recentpopularlog in

nhaliday : acmtariat   247

« earlier  
A Recipe for Training Neural Networks
acmtariat  org:bleg  nibble  machine-learning  deep-learning  howto  tutorial  guide  nitty-gritty  gotchas  init  list  checklists  expert-experience  abstraction  composition-decomposition  gradient-descent  data-science  error  debugging  benchmarks  programming  engineering  best-practices  dataviz  checking  plots  generalization  regularization  unsupervised  optimization  ensembles  random  methodology  multi  twitter  social  discussion  techtariat  links  org:med  pdf  visualization  python  recommendations  advice  devtools 
april 2019 by nhaliday
Surveil things, not people – The sideways view
Technology may reach a point where free use of one person’s share of humanity’s resources is enough to easily destroy the world. I think society needs to make significant changes to cope with that scenario.

Mass surveillance is a natural response, and sometimes people think of it as the only response. I find mass surveillance pretty unappealing, but I think we can capture almost all of the value by surveilling things rather than surveilling people. This approach avoids some of the worst problems of mass surveillance; while it still has unattractive features it’s my favorite option so far.

...

The idea
We’ll choose a set of artifacts to surveil and restrict. I’ll call these heavy technology and everything else light technology. Our goal is to restrict as few things as possible, but we want to make sure that someone can’t cause unacceptable destruction with only light technology. By default something is light technology if it can be easily acquired by an individual or small group in 2017, and heavy technology otherwise (though we may need to make some exceptions, e.g. certain biological materials or equipment).

Heavy technology is subject to two rules:

1. You can’t use heavy technology in a way that is unacceptably destructive.
2. You can’t use heavy technology to undermine the machinery that enforces these two rules.

To enforce these rules, all heavy technology is under surveillance, and is situated such that it cannot be unilaterally used by any individual or small group. That is, individuals can own heavy technology, but they cannot have unmonitored physical access to that technology.

...

This proposal does give states a de facto monopoly on heavy technology, and would eventually make armed resistance totally impossible. But it’s already the case that states have a massive advantage in armed conflict, and it seems almost inevitable that progress in AI will make this advantage larger (and enable states to do much more with it). Realistically I’m not convinced this proposal makes things much worse than the default.

This proposal definitely expands regulators’ nominal authority and seems prone to abuses. But amongst candidates for handling a future with cheap and destructive dual-use technology, I feel this is the best of many bad options with respect to the potential for abuse.
ratty  acmtariat  clever-rats  risk  existence  futurism  technology  policy  alt-inst  proposal  government  intel  authoritarianism  orwellian  tricks  leviathan  security  civilization  ai  ai-control  arms  defense  cybernetics  institutions  law  unintended-consequences  civil-liberty  volo-avolo  power  constraint-satisfaction  alignment 
april 2018 by nhaliday
The Hanson-Yudkowsky AI-Foom Debate - Machine Intelligence Research Institute
How Deviant Recent AI Progress Lumpiness?: http://www.overcomingbias.com/2018/03/how-deviant-recent-ai-progress-lumpiness.html
I seem to disagree with most people working on artificial intelligence (AI) risk. While with them I expect rapid change once AI is powerful enough to replace most all human workers, I expect this change to be spread across the world, not concentrated in one main localized AI system. The efforts of AI risk folks to design AI systems whose values won’t drift might stop global AI value drift if there is just one main AI system. But doing so in a world of many AI systems at similar abilities levels requires strong global governance of AI systems, which is a tall order anytime soon. Their continued focus on preventing single system drift suggests that they expect a single main AI system.

The main reason that I understand to expect relatively local AI progress is if AI progress is unusually lumpy, i.e., arriving in unusually fewer larger packages rather than in the usual many smaller packages. If one AI team finds a big lump, it might jump way ahead of the other teams.

However, we have a vast literature on the lumpiness of research and innovation more generally, which clearly says that usually most of the value in innovation is found in many small innovations. We have also so far seen this in computer science (CS) and AI. Even if there have been historical examples where much value was found in particular big innovations, such as nuclear weapons or the origin of humans.

