388 karma · joined April 13, 2016
Sometimes I can't help but wonder how the world where there is no need to spend endless billions on "cybersecurity", "infosec" would look like. Perhaps these billions would be used to create more value for the people. I find it insane that so much money and manpower is spent on scrambling the data to "secure" it from vandal-ish script kiddies (sometimes hired by governments), there is definitely something unhealthy about it.
But look at what Deepmind does: it takes these ideas (and also ideas from systems neuroscience), implements them as differentiable modules and trains them on data in end-to-end fashion. This works really well.
Learning is very important, much more important than architecture. If you have a model that can learn you can add more structure later - again this is what modern deep learning is all about.
>Turing-completeness is quite broad and nonspecific, like I said.
It is, but feedforward models (and almost every Bayesian/statistical model) don't possess it even in theory, while RNNs do.
>Doing "some computation" is an obvious statement that doesn't add any information.
Let me be more specific: currently researchers think that later stages of CNNs do something that is more interpretable as computation than as mere pattern matching. Our world doesn't require 50-level hierarchy, but resnets with 50+ layers do good, looks like because they learn some non-trivial computation.
>the jury is still out on whether any of those RNN approaches will be the needed breakthrough.
Sure, we'll see. Maybe there won't be need in any breakthrough, just incremental improvement of models. And even current models when scaled up to next-gen hardware (see nervana) can surprise us again with their performance.
Even with language there is a whole spectrum of language skill. Some animals like parrots, crows, great apes, and then people can learn language at various levels.
Some deep learning models can already learn basic language skills too. The question is, how far can these techniques go. Maybe, pretty far.
Yes they are: https://indico.io/blog/sequence-modeling-neural-networks-par... https://github.com/harvardnlp/seq2seq-attn
It is funny how every AI post on HN turns into a speculative discussion forum full of words "I think", "likely", "I suspect", "My guess" etc, when all the research is available for free and everyone is free to download and read it to get a real understanding of what's going on in the field.
>what I see is the fact that networks have huge capacity to fit to data and are deep (rely on a hierarchy of features).
Actually recurrent neural networks like LSTM are turing-complete, i.e. for every halting algorithm it is trivial to implement an RNN that computes it. It is non-trivial to learn these parameters from algorithm IO data, but for many tasks it is possible too.
>I suspect that most of what is being learned (and solved) is perception of useful features from the raw game images.
It is not this simple, deep enough convnets can represent computations, the consensus is that middle and upper layers of convnets represent some useful computation steps. Also note that human brain can only do so much computation steps to answer questions when in dialogue, due to time and speed limits.
>My guess is we may need a fundamental breakthrough in a newfangled hierarchical learning system that is better suited for language to “solve” NLP.
This is being worked on, see the first link for Memory Networks and Stack RNNs, DeQue RNNs, Tree RNNs. Deep learning is a very generic term, there are dozens of various feedforward and recurrent architectures that are fully differentiable. The full potential of such models has not been nearly reached yet and maybe language understanding will be solved in the coming years (again, the first link shows that it is in process of being solved).
>NNs that manipulate existing code? e.g. parse it, find design pattern, optimize
Also: you probably shouldn't need special architectures for such problems, these can be solved by seq2seq, NTM, RL-NTM and such, given a good dataset. Such tasks are still beyond current state of art though, it seems.
Every ML problem has a lot of algorithms that can be applied to it to solve it. Previously unsolvable problem (e.g. learning playing atari game from pixels) can become solvable if a better ML algorithm is devised, more compute power is applied, or the problem is simplified (maybe with data augmentation or feature engineering, or constraining the domain).
I think the problem you are stating - finding the best configuration of website to extract more ad revenue - can be simplified until it is solvable by currently available algorithms and hardware. Big companies are already doing similar things - Google uses reinforcement learning to recommend new items on its services, Facebook is known to use machine learning to somehow make more profit from users.
Given expert knowledge of the problem and good algorithms, it can be solved in some form, that's my point here.
Do not underestimate what good ML model can learn given enough data. And in case of online shopping there usually is more than enough data, and big companies like Amazon have been using ML (context bandits) for quite some time.
>To accomplish this goal, we leverage a dataset from a pedagogical programming online judge (OJ) system,2 intended for the undergraduate course, Introduction to Computing. The OJ system comprises different programming problems. Students submit their source code to a specific problem, and the OJ system judges its validity automatically (via running).
And trained seq2seq model on it on character level and it almost works, barring a few typos. IMHO this is an underappreciated breakthrough. With more data and a better model generating competition-grade programs seems possible.
