chimtim

Interesting to see the speech, vision API among others. I wonder what this means for metamind, clarifai and many other startups offering an API for doing something close?

This is a great experiment for computer vs computer simulations. When I played FIFA, I tried to buy cheap players with properties that worked well with my playing abilities. For me, at least until FIFA14 (did not try beyond that), a player's speed had the most net positive in my scoring ability and I used to buy the fastest yet cheaper players instead of buying a really skilled but slow players[1].

[1] http://fifabeast.com/top-50-fastest-players-in-fifa-14-overa...

AlphaGo can be beaten. It uses reinforcement learning so it will perform the set of moves that in the past led to its win. So predictable. Sedol just needs to take control and make it play in a predictable fashion. Also, perhaps play obscure moves that AlphaGo wouldn't have trained on. Perhaps next year's Go winner will have a PhD in computer science.

Another way to look at it would be that best-paper awards deserve much LESS credit than they are given. E.g. MapReduce did not get a best paper award but has been more influential than other best paper award winners.

"This year, I'll teach my simple AI to recognize patterns. I'll train it to recognize my voice so I can control my home through speaking. I'll train it to recognize my face so it can open the door when I'm approaching, and so on."

The announcement is slightly disappointing, in the sense that face and voice recognition are fairly solved and there is code on github that already allows you to achieve this. I was hoping for more, like at least an AI engine that can scan his shirts and recommends what to wear but I guess he does not have that problem.

I believe the Chess/Go engines are not really AI. They are operating on a set of rules written in code. It can compute and search the moves faster than a human. It can also remember a longer chain of moves better than a human can. But all those rules have been fed via code and it is just exploring that state space.

Only recently researchers have started to expose chess to reinforcement learning (i.e. a true AI engine that learns chess on its own and beats humans). But the existing commercial engines that beat humans (like stockfish) are anything but AI.

what fault tolerance does spark give you in this scheme? It cannot look into TF progress and checkpoint all state. Using Spark with TF, seems like an overkill -- you need to manage and install two framework what should ideally be a 200 line python wrapper or small mesos framework at most.

A good measure of any of the techniques are the results. We see record object recognition and speech recognition results. How many of these results are from Numenta or IBM neomorphic chips? Most successes have been on deep learning architectures over GPUs (and large datasets). While these fancy architectures may have their applications, they have nothing to do with recent advancements in last 3 years.

This is not really true. ML applications in general do not scale linearly. There is the systems (scaling) overhead, and then there are the algorithm payoffs, which start to diminish depending on the algorithm. If they went from 4 to 8 GPUs, the only thing they can guarantee to double is how fast they burn power.

Is this a machine with GPUs? You can order a similar looking design from acmemicro since last 4 years atleast?

Thanks a lot for your hard work! This is amazing work. I was wondering if you had any sense of which one of the models works best (or when one is better than another) -- Stanford, Toronto (Show, attend, tell), and Google (Show and tell) ?

IMHO, Google released Tensorflow because AI is currently being driven by research, and researchers were mostly writing code for Torch that is used at Facebook. So FB folks were enjoying lots of new algorithms, benchmarked against their systems. Google open-sourcing TF shows that benefits of open-source outweigh the disadvantages of being closed source. Even if the future is data and not code, companies will not open-source their code if they have nothing to benefit.

Appears to be only CUDA from the code. Also no distributed systems support released. Also, I cannot get it to work since this morning.

Do 10% speedups really matter for its success, if both libraries are ultimately calling into CUDNN? I do not think so.

Right now, Torch/Caffe adoption is really good with lots of research code being released for it (even from within Google DeepMind). So it will be interesting to see how this catches up amongst the research and developer community.

My comment and the article are about AI research specifically, where there is lots of exciting new research due to the success of deep learning. And since the area is new, there is lot of low hanging fruit and challenging problems, which researchers are eager to solve and publish, and as a result Apple is suffering.

