slashcom

If you're well experienced with deep learning already, maybe only the last 2-3 minutes on the future is worth watching.

It will take at least a few more years before submissions are allowed in Chinese: reviewers can't be assumed to know Chinese. The momentum is there though.

Fake ones like this make it off the $800 publishing fee. Since it's "open access" they just put the PDF online and call it a day. $800 isn't bad for a few hours of proofreading and updating one web page.

That said, it's probably reasonable snow trucks get extremely high priority.

If it supported, then the major difference is GPU memory, which limits the size of the network you can train. The newest models are faster than some 1-2 year old ones, but older hardware does the job fine.

I did not read it very closely though.

non-SV male median: 97000

non-SV female median: 90000 (92% of male)

SV male median: 137120

SV female median: 99187 (72% of male!)

Unless there's some external factor here, things look pretty damning in SV...

And simultaneously, use all the emotions to train a ridiculously good sentiment analysis system.

They're, unfortunately, not based on intuition. Just statistics.

Basically, the machine plays many, many random games. The more winning games which play a particular stone, the more valuable that particular position is. Then the position with the highest value is chosen.

This alleviates the need of brute force search (which is just too large for Go).

As far as I understand, most of the more successful AIs use UCT for the general game, but then fall back to heuristics and brute search for small, local conflicts (like if forced some stones to be played until death) and counting stones.

Sensei's library has some nice high level information written about the topic: http://senseis.xmp.net/?UCT

More recently, DeepMind (and someone else independently at Edinburgh, I believe) has published a couple papers about training a neural network to play by teaching it to predict the next move of pros given the board state. These use modern computer vision techniques and data. Last paper I saw, this did much better than GNU Go (which doesn't use UCT, and is very weak), but still not as good as monte carlo based methods. There has yet to be a combination of the Neural Network and Monte Carlo methods, but they're quite complimentary.

I believe we'll see that combination in the next year or so, and this will nearly close the AI-human gap.

Edit: Monte Carlo. :P Hooray for autocorrect.

Distributional vectors are a proxy to word meaning. Words with similar vectors have similar meaning or are semantically related in some way. But usually, you just measure similarity by a single number: the cosine similarity between the two words' vectors.

This number can tell you words are related or not, but it can't tell you how they're related [1]. There's been a good deal of work in automatically identifying that "ship is-a boat" (which is called hypernymy) or cats and dogs are unrelated animals (cohyponomy), but it's still being perfected.

But it is useful. Words that are similar in meaning can be treated similarly. As a bad example: maybe I know that "anger" has negative sentiment, but I don't know what sentiment "furious" has, but I can infer it probably has negative sentiment since it's so similar to "anger".

[1] There's a good deal of evidence that words that have high cosine similarity are more likely to be cohyponyms.

The editor picks are usually extremely good, and if I didn't read them already, I know that I should.

I specialize in word representations and using them for various tasks. Word Similarity prediction has been used as a basic first evaluation for many years now, with analogies becoming an additional standard task in the past couple of years. But it's worth noting that word representations have a LOT of open parameters (which model? how many dimensions? do I remove stopwords and low frequency words prior? do I use a bag-of-words context or a syntactic context?).

The optimal parameter choices for one task are very frequently not the optimal parameters for another. While there are usually "reasonable defaults" for when you don't want to optimize everything, a solid standardized approach risks vastly overfitting to one task, possibly at the expense of more useful tasks.

http://www.nsf.gov/statistics/seind14/index.cfm/chapter-5/c5...