Friday Assorted List
November 18, 2015
I read and learn about a lot of things on the internet. I share with you some of my findings from this week:
- The engineers at Sift recently published
an interesting article with a great insights
about their experience with decision forests trees and a few past
models for their fraud detection engine. They ended up with an
in-house customized solution. Here some of the resulting statistics:
- Hundred of trees
- Over 10s of million of samples
- A few hours of training
- Significant infrastructure cost saving
Science Isn’t Broken: Would you like to hack some p-values? This great article go in depth into why p-values, which is defined as the probability under the null hypothesis, can in a lot of cases be harmful and damage your conclusions. P-values tend to overstate the evidence against the null, and are thus very “trigger happy”.
Streaming data processing is attracting a lot of attention these days. This is a big deal. But a lot of streaming systems are either immature or lack in essential abilities. Their complexity lies behind the very wide range of applicability. I recommend you to take a look at Spark Streaming which is a well maintained and promising data processing project. Some engineers at Google wrote a paper about Dataflow and provide us with a lot more details on the subject.
The Person I Call Most for Startup Advice. Insightful wide-ranging conversation between Tim Ferriss (@tferriss) and Naval Ravikant (@naval) on startups, ethics, mindfulness, self-improvement, evolution and more. There is plenty of references you might want to take some notes.
LSTM Recurrent Neural Networks predominantly occupied of lot of my time last year at Google. They look quite complex but are very fascinating. Traditional recurrent neural networks naturally incorporate memory and time-based dependencies which allow them to learn and predict from sequences. (Speech recognition, text processing,…) However, they struggle to learn long term dependencies as the error signal vanishes rapidly when getting backpropagated in the past. This is where LSTM RNNs come in action: Understanding LSTM Networks.
A/B Testing with Hierarchical Models in Python: There were a lot of hype in the industry lately about A/B testing methods. As useful as they can be to your business, a few practices will tend to increase the likelihood of false positives and mislead your decisions.
Eigenvalues and Eigenvectors were one of the most counter-intuitive and weird notion I’ve learned in linear algebra. And yet, they represent the underlying foundation of so many concepts in machine learning. Here is good introduction by deeplearning4j.
- Very complete post about How does a relational database work.