Apparently many people associated with AI risk, including the star machine learning (ML) researchers that they often idolize, find it intuitively plausible that AI and ML progress is exceptionally lumpy. Such researchers often say, “My project is ‘huge’, and will soon do it all!” A decade ago my ex-co-blogger Eliezer Yudkowsky and I argued here on this blog about our differing estimates of AI progress lumpiness. He recently offered Alpha Go Zero as evidence of AI lumpiness:

...

In this post, let me give another example (beyond two big lumps in a row) of what could change my mind. I offer a clear observable indicator, for which data should have available now: deviant citation lumpiness in recent ML research. One standard measure of research impact is citations; bigger lumpier developments gain more citations that smaller ones. And it turns out that the lumpiness of citations is remarkably constant across research fields! See this March 3 paper in Science:

I Still Don’t Get Foom: http://www.overcomingbias.com/2014/07/30855.html
All of which makes it look like I’m the one with the problem; everyone else gets it. Even so, I’m gonna try to explain my problem again, in the hope that someone can explain where I’m going wrong. Here goes.

“Intelligence” just means an ability to do mental/calculation tasks, averaged over many tasks. I’ve always found it plausible that machines will continue to do more kinds of mental tasks better, and eventually be better at pretty much all of them. But what I’ve found it hard to accept is a “local explosion.” This is where a single machine, built by a single project using only a tiny fraction of world resources, goes in a short time (e.g., weeks) from being so weak that it is usually beat by a single human with the usual tools, to so powerful that it easily takes over the entire world. Yes, smarter machines may greatly increase overall economic growth rates, and yes such growth may be uneven. But this degree of unevenness seems implausibly extreme. Let me explain.

If we count by economic value, humans now do most of the mental tasks worth doing. Evolution has given us a brain chock-full of useful well-honed modules. And the fact that most mental tasks require the use of many modules is enough to explain why some of us are smarter than others. (There’d be a common “g” factor in task performance even with independent module variation.) Our modules aren’t that different from those of other primates, but because ours are different enough to allow lots of cultural transmission of innovation, we’ve out-competed other primates handily.

We’ve had computers for over seventy years, and have slowly build up libraries of software modules for them. Like brains, computers do mental tasks by combining modules. An important mental task is software innovation: improving these modules, adding new ones, and finding new ways to combine them. Ideas for new modules are sometimes inspired by the modules we see in our brains. When an innovation team finds an improvement, they usually sell access to it, which gives them resources for new projects, and lets others take advantage of their innovation.

...

In Bostrom’s graph above the line for an initially small project and system has a much higher slope, which means that it becomes in a short time vastly better at software innovation. Better than the entire rest of the world put together. And my key question is: how could it plausibly do that? Since the rest of the world is already trying the best it can to usefully innovate, and to abstract to promote such innovation, what exactly gives one small project such a huge advantage to let it innovate so much faster?

...

In fact, most software innovation seems to be driven by hardware advances, instead of innovator creativity. Apparently, good ideas are available but must usually wait until hardware is cheap enough to support them.

Yes, sometimes architectural choices have wider impacts. But I was an artificial intelligence researcher for nine years, ending twenty years ago, and I never saw an architecture choice make a huge difference, relative to other reasonable architecture choices. For most big systems, overall architecture matters a lot less than getting lots of detail right. Researchers have long wandered the space of architectures, mostly rediscovering variations on what others found before.

Some hope that a small project could be much better at innovation because it specializes in that topic, and much better understands new theoretical insights into the basic nature of innovation or intelligence. But I don’t think those are actually topics where one can usefully specialize much, or where we’ll find much useful new theory. To be much better at learning, the project would instead have to be much better at hundreds of specific kinds of learning. Which is very hard to do in a small project.

What does Bostrom say? Alas, not much. He distinguishes several advantages of digital over human minds, but all software shares those advantages. Bostrom also distinguishes five paths: better software, brain emulation (i.e., ems), biological enhancement of humans, brain-computer interfaces, and better human organizations. He doesn’t think interfaces would work, and sees organizations and better biology as only playing supporting roles.

...

Similarly, while you might imagine someday standing in awe in front of a super intelligence that embodies all the power of a new age, superintelligence just isn’t the sort of thing that one project could invent. As “intelligence” is just the name we give to being better at many mental tasks by using many good mental modules, there’s no one place to improve it. So I can’t see a plausible way one project could increase its intelligence vastly faster than could the rest of the world.

