Hunch.net is John Langford's machine learning blog, covering research practice, conferences and evaluation methods.
Commenters recommend it as a pure machine learning blog and cite it on the Netflix Prize scoring and peer review debates.
A few others:
Machine Learning (Theory), http://hunch.net/, is a good place to learn about new machine-learning papers. It doesn't really do research exposition on the blog, but it posts about recent ML conferences, highlighting some of the papers the author finds interesting.
Embedded in Academia, http://blog.regehr.org/, is about 50% personal stuff, but 50% posts on John Regehr's work on C-compiler fuzzing, with some interesting examples if you're into compilers or the finer points of C semantics.
Proper Fixation, http://www.yosefk.com/blog/, is by an embedded developer (not academic), and not always about research, but it has some good researchy and expository posts. For example, it has the best concise overview I've found of how SIMT/SIMD/SMT relate (http://www.yosefk.com/blog/simd-simt-smt-parallelism-in-nvidia-gpus.html).
While it's a mathematics blog, Terence Tao's blog, http://terrytao.wordpress.com/, has a lot of content likely of interest to computer scientists as well. In particular, his blog-exposition versions of papers are often a better introduction to recent research for nonspecialists than anything in the official published literature is.
Tomasz Malisiewicz's computer-vision blog, http://quantombone.blogspot.com/, has intermittent but often quite good posts on object recognition and similar topics.
Of course I can't refrain from mentioning my own quasi-blog, http://www.kmjn.org/notes/, though only about 1/4 of it is on computer science (about 4/5 of my…
Preface: AlphaGo is an amazing achievement and does show an interesting advancement in the field.
Yet ... it really doesn't mean almost anything that people are predicting it to mean. Slashdot went so far as to say that "We know now that we don't need any big new breakthroughs to get to true AI".
The field of ML/AI is in a fight where people want more science fiction than scientific reality. Science fiction is sexy, sells well, and doesn't require the specifics.
Some of the limitations preventing AlphaGo from being general:
+ Monte Carlo tree search (MCTS) is really effective at Go but not applicable to many other domains we care about. If your problem is in terms of {state, action} pairs and you're able to run simulations to predict outcomes, great, but otherwise, not so much. Go also has the advantage of perfect information (you know the full state of the board) and deterministic simulation (you know with certainty what the state is after action A).
+ The neural networks (NN) were bootstrapped by predicting the next moves in more matches than any individual human has ever seen, let alone played. It then played more against itself (cool!) to improve - but it didn't learn that from scratch. They're aiming to learn this step without the human database but it'll still be very different (read: inefficient) compared to the type of learning a human does.
+ The hardware requirements were stunning (280 GPUs and 1920 CPUs for the largest variant) and were an integral part to…
To use the test set explicitly for evaluation is a deadly sin. When found out, you'd face serious damage to your reputation (Like Baidu did a few years ago). [1] What the decreasing performance results on the remade CIFAR-10 test set shows, is probably more akin to a subtle form of overfitting (Due to these datasets being around for a long time, the good results get published, and the bad results discarded, leading to a feedback loop). [2] It is also possible the original test set was closer in distribution to the train set, than the remade one. The ranks stay too consistent for test set evaluation cheating.
I also think the "do not trust saliency maps" is too strongly worded. The authors of that paper used adversarial techniques to change the saliency maps. Not just random noise or slight variation, but carefully crafted noise to attack saliency feature importance maps.
> For example, while it would be nice to have a CNN identify a spot on an MRI image as a malignant cancer-causing tumor, these results should not be trusted if they are based on fragile interpretation methods.
Interpretation methods are as fragile as the deep learning model itself, which is susceptible to adversarial images too. If you allow for scenario's with adversarial images, not only should you not trust the interpretation methods, but also the predictions themselves, destroying any pragmatic value left. It is hard to imagine a realistic threat scenario where MRI's are altered by an adversary,…
> AIXI is a pretty deeply flawed theory of "universal intelligence"
What would you propose as an alternative? If nothing, fine, but how can we relate, when we only have a single best (or if you prefer: flawed) thing?
> It can't handle modelling itself
It can add its internal states to the environment, and hence model these internal states.
> the large dimensionality of its hypothesis space
Its bounded by how many compressors / programs are available to it. Calculating the length of these programs that are consistent with the environment is feasible.
> handle imprecise stimulus information
I'm not sure if you mean imprecise stimuli here or imprecise sensors.
> can be made arbitrarily "stupid"
This can be seen as a flaw, or as a simple property (or even a feature). Something that can be optimal, and "stupid", does not detract much from its ability to be optimal.
> before we get to mere computational limitations or the uncomputability issue
Just because it is uncomputable, does not mean the theory is flawed. Sure, it is not practical, and we like practical things, but it is still valuable to have such a theory. Especially when approximations do yield practical applications.
Marcus Hutter on these issues: http://hunch.net/?cat=14
Best of luck. Do you have a hypothesis? More experienced ML folks don't seem to be on board [1] [2]. I haven't seen anyone posit how to build a machine that can set its own goals.
I'm sure you'll build and learn something cool nonetheless
Another organization that tried to build AI was called Thinking Machines [3]. They were good at parallel programming. A few of them would later form "Ab Initio Software" [4]. Their motto is "from first principles," which Elon likes to say. Socrates did too.
My first job out of college trained us on that software. It's easy to use even if you're not a programmer, which was impressive for parallel software in 2004. I still get job emails for having it on my resume. The company is private and does not share its training resources. Boston mentality, I guess. Despite being so closed, they appeared to be quite successful in 2004 and have managed to stick around.
[1] http://hunch.net/?p=3692542
[2] https://www.facebook.com/yann.lecun/posts/10153442884182143?__mref=message_bubble
[3] https://www.technologyreview.com/s/406781/thinking-machines/
[4] http://www.abinitio.com/
Yes, it is the case here that a bayesian approach can lead to a previous empirical formula falling out. What I was saying as well is that this regularization + test set approach is also valid (and sometimes slightly more or less general than the bayesian approach, since, for example, SVMs fall naturally out of thinking about regularization but they have no analogue in bayesian classifiers).
It also goes the other way, and some formulas are first proposed in a more bayesian-ish context and then extended to some simpler-looking empirical formulas (for example, the jumps from hidden Markov models to max-ent Markov models to conditional random fields to max-margin Markov networks to structured SVMs).
There are more approaches to machine learning, and in John Langford's blog there is a nice table showing the merits and flaws of many of them: http://hunch.net/?p=224 . But you must keep in mind that you can find many equivalencies between these approaches (boosting for example can be seen as a loss minimization with regularization, and max-ent can be seen as a special case of a bayesian model, etc).