The Dirac Sea

Machine Learning and CS stuff

[mgsroom]

Aprendizaje más allá del aula

Cuando pensamos en capacitación usualmente nuestra mente nos lleva de vuelta a las experiencias escolares. A nivel profesional, no hay más que un aula fría con un instructor al frente. Una capacitación más obligada en tiempo y forma por el jefe.

Sin embargo, esto ya no es del todo cierto, hoy en día el acceso al conocimiento es diferente; todos hemos visto algún tutorial en internet, nos hemos inscrito a un webinar o  a un curso en línea. El avance tecnológico nos ha enseñado que la información se encuentra  disponible para cuando necesitemos de ella.

Keep reading

[leonpalafox]

It's no mystery that Zoubin is one of my favorite researchers, his papers are usually insightful and well explained.

This is hands down the best non-technical survey I've read on the topic.

[leonpalafox]

Just an amazing visualization of some of the people related to Data Science and Machine Learning on Twitter

[leonpalafox]

Some years ago, I was trying to understand how message passing worked, and Yaroslav answered my question in an incredibly detailed way.

[leonpalafox]

[leonpalafox]

In NLP, there are usually large datasets that have many sparse values. Often, it is useful to reduce the dimensionality of the data so we have more compact representations of the data.

This paper presents some nice results on the effect of reducing the dimensions in SVMs.

[leonpalafox]

If you want to know a bit more on what do the eigenvalues in the matrix mean, this is a nice place to start

[leonpalafox]

Whoooa, here I go for the motherload of scientific key words this week, this is one of the best, deep and most gentle tutorials to learn and implement deep learning algorithms, read this, understand it, and you are ready to go.

[leonpalafox]

Some of you know that my PhD work is based on finding regulations between genes. The problem is very interesting, and one of the benchmark datasets everyone uses is the E.Coli genome. 

[leonpalafox]

Recursive Neural Networks are a nice way to use a Neural Network structure to represent time series data.

However, more often than not, training is not as straight forward, because you can have problems with the gradients, as the whole thing is essentially a big loop.

This is a nice explanation on why does this happen, I would add to the paper that not surprisingly, a couple of cluster nodes and Differential Evolution are usually good to solve small RNNs

[leonpalafox]

A problem of many Non Parametric Bayesian approaches is that the data is usually exchangable, which for NLP and classification is usually a nice property, however, when we deal with problems like EEG classification or Gene Regulations, the anture of our data makes it hard to impose exchangable priors.

In this survey, we have a good explanation with nice references to Non Exchangeable priors

[leonpalafox]

The best intro to Gibbs Sampling you will ever read

This paper by Resnick does what many papers don't, explain things well, slow and with examples.

It explains 3 fundamental concepts very well: Bayesian Inference, Gibbs Sampling and Probabilistic Inference.

Be sure to read it.