Takeoff speeds: https://sideways-view.com/2018/02/24/takeoff-speeds/
Futurists have argued for years about whether the development of AGI will look more like a breakthrough within a small group (“fast takeoff”), or a continuous acceleration distributed across the broader economy or a large firm (“slow takeoff”).

I currently think a slow takeoff is significantly more likely. This post explains some of my reasoning and why I think it matters. Mostly the post lists arguments I often hear for a fast takeoff and explains why I don’t find them compelling.

(Note: this is not a post about whether an intelligence explosion will occur. That seems very likely to me. Quantitatively I expect it to go along these lines. So e.g. while I disagree with many of the claims and assumptions in Intelligence Explosion Microeconomics, I don’t disagree with the central thesis or with most of the arguments.)
ratty  lesswrong  subculture  miri-cfar  ai  risk  ai-control  futurism  books  debate  hanson  big-yud  prediction  contrarianism  singularity  local-global  speed  speedometer  time  frontier  distribution  smoothness  shift  pdf  economics  track-record  abstraction  analogy  links  wiki  list  evolution  mutation  selection  optimization  search  iteration-recursion  intelligence  metameta  chart  analysis  number  ems  coordination  cooperate-defect  death  values  formal-values  flux-stasis  philosophy  farmers-and-foragers  malthus  scale  studying  innovation  insight  conceptual-vocab  growth-econ  egalitarianism-hierarchy  inequality  authoritarianism  wealth  near-far  rationality  epistemic  biases  cycles  competition  arms  zero-positive-sum  deterrence  war  peace-violence  winner-take-all  technology  moloch  multi  plots  research  science  publishing  humanity  labor  marginal  urban-rural  structure  composition-decomposition  complex-systems  gregory-clark  decentralized  heavy-industry  magnitude  multiplicative  endogenous-exogenous  models  uncertainty  decision-theory  time-prefer 
april 2018 by nhaliday
Sequence Modeling with CTC
A visual guide to Connectionist Temporal Classification, an algorithm used to train deep neural networks in speech recognition, handwriting recognition and other sequence problems.
acmtariat  techtariat  org:bleg  nibble  better-explained  machine-learning  deep-learning  visual-understanding  visualization  analysis  let-me-see  research  sequential  audio  classification  model-class  exposition  language  acm  approximation  comparison  markov  iteration-recursion  concept  atoms  distribution  orders  DP  heuristic  optimization  trees  greedy  matching  gradient-descent  org:popup 
december 2017 by nhaliday
[1709.06560] Deep Reinforcement Learning that Matters
https://twitter.com/WAWilsonIV/status/912505885565452288
I’ve been experimenting w/ various kinds of value function approaches to RL lately, and its striking how primitive and bad things seem to be
At first I thought it was just that my code sucks, but then I played with the OpenAI baselines and nope, it’s the children that are wrong.
And now, what comes across my desk but this fantastic paper: (link: https://arxiv.org/abs/1709.06560) arxiv.org/abs/1709.06560 How long until the replication crisis hits AI?

https://twitter.com/WAWilsonIV/status/911318326504153088
Seriously I’m not blown away by the PhDs’ records over the last 30 years. I bet you’d get better payoff funding eccentrics and amateurs.
There are essentially zero fundamentally new ideas in AI, the papers are all grotesquely hyperparameter tuned, nobody knows why it works.

Deep Reinforcement Learning Doesn't Work Yet: https://www.alexirpan.com/2018/02/14/rl-hard.html
Once, on Facebook, I made the following claim.

Whenever someone asks me if reinforcement learning can solve their problem, I tell them it can’t. I think this is right at least 70% of the time.
papers  preprint  machine-learning  acm  frontier  speedometer  deep-learning  realness  replication  state-of-art  survey  reinforcement  multi  twitter  social  discussion  techtariat  ai  nibble  org:mat  unaffiliated  ratty  acmtariat  liner-notes  critique  sample-complexity  cost-benefit  todo 
september 2017 by nhaliday
Superintelligence Risk Project Update II
https://www.jefftk.com/p/superintelligence-risk-project-update

https://www.jefftk.com/p/conversation-with-michael-littman
For example, I asked him what he thought of the idea that to we could get AGI with current techniques, primarily deep neural nets and reinforcement learning, without learning anything new about how intelligence works or how to implement it ("Prosaic AGI" [1]). He didn't think this was possible, and believes there are deep conceptual issues we still need to get a handle on. He's also less impressed with deep learning than he was before he started working in it: in his experience it's a much more brittle technology than he had been expecting. Specifically, when trying to replicate results, he's often found that they depend on a bunch of parameters being in just the right range, and without that the systems don't perform nearly as well.

