factor in our decision to add random forests to our modeling stack at Sift Science, which you can read more about here: http://blog.siftscience.com/blog/2015/large-scale-decision-f...

[1]: http://statweb.stanford.edu/~tibs/ElemStatLearn/

[2]: http://www.massey.ac.nz/~mkjoy/pdf/Olden,Joy&DeathEM.pdf

Author here, feel free to ask me any questions.

It seems like this would be much more useful if the front and back of the card were OLED screens, or something similar, so that you could toggle the appearance of the card. I feel like that would solve most of the problems people are suggesting here - i.e., it could replace driver's license, photo id's, wouldn't be a problem at bars that hold on to your CC, etc. Personally, I'm going to hold off on getting this until there's some functionality similar to what I outlined above, because currently this could only replace 2 out of the 5 cards in my wallet.

It's funny to see this on HN, because I currently have a copy of one of Dr. Poggio's papers in front of me. I've got to say, from what I've read so far I've gotten more excited by his group's approach than any other's I've seen. If you are mathematically inclined and interested in AI, I highly recommend reading through some of his group's papers, which are available on their site[1]. For those of you commenting that we don't have any theory on how the brain does its magic, this new initiative should excite you because this group is actually approaching the problem from a far more rigorous theoretical viewpoint than before. Specifically, their theory of neurons storing both sample images and common transformations on them fits rather elegantly with what we know about the brain, and seems like a promising route to matching humans' ability to "learn" a new object from just one instance of it.

[1]: http://cbcl.mit.edu/publications/index-pubs.html

Actually, this may be very good news for Kurzweil. In his last book "How to Create a Mind", he lays out a theory centered around a "pattern-recognizer" unit that is repeated throughout the columns and regions of the neocortex. In his book, he assumed it to be made up of several neurons wired in a specific manner, but if each neuron can do some hierarchical processing of its own then the pattern-recognizer might be reducible to a single neuron.

Biology may be very "unabstractable", but that doesn't mean we can't still learn enough about its high level behavior to create practical systems in its image. For example, it is unlikely that whatever structure or property of the brain makes it intelligent exists only at the molecular level, so we likely won't need to simulate the brain at that low of a level to create an intelligent machine.

Pure speculation, but I have to wonder if Sinofsky's firing had something to do with him not agreeing with this decision. Seeing as a key selling point for the Surface over the iPad is the inclusion of Office, and the Surface was Sinofsky's baby, I imagine he wouldn't like this decision at all. That, and one of the reasons Ballmer cited for Sinofsky leaving was his lack of cooperation with other senior leadership.

Maybe I'm misunderstanding what you mean by 'swiping the other direction', but when I swipe from right to left on a message it displays an 'Archive' button over the message.

Google must be all-over this for their Glasses. Seems like it would be a great solution to the input problem with them.

The only way I can really see a big-box chain like Best Buy surviving alongside Amazon and the other online retailers is if they were to focus heavily on training their employees to be knowledgeable on all the products they are selling, similar to Apple's "Geniuses". I think the original problem stores like Best Buy were created to solve no longer exists - namely, a _physical_ one-stop shop for all your electronic needs. Amazon is clearly better for this.

Where I see opportunity for Best Buy to differentiate is in customer experience. When the typical consumer decides they want to buy a new laptop, and don't want a mac, they basically have a couple of options. They can either research laptops on their own (through Amazon, Engadget, et al) or they can just drive to a Best Buy store and buy whatever the employee tells them. The thing is, the Best Buy option really isn't much better currently because their average employee isn't very knowledgeable on the laptops they carry.

For this to work, Best Buy would have to both pay employees more and spend more time/money training them. Not to mention the costs associated with re-branding the company as a whole. However, IMHO I think something this drastic needs to be done for them to survive.

Kind of tangential to the article, but I often wonder why facebook has not tried the "freemium" model at all. I'm not suggesting they do the typical freemium model, which would be to offer users the option of paying to get rid of ads. Instead, I think they could try charging based on usage (kind of like heroku). They would have to keep the service free for the majority of users, but I think they could get by with charging super-power users; I.e. the ones who can't live without the service. I know many people that would fall into this category, and I'm sure they would be willing to pay $5/month if it meant keeping their facebook.

As a college student and St. Louis native, this really excites me. I wish all the startups in these programs the best of luck.