emcq

26 ms·

241.
by emcq 11y ago
Ah bummer, looks like my response was cut off :( If you can bundle all your dependencies into a static file you're right, C/C++ is actually better as you have no install requirements. However in terms of maintaining your dependenc
242.
by emcq 11y ago
Do you believe there is a limit to what models we can train with layers of differentiable units?
243.
by emcq 11y ago
It's refreshing to hear other people think the same thing. Most of development time isn't spent physically going code. By picking C you get a high performance layer which can save a lot in infrastructure headaches and has the best
244.
by emcq 11y ago
Or the window manager should zero things out? It seems more consistent for the OS to be responsible for this then say the driver which also does things like supporting OpenGL and GPGPU usecases.
245.
by emcq 11y ago
The article you mention names RichRelevance, but there are others who implement Thompson Sampling or other forms of Bayesian Bandits, such as SigOpt [0] and Dynamic Yield [1]. Various adtech companies also use it underneath the hood. [0] h
246.
by emcq 11y ago
I like the article, but not your title. It implies that this trend of I/O becoming highly performant has occurred recently, when in fact has been observed and studied for quite some time [0, 1]. Even before SSDs Gigabit ethernet was sa
247.
by emcq 11y ago
Reinforcement learning isn't magic pixie dust that suddenly solves AI. One thing we often do is assume for some compact subproblem that AI will offer massive optimizations relative to our current human baseline. However it often turns
248.
by emcq 11y ago
Thanks I was wondering how he would stay unemployed with his reputation and prior experiences at Apple :)
249.
by emcq 11y ago
A bay area resident who has several thoughtful writings and talks about human computer interaction and more recently a piece on global climate change. AFAIK he is unemployed so he has lots of free time to work on personal projects which may
250.
by emcq 11y ago
In an optimized system the bottleneck will be I/O. However elements like serialization, many function calls, and extraneous memory copies chips away at performance to the point it matters and its difficult to address these problems dir
251.
by emcq 11y ago
I think he was using a new acronym for Dynamic Memory Allocation rather than Direct Memory Access. I also was confused the first time I read it.
252.
by emcq 11y ago
One perspective is that you are making 20x the average household income but keeping 1/2. Not too shabby. However I agree that it seems ridiculous that income is taxed so much more than capitol gains and likely provides much more positi
253.
by emcq 11y ago
I would love to hear what people think of mxnet. I haven't had time to do anything real with it but at a high level view seems super fast and flexible.
254.
by emcq 11y ago
I'm excited about this too, but the article suggests this is an industry wide thing; IBM is doing this with Xilinx, Qualcomm is experimenting with ARM stuff (not sure how this is different than the Zynq), AMD also with Xilinx. If there
255.
by emcq 11y ago
Not if you're sitting around waiting for data stored on some DB or distributed file system. FPGAs don't do much to address that. However there are certainly cases where you could design very lean systems where you could win with a
256.
by emcq 11y ago
I have a friend at space x who can earn performance bonuses relative to the improvements within the company, such as reduction in cost or construction time. This feels fair; if you find a way to save a million dollars a year you should get
257.
by emcq 11y ago
That's true, there are certainly many optimization objectives computationally intractable, or perhaps too abstract to be useful for learning. However, I would argue the prior of Bayesian modeling can be just as nebulous and computation
258.
by emcq 11y ago
Even when you specialize, you aren't privy to the same sort of functionality or information exposed to the inner circle of lead game developers, so you have to spend a lot more effort to gain the same insights.
259.
by emcq 11y ago
Perhaps Gary Marcus said it best himself when critiquing Numenta, yet another brain inspired company [0]: [Brain inspired] models are arguably closer to how the brain operates than artificial neural networks. "But they, too, are oversi
260.
by emcq 11y ago
This is perhaps a bit philosophical, but assuming we had a "meaning maximization" function, what stops us from writing it as a loss function and using our current supervised machine learning frameworks?
261.
by emcq 11y ago
I wonder how much of this was motivated by Nervana System's successes building their own assembler for Neon [0], or if perhaps this work started before it. In either case more openness for low level hardware optimizations is always wel
262.
by emcq 11y ago
While a good point that simple is relative, if you specifically look at deep neural networks we don't understand how training a non convex function converges with gradient descent - the fundamental component to create a usable model. I
263.
by emcq 11y ago
On that note, one of the best books I've read for teaching the concepts is Learning From Data by Yaser Abu-Mostafa et al. I highly recommend it as a way to refresh or introduce the fundamentals: http://amlbook.com/
264.
by emcq 11y ago
Some interesting points but it isn't limited just to academia. You unfortunately see similar cliques in industry for things like databases, languages, MVC frameworks, etc which is rather thin knowledge that can be picked up quickly by
265.
by emcq 11y ago
Are you sure that's not just the adderall? Asking mostly as a joke, but I know a few people who use it specifically for these purposes.
266.
by emcq 11y ago
Both good points. Yann LeCun (et al) have developed CNN FPGA designs where they project future ASIC performance, but I'm not aware that this has ever been taped out, nor any power numbers released: http://yann.lecun.com&#x2F
267.
by emcq 11y ago
I agree that neuromorphic is often specious, but it's not fair to say that neuromorphic hardware fails if it can do non brain like computations more efficiently than conventional hardware. The thesis may be misguided at times but build
268.
by emcq 11y ago
To be fair, the goal of those new computing architectures is not to advance algorithms. That at best is a side benefit. Those algorithmic performance improvements seen in DNN have come from improved datasets and training. Numenta and IBM ar
269.
by emcq 11y ago
This is not very accurate. The true North chip places memory next to compute elements on chip. GPUs have most memory off chip. The true North hardware is specialized to neural net stuff, and doesn't allow general floating point comput
270.
by emcq 11y ago
Great points. While it's market is limited, I think Skully [0] is an interesting case of where AR can be more than just an entertainment device, in fact becoming a much more valuable safety device. [0] https://www.youtube.co

More ›