The bottom line, to him, was that since we are still many breakthroughs away from getting to AGI, we can't productively work on reducing superintelligence risk now.

He told me that he worries that the AI risk community is not solving real problems: they're making deductions and inferences that are self-consistent but not being tested or verified in the world. Since we can't tell if that's progress, it probably isn't. I asked if he was referring to MIRI's work here, and he said their work was an example of the kind of approach he's skeptical about, though he wasn't trying to single them out. [2]

https://www.jefftk.com/p/conversation-with-an-ai-researcher
Earlier this week I had a conversation with an AI researcher [1] at one of the main industry labs as part of my project of assessing superintelligence risk. Here's what I got from them:

They see progress in ML as almost entirely constrained by hardware and data, to the point that if today's hardware and data had existed in the mid 1950s researchers would have gotten to approximately our current state within ten to twenty years. They gave the example of backprop: we saw how to train multi-layer neural nets decades before we had the computing power to actually train these nets to do useful things.

Similarly, people talk about AlphaGo as a big jump, where Go went from being "ten years away" to "done" within a couple years, but they said it wasn't like that. If Go work had stayed in academia, with academia-level budgets and resources, it probably would have taken nearly that long. What changed was a company seeing promising results, realizing what could be done, and putting way more engineers and hardware on the project than anyone had previously done. AlphaGo couldn't have happened earlier because the hardware wasn't there yet, and was only able to be brought forward by massive application of resources.

https://www.jefftk.com/p/superintelligence-risk-project-conclusion
Summary: I'm not convinced that AI risk should be highly prioritized, but I'm also not convinced that it shouldn't. Highly qualified researchers in a position to have a good sense the field have massively different views on core questions like how capable ML systems are now, how capable they will be soon, and how we can influence their development. I do think these questions are possible to get a better handle on, but I think this would require much deeper ML knowledge than I have.
ratty  core-rats  ai  risk  ai-control  prediction  expert  machine-learning  deep-learning  speedometer  links  research  research-program  frontier  multi  interview  deepgoog  games  hardware  performance  roots  impetus  chart  big-picture  state-of-art  reinforcement  futurism  🤖  🖥  expert-experience  singularity  miri-cfar  empirical  evidence-based  speculation  volo-avolo  clever-rats  acmtariat  robust  ideas  crux  atoms  detail-architecture  software  gradient-descent 
july 2017 by nhaliday
How to Escape Saddle Points Efficiently – Off the convex path
A core, emerging problem in nonconvex optimization involves the escape of saddle points. While recent research has shown that gradient descent (GD) generically escapes saddle points asymptotically (see Rong Ge’s and Ben Recht’s blog posts), the critical open problem is one of efficiency — is GD able to move past saddle points quickly, or can it be slowed down significantly? How does the rate of escape scale with the ambient dimensionality? In this post, we describe our recent work with Rong Ge, Praneeth Netrapalli and Sham Kakade, that provides the first provable positive answer to the efficiency question, showing that, rather surprisingly, GD augmented with suitable perturbations escapes saddle points efficiently; indeed, in terms of rate and dimension dependence it is almost as if the saddle points aren’t there!
acmtariat  org:bleg  nibble  liner-notes  machine-learning  acm  optimization  gradient-descent  local-global  off-convex  time-complexity  random  perturbation  michael-jordan  iterative-methods  research  learning-theory  math.DS  iteration-recursion  org:popup 
july 2017 by nhaliday
Unsupervised learning, one notion or many? – Off the convex path
(Task A) Learning a distribution from samples. (Examples: gaussian mixtures, topic models, variational autoencoders,..)

(Task B) Understanding latent structure in the data. This is not the same as (a); for example principal component analysis, clustering, manifold learning etc. identify latent structure but don’t learn a distribution per se.

