abhgh

69 ms·

331.
by abhgh 8y ago
I am not sure if the author did this (he certainly doesn't mention it), but a lot of what constituted Watson at the time of Jeopardy was published as a set of 17 papers called "This is Watson" in the IBM Journal Of Research a
332.
by abhgh 8y ago
Oh wow, Roger Schank [1]! Haven't heard that name in a while - he was quite famous in the early days of AI. I wonder if he has figured out a good way to marry ML to his theory of Conceptual Dependency (CD) [2] - because that would coul
333.
by abhgh 9y ago
I think, not much. This is an optimization problem to which you can apply any algorithm that "fits" --- no closed-form functional form, interactive feedback --- so bayesian optimization (with GPs), some form of RL, evolutionary al
334.
by abhgh 9y ago
The AI that is not explainable is not because it cannot log things, its because the semantic interpretation of what it can log is hard. Starting with the real world input (which we understand) a lot of algorithms progressively apply mathema
335.
by abhgh 9y ago
My school had a two year course on programming in C\C++; fairly basic, being an introduction to programming itself. At that time, I didn't own a computer, and to use the computers in the lab you needed to wait an extra hour and arrange
336.
by abhgh 9y ago
Adding to the credibility, this was also accepted to NIPS-2017.
337.
by abhgh 9y ago
I completely agree with you on knowing the math part. I am kind of emphatic about it too when doling out advice. Without that, getting good results just becomes a blind exercise in tweaking parameters. With that, you can really troubleshoot
338.
by abhgh 9y ago
Support Vector Machine. A Machine Learning technique that's used typically for classification and regression, but also has been adapted to novelty detection, stuctured prediction, ranking etc. I wrote up a beginner's tutorial some
339.
by abhgh 9y ago
I was reading the discussion on my mobile, but opened up the page on my desktop (where I am logged in), just to say that this is an amazing little tool! Search is blazing fast. I have been a happy user for the past 5 years.
340.
by abhgh 9y ago
Over time I have settled into the following "personal best practices" of sorts to track my research: 1. I add papers that I read to Mendeley. This makes them accessible across devices. Also makes them searchable. Before: Used to k
341.
by abhgh 9y ago
TD-Gammon was taught to us as a part of a classroom course on Reinforcement Learning (RL) in 2007 . ML was known to a small set of people back then, there weren't many jobs in the area (this is in India), and even to many in this set,
342.
by abhgh 9y ago
For me, I think, the primary reason an open office often doesn't work (I say 'often' because sometimes it does) is that being physically and visibly present seems to signal that you are available for answering sporadic querie
343.
by abhgh 9y ago
K-NN is one of the more concise classifiers to implement. I did a Python implementation a while ago that can fit into a tweet -[1]. Since maps/lambdas are available in Ruby, this should be possible in Ruby too. Sorry for the bad presen
344.
by abhgh 9y ago
I'm sorry I didn't have a chance to respond before -you've been very patient with your responses! Thanks! Yes, this is quite neat! I will look up the rule/theorem you mention. It seems "kind of" reasonable in t
345.
by abhgh 9y ago
Oh, I see it now. Thanks for clarifying! In that case, you needn't have kept the denominator around at all. Since P(w) is not based on a class i.e. its not class conditioned, the classifier could have directly calculated P(C_1|w)/
346.
by abhgh 9y ago
> It is hard for me to empathize with Musk’s fixation on evil super-intelligent AGI killer robots in a very distant future. This reminded me of a talk by Andrew Ng I watched recently [1]: "Worrying about evil AI robots today is a li
347.
by abhgh 9y ago
Deep Learning is in vogue so I think we see a lot around it written about in blogs, media etc. GBDTs are used in the industry though.
348.
by abhgh 9y ago
I may be missing something, but how do you take a log on the RHS of p(c_1|w)? The denominator has a sum i.e. "1+ something", which the log doesn't distribute over.
349.
by abhgh 9y ago
Accuracy is a problem. NBs are linear classifiers, and more so, the linear boundary is found by assuming features to be independent. So, in cases: (1) data cannot be separated linearly - NBs don't do well for the same reason as any oth
350.
by abhgh 9y ago
A side note here: the classification probabilities NB produces are not very accurate and usually need some correction using a process called "calibration".
351.
by abhgh 9y ago
From experience, I suspect NB does well on text primarily for two reasons: (1) High dimensionality - this is also the reason why SVMs with linear kernels do so well on text (NBs find a linear boundary too - but a different linear boundary t
352.
by abhgh 9y ago
Probably relevant, there has been work on converting feed-forward networks to decision trees e.g. [1]. EDIT: why is this relevant to the parent comment? You can learn a neural net, convert it to a tree, and then integrate rules from domain
353.
by abhgh 9y ago
Note though that the moment you go beyond single decision trees, with say random forests, boosted trees or bagged decision trees, the models become less and less interpretable. They "look" like rules, but with multiple disjoint se
354.
by abhgh 9y ago
I agree. I have a hard time understanding pricey paywalls for academic publishing. The research is not funded by the company that paywalls, the researcher is not paid to do research or paid royalties by the company, and most importantly, th
355.
by abhgh 9y ago
I did exactly the same. Not on a regular basis though, but on days when work was tiring I used to head out to the parking lot and sleep in the back of my car. Typically 20 min. That made me much more productive than I should have been w&#x2
356.
by abhgh 9y ago
This is not really true. Aside from ensemble models they tend to perform pretty much at par or better. Here's an extensive comparison by Rich Caruana [1] [1] https://www.cs.cornell.edu/~caruana/ctp/ct.papers&#
357.
by abhgh 9y ago
Have been in the industry for quite a while, and I think one of the often-overlooked non-technical reasons is they are hard to get an intuitive feel of, compared to ANNs. This might seem irrelevant, but think about it - someone who doesn&#x
358.
by abhgh 9y ago
Actually I used to read in lunch hour. I used to specifically eat lunch later than my peers so that I don't end up going in a group (I know, I know, I was picking reading over society ...). But of late my workload has increased a lot,
359.
by abhgh 9y ago
I've heard good reviews. Have had this downloaded on my mobile for a few years now ... should get around to reading it "soon" -_-
360.
by abhgh 10y ago
Serre understands the nut, Grothendieck understands the universe :-)

More ›