thearn4

It's interesting how amenable image classification neural networks are to the "take working model, peel off last layer or two, retrain for a new application" approach. I've seen this suggested as working pretty well in a few instances.

I guess the interpretation is that the first few normalize->convolution->pool->dropout layers are basically achieving something broadly analogous to the initial feature extraction steps that used to be the mainstay in this area (PCA/ICA, HOG, SIFT/SURF, etc.), and are reasonably problem-independent.

My former advisor keeps his .edu webpage full of pre-print versions of every paper that he has published. I'm not sure if every journal is okay with this, but I'm also not positive that he would stop if prompted. He'd probably just target other journals.

Well, I still think Stallman is quite crazy still, but makes some excellent points too. He's definitely not stupid.

Shouldn't it be possible to backpropogate those categorical outputs all the way back to the inputs/features (NOT weights) after a forward pass, to localize the sensitivity of them with respect to the actual pixels for a prediction? I imagine that would have to give at least some insight.

Beyond that, the convolution/max pool repeated steps could be understood to be applying something akin to a multi-level wavelet decomposition, which is pretty well understood. It's how classical matched filtering, Haar cascading, and a wide variety of proceeding image classification methods operated at their first steps too.

CNNs/Deep learning really doesn't seem like a black box at all when examined in sequence. But to me at least, randomized ensemble methods (random forest, etc.) are actually a bit more mysterious to me in their performance out of the box, with little tuning.

CFD and other high-fidelity multiphysics simulations can be efficient when GPU-bound. It all comes down to being able to write a matrix and stream a bunch of vectors to multiply it by. When considering Krylov subspace methods, this extends to covering a large part of the problem space of linear algebra.

I think the Tesla cards in particular are double precision, which is not necessarily the best for deep learning applications.

For packaging pains, do you mean the fact that there is no single agreed on packaging/provisioning method (easy_install, pip, conda, apt, etc.), or that these solutions are terrible?

I do feel like there is a middle ground somewhere there between the "format all content to fit A4 paper" idiom of Word/TeX, and the more interactive markdown + simulation environment of Jupyter. The latter is really growing on me, but conferences really still want their paper articles and pdf/pptx slide decks.

seaborn is actually pretty nice, though not as diverse.

I use Python 3.6 in an Anaconda env

I think the biggest issue with matplotlib is it's straddling of the fence between MATLAB and pythonic syntax throughout the library. It makes it actually pretty hard to predict what the syntax/call signature of a particular method is if you don't use it fairly regularly.

> I remember when I was a kid we used to borrow each other's NES games all the time and never give them back.

I definitely remember it being a very collective ownership thing in my neighborhood when it came to NES games.

On a per-day basis, generally only based on my electronic calendars (personal=gcal, work=outlook).

On a week to week or month to month (and longer term) basis, I have rough notes sketched out in a plaintext file.

Scientific and engineering ethics kind of stand in the way of letting people throw away their lives just because it's cool.

That's a pretty neat concept.

An audiobook format would be interesting too. Might be tricky to do in the car though, ha.

My comment was less about FOSS actually. I think it has more to do with the branding of Arduino. When people think educational or approachable microcontrollers, they think Arduino. As soon as that gets muddied a bit, I think folks may see that there are much better alternatives, and will shift to that.

This looks like a ton of fun!

I think the Arduino civil war is inevitably going to lead to ESP8266/NodeMCU boards taking over this market.

> conda install matplotlib

Anaconda Python really should be more popular on OS X and Linux than it is. Homebrew is okay, but a little bit wonky for Python packages.

I did for awhile, until making the resolution this year to use 3.6+ whenever and wherever possible. So far it turns out, I had very few incompatibilities.

Yeah ~3 FTE sounds about right. But if offered as grants for students to work on over internships or something, it would go a bit further. But I don't know enough about the GIMP codebase to know how feasible that would really be.

Want to? Near zero. Have to? Maybe a bit larger, hard to say.

As much as I hate having to work in Word for small things that could so easily be a plain-text format, the versioning system within Word isn't exactly the worst thing ever.

I have had the same thoughts too. What also had me a bit puzzled in the ML community is the lack of discussion on local optimality, and constraint/Lagrangian loss functions. But there is probably a subtlety that I'm missing.

The engineers (aerospace, etc) also prefer the term "adjoint mode"

Thanks for the clarification, Travis, that's pretty exciting to hear.

I'm building course notes right now for a college class using markdown & MathML/LaTeX, that i'll be hosting on Github pages for the students to browse. It's by far the best balance that I've come across yet.

I could have sworn that Continuum had gov't funding for numpy development, but maybe that was just for Blaze?

The output looks like transcripts of some of the poorer TEDx talks that I've watched. Excellent job.

The same thing led me to stop listening to Radiolab too.

You and me both, but there was a time when I would have been perfectly fine with this, because I lived it.

Well, it was the military, and you'd have to replace "hack away" with "non-stop exercise and dumb chores", but the analogy still fits I think.

It rings true to me. I have a math PhD, and taught at various levels. Something I can say is that while I'm not sure that everyone is capable of groking advanced math, a person who can do that could actually come from anywhere, at least. And our system isn't really all that great at seeking that capability out. Unless you are in a fairly privileged group.