(Task C) Feature Learning. Learn a mapping from datapoint → feature vector such that classification tasks are easier to carry out on feature vectors rather than datapoints. For example, unsupervised feature learning could help lower the amount of labeled samples needed for learning a classifier, or be useful for domain adaptation.

Task B is often a subcase of Task C, as the intended user of “structure found in data” are humans (scientists) who pour over the representation of data to gain some intuition about its properties, and these “properties” can be often phrased as a classification task.

This post explains the relationship between Tasks A and C, and why they get mixed up in students’ mind. We hope there is also some food for thought here for experts, namely, our discussion about the fragility of the usual “perplexity” definition of unsupervised learning. It explains why Task A doesn’t in practice lead to good enough solution for Task C. For example, it has been believed for many years that for deep learning, unsupervised pretraining should help supervised training, but this has been hard to show in practice.
acmtariat  org:bleg  nibble  machine-learning  acm  thinking  clarity  unsupervised  conceptual-vocab  concept  explanation  features  bayesian  off-convex  deep-learning  latent-variables  generative  intricacy  distribution  sampling  grokkability-clarity  org:popup 
june 2017 by nhaliday
Prékopa–Leindler inequality | Academically Interesting
Consider the following statements:
1. The shape with the largest volume enclosed by a given surface area is the n-dimensional sphere.
2. A marginal or sum of log-concave distributions is log-concave.
3. Any Lipschitz function of a standard n-dimensional Gaussian distribution concentrates around its mean.
What do these all have in common? Despite being fairly non-trivial and deep results, they all can be proved in less than half of a page using the Prékopa–Leindler inequality.

ie, Brunn-Minkowski
acmtariat  clever-rats  ratty  math  acm  geometry  measure  math.MG  estimate  distribution  concentration-of-measure  smoothness  regularity  org:bleg  nibble  brunn-minkowski  curvature  convexity-curvature 
february 2017 by nhaliday
Predicting with confidence: the best machine learning idea you never heard of | Locklin on science
The advantages of conformal prediction are many fold. These ideas assume very little about the thing you are trying to forecast, the tool you’re using to forecast or how the world works, and they still produce a pretty good confidence interval. Even if you’re an unrepentant Bayesian, using some of the machinery of conformal prediction, you can tell when things have gone wrong with your prior. The learners work online, and with some modifications and considerations, with batch learning. One of the nice things about calculating confidence intervals as a part of your learning process is they can actually lower error rates or use in semi-supervised learning as well. Honestly, I think this is the best bag of tricks since boosting; everyone should know about and use these ideas.

The essential idea is that a “conformity function” exists. Effectively you are constructing a sort of multivariate cumulative distribution function for your machine learning gizmo using the conformity function. Such CDFs exist for classical stuff like ARIMA and linear regression under the correct circumstances; CP brings the idea to machine learning in general, and to models like ARIMA when the standard parametric confidence intervals won’t work. Within the framework, the conformity function, whatever may be, when used correctly can be guaranteed to give confidence intervals to within a probabilistic tolerance. The original proofs and treatments of conformal prediction, defined for sequences, is extremely computationally inefficient. The conditions can be relaxed in many cases, and the conformity function is in principle arbitrary, though good ones will produce narrower confidence regions. Somewhat confusingly, these good conformity functions are referred to as “efficient” -though they may not be computationally efficient.
techtariat  acmtariat  acm  machine-learning  bayesian  stats  exposition  research  online-learning  probability  decision-theory  frontier  unsupervised  confidence 
february 2017 by nhaliday
Unlearning descriptive statistics | Hacker News
For readers who are OK with some math, I recommend John Myles White's eye-opening post about means, medians, and modes: http://www.johnmyleswhite.com/notebook/2013/03/22/modes-medians-and-means-an-unifying-perspective/. He describes these summary descriptive stats in terms of what penalty function they minimize: mean minimizes L2, median minimizes L1, mode minimizes L0.
hn  commentary  techtariat  acmtariat  data-science  explanation  multi  norms  org:bleg  nibble  scitariat  expectancy  traces 
february 2017 by nhaliday
« earlier      
per page:    204080120160

Copy this bookmark:





to read