reader5000

This viewpoint is effectively British and Canadian law.

Well certainly, but interpretability is obsolete.

Yeah of course. I can't explain to you why NNs outperform bayesian approaches, probably just NNs are capturing the correct type of prior for vision/language tasks. And yeah bayesian models are more interpretable but when you have millions of latent variables I'm not sure interpretability is a thing.

You're right, but a) you have comp efficiency issues with MCMC, and b) just empirically MCMC models don't work as well as gradient descent + NN for many tasks.

Thats true. There is evidence internet addiction shows up on brain scans though.

> But internet addiction is invisible; it masks itself as the person being just slower than their peers, or lazier.

Socially isolated, failure to obtain/hold jobs, failure to reach traditional milestones (marriage, own house, children, etc.).

I thought I was weird in noticing "walking=less internet". Another thing is meditation, but it's no silver bullet. I was supposed to not be on the internet right now.

Try to derive e.g. a face detector from bayes theorem. You immediately arrive at computationally intractable sums/integrals. Yet, we have super-human image classifiers. Therefore, bayes theorem is obsolete. Sure, you can try to retrofit bayes theorem on top of a neural net, but who cares?

>I think it's more like not hiring a big name coding competition winner because they never bothered to learn how to use version control, or any coding best practice, or any language other than C.

Depends on what you're hiring for, but I'll take "competition winner with no version control" over "average programmer with expert VC capabilities".

>Bayes rule isn't some kind of deep magic

Yes, it's largely conceptually obsolete.

The people jamming out weekly SOTA machine learning models on arxiv aren't sitting around meditating on conditional probabilities. They're making little tweaks to giant models that are basically impossible for a human to comprehend.

"Data science" is a field with so much conceptual churn and fads that interviewing for it is a completely ridiculous notion.

This is like not hiring a [big name coding competition] winner because he didn't know radix sort.

Technically anybody who uses Excel is a "data scientist". Just got to get the right buzzwords.

Recurrent models do not simply map from one vector space to another and could very much be interpreted as reasoning about their environment. Of course they are significantly more difficult to train and backprop through time seems a bit of a hack.

Effectively it looks like the guy was looking for a suicide-buddy. Initially she told him to get help, but he kept pestering her about wanting to commit suicide. So she helped him through it in a non-material way. That's one interpretation, and there's nothing criminal about that. Just a poor decision from the judge here.

Reminiscent of a case from years ago where roommates secretly filmed a roommate having a gay encounter, and the roommate killed himself on finding out. I don't get how a person committing suicide in a non-intoxicated state is anybody's fault but their own.

I guess you could argue if you find a drunk person and tell them to do something reckless, but none of these cases seem to fit that.

There isnt really any math to deep learning other than the concept of a derivative which is taught in high school calculus. The reason deep learning papers seem mathy is people take network architectures and various elementary operations on them and try to express them symbolically in latex using summations and indexing-hell. For example the easy concept of "updating all the neurons in one layer based on the neurons in the previous layer and connecting weights" is expressed as matrix-vector multiplication for not really any apparent reason other than it is technically correct and makes for slicker notation, and I guess makes it easier to use APIs that compute gradients for you. Deep learning however is broadly an experimental science, which in many ways is the opposite of math as traditionally envisioned, in which great insights follow deductively from prior great insights. If you ask a basic question like "why should use 4 layers instead of 3?" there is no answer other than "4 works better". Similarly with gradient descent versus random search in weight space. There are many problem domains where random search is as good as any known hill-climbing heuristic search (like gradient descent). Why is GD so effective when learning image classifiers expressed as stacked weight sums? Who knows.

Would meeting strangers recommended to you by a deep learning bot be enriching??

I believe it is the case that deep NNs at the highest/deepest levels have interpretable feature-specific cells as well. Not sure what they think the difference is.

Never heard of this book, thanks for sharing.

Isn't rule 30 also Turing complete? Therefore GoL and rule 30 can both emulate the other, modulo possible exponential blowups in time/space requirements.

Cruel and unusual punishment. The judges kept alluding to the volume of the sales; thats just the nature of the technology, not the nature of the crime. There was also language in the original sentencing that he was "privileged" (i.e., white) and therefore deserved an absolutely absurd sentence. You can rape and murder somebody and spend less time in jail.

But a lot of the reason it's useful/interesting that say sorting is n log n is because we have computers to actually implement sorting. I'm not sure the problem was even posed before computers.

If you think spending power is the only relevant form of power in a marriage you are incorrect.

Yes because the entire value a husband receives from his wife is her housekeeping contributions.

In the sjw-religion, why is "homemaker" considered inferior to "computer programmer"? One of the oldest and most important human occupations versus hunched over at a desk slaving for a salary until being outsourced to a bot in 5 years? I've never understood the default sjw/"feminism" assumptions that anything feminine is "bad".

Human intelligence is rapidly becoming obsolete.

Are you referencing https://arxiv.org/abs/1412.0233?

That's not correct though, theres plenty of reasons why on an intuitive level neural networks should not work. Also, genetic algorithms are intuitively appealing but do not work well.

The article meant elegant in the sense of "nn training provably converges in subexponential time and doesnt overfit the data". E.g., like the theory for convex SVMs.

Mathematically, nn training is an unproven algorithm, it just works empirically surprisingly well. It's analogous to the simplex method which worked well for years without theoretical justification.

There is no understanding of why NN training converges in reasonable time and does not over/underfit. (Unlike for say convex SVMs)

Great. Explain why training converges in subexponential time despite the nonconvex cost function and does not overfit the data.