RT @shivon: Cards Against Machine Learning

RT @latentjasper: Bayesian optimization played a significant role for AlphaGo! "prior to the match with Lee Sedol, we tuned the latest Alp….

RT @Reza_Zadeh: Best paper award #NeurIPS2018 main idea: Defining a deep residual network as a continuously evolving system & instead of up….

a very inspiring personal story on AI research. It’s not a privilege reserved solely for big organizations anymore. People can do it alone nowadays as long as they persevere and work hard. Kudos to Alexia‘s well-deserved success 👏👏.

There’s a lot more to learn from community, though. I don’t think incremental research is bad. Success/failure, we need them to move forward. But we should step back sometimes and see a big picture. Simplifying is the core of (Bayes/AI) learning, I think.

Anyway, I’m delighted that Bayes got revived across many fields. 10 years ago, I went to Paris to learn about ML/AI (I studied robotics before that). I got stuck with Bayes ever since. Bayes AI still has a very long way to go, but I think it’s the only way we can achieve Human AI.

he also used Negative Binomial (NB) dist. to fit and deduce the rate of referee’s acceptence (35%).It seems interesting. Without NB, I couldn’t derive closed form MAP (i.e optimal) number of PCA components in my recent paper either.

In paper “You’ve come a long way, Bayesians”, he summarized 40 years of Bayesian phylosophy. Here he also explained intuitively why Bayes is consistent with Axiom of Probability (check out his paper “Varieties of Bayesianism” or my PhD thesis for details).

When looking for NeuralNet python tutorial for beginners, I came across a nice blog of Prof. @jweisber at Uni. Toronto @UofT . He used the same excellent Coursera of @AndrewYNg but with simpler form. What’s more. it turns out that he is a cool Bayesian!.

.“The most powerful A.I. systems… use techniques like deep learning as just one element in a very complicated ensemble of techniques, ranging from the statistical technique of Bayesian inference to deductive reasoning”. A good read from Prof.@GaryMarcus !.

A nice illustration of SGD’s drawback. Extracted from their recent paper:.

A really cool paper on how to run Gradient Descent on Bayesian NeuralNet efficiently (and, as extra point, it’s relevant to my Copula Variational paper 😀). They also showed that traditional SGD over parameter has a fundamental flaw (it may not converge to the optimal dist.) 1/2.

“Two approaches in particular stand out to Herbrich. One is the revival of a technique called Bayesian learning (. ).The other machine learning technique that Herbrich sees drawing increased attention is something called a spiking neural network”.

RT @Reza_Zadeh: Best Paper Award #ICLR18: finds flaw in ADAM proof of convergence, shows failure on specific toy convex problem & suggests….

I solved a 50-year-old challenge for closed-form of optimal estimation of VA in Y=VA+Z in PCA, without overfitting. We found SNR > -10 dB is limit. This is my first step toward avoid overfitting in #NeuralNet completely!. paper code

Our ICASSP paper is published! Journal version is up tomorrow. Given Y = VA + Z, we found that non-overfitting limit for estimating V,A is SNR > -10 (dB) for PCA/MUSIC. paper: code ( + journal): explanation:

I have a feeling that Sir Atiyah has solved it. His approach is pretty clever, since pi,e and alpha is a fundamental connection between maths and physics. I am now reading his proof carefully, anyone cares to join me xD? . His proof is already up:

the biggest hit-or-miss of maths today.current livestreams on one of hardest Riemann Hypothesis proof of Sir Atiyah!. finger crossed!!!!! I wish him wellll.

“People say ‘we know mathematicians do all their best work before they’re 40’”, says Atiyah. “I’m trying to show them that they’re wrong. That I can do something when I’m 90.”.

RT @nolimits: When interviewing job candidates, I always refer to this passage from Jeff Bezos' 1998 shareholder letter. It's a great, simp….