In fact there is already a lot of solid work just on this subject:
* Learning algorithms from examples http://arxiv.org/abs/1511.07275 https://arxiv.org/abs/1410.5401
* Generating source code from natural language description http://arxiv.org/abs/1510.07211
* And, the most closest work to what author probably wants, a way to write a program in forth while leaving some functions as neural blackboxes to be learned from examples: http://arxiv.org/abs/1605.06640
* Also there is a whole research program by nothing less than Facebook AI Research that explicitly aims at creating a conversational AI agent that is able to translate user's natural language orders into programs (asking to the user additional questions if necessary): http://arxiv.org/abs/1511.08130 (there is also a summary here http://colinraffel.com/wiki/a_roadmap_towards_machine_intell... )
And deepmind is also working on conversational agents: https://youtu.be/vQXAsdMa_8A?t=1265
Given current success of such models, automating simple programming tasks maybe not as much research as engineering and scaling up problem.
There is a lot of exciting machine learning research out there nowadays. Almost all of this research is available for free from papers posted on arxiv. It is a really good idea to read more about state of the art before coming with new ideas.
Modern models learn to achieve goals in increasingly sophisticated 3d environments and even learn execute commands in some very limited form of natural language. You could say that the question of achieving more humanlike performance may be "just" a question of engineering and scale.
>This paper envisions an end-to-end program generation scenario using recurrent neural networks (RNNs): Users can express their intention in natural language; an RNN then automatically generates corresponding code in a characterby-by-character fashion. We demonstrate its feasibility through a case study and empirical analysis. To fully make such technique useful in practice, we also point out several cross-disciplinary challenges, including modeling user intention, providing datasets, improving model architectures, etc. Although much long-term research shall be addressed in this new field, we believe end-to-end program generation would become a reality in future decades, and we are looking forward to its practice.
Also you don't need the management cores to be fast, it suffices to have lots of GHz+ cores controlling ultra-wide SIMD (as Intel's Xeon Phi has shown).
Arguably, one could get another order of magnitude speedup by implementing XNOR-net (http://arxiv.org/abs/1603.05279) hardware accelerator, but it hasn't been done yet.
Also there is a free book that contains descriptions of various common and exotic compression formats http://www.mattmahoney.net/dc/dce.html
Overall I think we are yet to see the full potential of deep learning unleashed on data compression. For example the neural network in cmix compressor is quite primitive compared to modern architectures. Someone will certainly find a way to do better than that!
Also see the graphs for atari environments: many games are played by rl agents "at human-level or above".
This has led Solomonoff to investigation of a question of universal sequence prediction. A couple of years later Solomonoff wrote a paper about such a system for prediction that used algorithmic probability (he is cited later as the original developer of algorithmic information theory which was later independently discovered by Kolmogorov who later acknowledged that the Solomonoff was the first). This method, Solomonoff's induction, is proven to be the most optimal (though incomputable) machine learning method possible.
He has never abandoned this project and for the rest of his life he focused on making more sophisticated system designs that are computable while still being proven to be optimal.
His latest system is called "Alpha", and it is designed as a machine for solving a sequence of function inversion and time limited optimization problems (a majority of science/engineering problems can be formulated this way) in a way that exploits experience gathered while solving these problems. This system, again, is proven to be optimal in a certain sense. He also tried to implement this system with various practical optimizations, but it didn't converge fast enough on his training sequences and on the hardware of that time.
Still, with modern hardware it is a possibility that it could work. And the whole design is described in the papers, so people can (and actually do, though privately and perhaps without much success) implement this system.
Here are the relevant papers: http://world.std.com/~rjs/publications/IncLrn89.pdf http://world.std.com/~rjs/nips02.pdf
Also it should be possible to train the NN model to explain its decision, even in plain language.
I think the "right to explanation" law is pretty sensible, people should have a right to question the algorithmic judgment that impacts their life. It will be somewhat inconvenient for big companies applying ML to make these decisions, but it won't ruin their business.
How to make an easily approachable tutorial that combines game-like world and real learning AI (for example Q-learning or maybe evolutionary policy search) remains and open question. Maybe someone will build it.
2) Hype about "superintelligent AI coming soon" is pretty much unfounded, current deep learning approaches have a lot of limits, hardware progress stalled as well (moore's law is officially dead). AI technology is like a fragile flower on a life support, it is so weak you can easily kill it with regulation and FUD. Instead of questioning if we suddenly achieve "real AI" one could wonder if the tech will ever reach any level of moderate general AI, considering the death of moore's law.
3) Nobody will listen to scaremongering, people will just continue doing their thing anyway.