I don't think MobileNetV2 is designed to train on GPUs - according to this https://azure.microsoft.com/en-us/blog/gpus-vs-cpus-for-depl... MobileNetV2 gets bigger gains from GPUs vs several CPUs than ResNet. You could argue the batch size doesn't fully use the V100 but these comparisons are tricky and this looks like fairly normal training to me.
It's pretty surprising to me that an M1 performs anywhere near a V100 on model training and I guess the most striking thing is the energy efficiency of the M1.
Could you say a little more? I think I understand the "scaler" and it's how I learned the scales and how I practice, but I'm curious what the pentanizer is suggesting to do. Picking a root and an interval and finding it all over the fretboard?
One of the coolest parts of teaching these classes is how awesome the people are that show up. The engineers that want to learn new things mid career are exactly the kind of people I want to work with and hang out with. I think there's a real opportunity for more classes like this.
I really appreciate the author's critical analysis of this correlation presented as "fact" by Radiolab and I love how Hacker News and other blogs take these types of scientific findings and dig in for the truth. I think the PNAS paper refutes the original conclusion pretty thoroughly - I wish the Nautilus author would just explain that.
I don't think we should dismiss effects just because they seem really large (as the Nautilus author claims) but I do think that it's incredibly irresponsible of Sapolsky and Radiolab to be uncritically citing a study that looks like it was debunked in 2011.
I also think it's strange that the author cites the SJDM paper which is much, much less convincing, claiming that it refutes the original experiment. It looks to me like that paper just shows that by simulating a non-random order of parole requests they can create data that looks like the original experiment.
I love that Hacker News posts these things and people go through and analyze the papers. No one outside of the specialized field could possibly have time to analyze all of these papers but they clearly have implications that matter for everyone. I wish that popular science shows would do a more thorough analysis of these results on their own.
I love learning math from books, I feel like being forced to do the visualization in my own head can be helpful. I've seen a ton of beautiful math visualizations and I always enjoy them but on the whole, I think I've learned more from textbooks.
That said, the three blue one brown youtube course on Linear Algebra is truly amazing. I highly recommend it.
This is a little off topic but these comments make me wonder: why are lawyers so reluctant to give informal advice? I really appreciate the two thoughtful and informed comments at the top of this article here - why the disclaimers? I look for and give informal advice about all kinds of other topics that have the same levels of ambiguity and sometimes the same levels of importance as legal issues but it's always a real challenge to get a lawyer to weigh in informally on a legal issue.
If someone asks me my opinion about an engineering or management issue they're facing, I'll give it to them knowing that I don't know the complete set of facts and they should take my opinion with a grain of salt. I might be missing important context and I might just be wrong. If I ask someone else advice on any topic, I assume that there is an implicit disclaimer. Is there something fundamentally different about the law? Is it because lawyers are in the business of giving advice?
Anyway, appreciate you guys weighing in and hope you do it more frequently :).
Marc - I've had so much fun with Raspberry Pis! Really appreciate the work you do. Have you considered setting up a kickstarter-style pre-order system? I think a lot of people would sign up and it might give you guys more predictability.
I've used both and both are great overall. Even if you are using TensorFlow, I would recommend AWS right now for someone just starting out because the documentation is more currently more thorough, although that will probably change. The CloudML service looks really cool (and I think it's really what everyone will ultimately use), but I hit enough problems/bugs getting my model trained and running that I plan to wait for it to come out of beta before trying again.
feelix - what kind of models do you typically run? I've spent a fair amount of time getting Neural Nets to run on Raspberry Pis and other platforms. In my experience it's possible to do inference with most models but often it's intolerably slow. For example the stock inception model that comes as a demo in the tensorflow code base takes about 10 seconds per image to do inference on my Pi 3. What domains are you typically working in? Do you have some tricks to make things run faster?
I'm not sure that the AI significantly under-performed humans. It looks to me like the labels that the accuracy number came from were labeled by watching video of the bats. The 61% understanding number was off of 7 possible topics and averaged over each topic so it's definitely better than guessing. I suspect there's a fair amount of ambiguity and mislabeling in these "topics" from humans trying to interpret bat motivations so 100% accuracy probably isn't really feasible.
I have no idea what the state of the art is in bat understanding but the results seems really impressive to me - maybe I'm easily impressed? :)
It's pretty surprising to me that an M1 performs anywhere near a V100 on model training and I guess the most striking thing is the energy efficiency of the M1.