So it does not matter how Google research worked earlier or if other areas of Google research continue to wait for a 5 year timer.

This is not how research works or a new fundamental technique is discovered. Any result needs to stand the rigor of peer-review before the said breakthrough can be conclusively proven (more unbiased minds trying to find the limitations of the approach and applying the said approach to various problems). Also this process is slow, and requires many PhDs, students building upon one another's work.

Also, going from 96-99 may not be really revolutionary, what really matters is how you got there e.g. was new research needed to detect new unicorn classes or scenes, or if it was simply solved by great engineering, more data, etc.

My guess is RankBrain is the personalization piece that operates on user data (location, history, etc.) while PageRank is the search index piece that operates on web data (web-pages, trends, etc.).

MSR has published lots of great papers but it has not yet replicated the success of PARC or Bell Labs by any means. I am not sure why this has happened, but it has been pretty obvious to everyone in the research community. Google, which came in with a research agenda pretty late, has more or less single handedly pushed the boundaries of industrial research in the last decade.

Hand-selecting a bunch of famous folks results in IAS Princeton effect which has the criticism of producing nothing significant. Usually the approach of hiring a towering/inspiring figure with an emerging research view works well (like FB did). They bring along a bunch of rookies and experts, and set up research groups quickly and rather smoothly.

This was a fascinating read. I found the initial bit interesting that other models were ahead because they simply had more raw computational power, though it later adds that there may be other subtle differences.

VLDB accepts 150 papers and SIGMOD perhaps another 150. This is just top tier. I am pretty sure science can live without about more than 50% of those papers.

I disagree with the tone of the original comment. But I do not disagree with the sentiment. Just having a large installation does not make it interesting. However, Google's large systems almost always push the boundaries of science -- MapReduce, GFS, Spanner, Distbelief and have been a joy to read.

Static analysis tools are great for simple bugs. However, most compiler tools already fix these simple bugs. If there is a class of bugs not addressed by a compiler, most developers have a script to catch these. It may not find the harder to detect bugs (such as those arising out of nested calls), but it gets the job done. The complex tools pushed out by the static analysis guys usually has a high false positive rate that at first becomes annoying to use and later is completely ignored.

As for more complex bugs, most developers are aware these bugs exist but do not want to fix them immediately. The reason is on many occasions, these bugs represent some bigger problem in the code base that requires significant re-factoring. And applying the quick fix as suggested by the code analysis tool simply buries these problems instead of fixing them the right way.

This looks cool but I'm somewhat skeptical. I would be more interested in seeing what problem the system solves better or decently (say even MNIST) rather than how it was built using memristors.

There is a lesson from IBM trying to mimic a rat's brain -- that is you try to solve a problem rather than just burn power.

There is another paper, XStream from SOSP 2013, which ran a facebook sized graph on a single machine. I wonder what is "hot" in this HotOS submission?

If the dataset and/or computation fits in your laptop, why would you use a cluster framework?

If you want to use multi-core, why would you not use the pthread library instead of using Spark, GraphX etc. The authors never show a pthread comparison.

The article shows just one algorithm for toy datasets. For certain algorithms such as stochastic gradient descent, multiple-cores can process data in parallel and go through the data very quickly. Again, if everything fits in memory, doing a gradient descent on entire dataset will be much faster and give a better quality result. This fact is pretty much well known to end-users i.e. folks who actually try to solve big-data problems.

However, most papers use small datasets like twitter or MNIST because their convergence behavior is well-understood (rather than to demonstrate scaling).

O(N^2) is not exponential but just polynomial. O(2^N) is still exponential. O(N!) and O(N^N) are faster than exponential but not that much; O(N^N) grows slower than O(N!)[1].

Hyper-exponential is most likely double exponential O(a^b^N). There is an exponential growth of apps on these platforms and another exponential growth of users on these apps.

[1] http://en.wikipedia.org/wiki/Stirling's_approximation