Data science musings

[graphicteadatascientist]

It turns out that most regex flavours don’t consider letters like æøå to be actual letters. Shortcuts like \w is predicated on the ASCII set, not Unicode. Except the Python regex engine does handle Unicode also. And I grew up on Python so I had never run into this issue before. I specifically ran into it in Amazon Sagemaker Data Wrangler where it stripped out some characters that I didn’t expect and then I was freaking out that my code might have been bollocks. But my code is Python and it’s fine. Data Wrangler seems to run on Spark so maybe it’s the Scala / Java regex flavour? Dunno.

[graphicteadatascientist]

by Daniel Godoy Explanations of the differences between glorot (xavier) and he kernel initializations, both normal and uniform.

Such a good description of the ins and outs of initializing weights in a neural network with lots and lots of plots (and also math). I loved it!

[graphicteadatascientist]

When I discovered Python’s new pathlib module a few years ago, I initially wrote it off as being a slightly more awkward and unnecessarily …

The Path class from Python’s built-in pathlib module is so very wonderful. Example:

from pathlib import Path

import pandas as pd

folder = Path(../data/)

filename = ‘mydata.csv’

df = pd.read_csv(folder / filename)

You write a directory path in Linux/MacOS style and Path will convert it to Windows when necessary. To add folders and filenames to a current Path object you simply use the division operator, /, and strings. Path has cwd and glob methods. It’s brilliant!

[graphicteadatascientist]

Oliver Sherouse writes on Data, Python, and Culture

If you are using Python you are going to be using Anaconda. And if you’re on Windows (because work is being mean to you or some such) I highly, highly recommend that you install the m2-base package. 

I first heard of m2-base on Stack Overflow but here’s a nice blog write-up from some person. 

m2-base has all the GNU core utils (cp, ls, mv, rm, head, wc, cut, join, cat, tr, sort, uniq, and so on) as well as essential basics like grep, sed, time, and gawk. 

[graphicteadatascientist]

Select multiple instances of the current selection using `Control+D` or `Command+D` in a Jupyter Notebook.

Same as the one for Jupyter Lab except this works with standard Jupyter Notebooks. Scroll all the way to the comments to find the part that you should actually paste into your custom.js - I couldn’t tell you why the post hasn’t been updated. 

[graphicteadatascientist]

A slightly opinionated Sublime notebook cell keybinding for JupyterLab. - ryantam626/jupyterlab_sublime

This seems trivial but tbh it is my no. 1 productivity tip - at least for Jupyter. If you use Jupyter Lab you can install this extension and get the sublime keybindings - for me it’s all about two things:

1. you get ctrl+L which selects the entire line where the cursor is

2. you get ctrl+D which adds the next occurrence of your selection to the selection. Sounds confusing but if you select/highlight ‘max’ in a cell and press ctrl+D then it will also select the next ‘max’. Do ctrl+D again and you have three selections of ‘max’. You can now delete all three at the same time, you can use ctrl+V to paste something over three times, you can overwrite all of them at once (maybe you need ‘min’ instead), or you can use the arrow keys to unselect but then have three cursors at the right locations (e.g. if you wanted to write ‘arg’ in front of all the ‘max’). MACS! Command+D, not ctrl+D.

[hackersandslackers]

Welcome to SQL: Modifying Databases and Tables

Brush up on SQL fundamentals such as creating tables, schemas, and views. Part of a multi-part series on learning SQL.

Brush up on SQL fundamentals such as creating tables, schemas, and views.

[datascience-from-the-trenches]

Joining columns across TSV files

This is a beautiful and useful tool developed by eBay:

Try it out: https://github.com/eBay/tsv-utils/tree/master/tsv-join

[graphicteadatascientist]

The AI Hierarchy of Needs – Hacker Noon

As is usually the case with fast-advancing technologies, AI has inspired massive FOMO , FUD and feuds. Some of it is deserved, some of it…

Data engineering is incredibly important and in many online resources about machine learning it gets stuffed into “Normally, you would not work with such a nice data set and would have to do a bunch of data cleaning first”. If your project is a one-shot then we could probably call it data cleaning - basically just creating some csv files with the features and labels of interest. 

However, in many projects you want to either continuously update your model or produce new models as more data becomes available. Here you would have to create a full data infrastructure, probably with a relational database or maybe a streaming data platform, before you have a data science product that creates value in your company. This is where tech start-ups have an advantage - often their whole existence is predicated on having well-structured data infrastructures.     

[graphicteadatascientist]

The rare form of machine learning that can spot hackers who have already broken in

Darktrace’s unsupervised-learning models sound the alarm before intruders can cause serious damage.

Machine learning and cyber security is of course very interesting in itself but I mainly wanted to share the part about unsupervised learning shining when working with data with only a few examples of the class of interest. 

“[...] supervised-learning algorithms work best with balanced data sets—in other words, ones that have an equal number of examples of what it’s looking for and what it can ignore. Cybersecurity data is highly unbalanced: there are very few examples of threatening behavior buried in an overwhelming amount of normal behavior.

Fortunately, where supervised learning falters, unsupervised learning excels. The latter can look at massive amounts of unlabeled data and find the pieces that don’t follow the typical pattern. As a result, it can surface threats that a system has never seen before and needs few anomalous data points to do so.”

[graphicteadatascientist]

Machine learning, meet quantum computing

A quantum version of the building block behind neural networks could be exponentially more powerful.

Quantum computing gets a lot of hype but people rarely go into the specifics. Here they are using qubits to encode perceptrons: 

“The big advantage of quantum computing is that it allows an exponential increase in the number of dimensions it can process. While a classical perceptron can process an input of  N  dimensions, a quantum perceptron can process 2N dimensions. 

Tacchino and co demonstrate this on IBM’s Q-5 processor. Because of the small number of qubits, the processor can handle N = 2. This is equivalent to a 2x2 black-and-white image. The researchers then ask: does this image contain horizontal or vertical lines, or a checkerboard pattern?“

[graphicteadatascientist]

Learn Functional Python in 10 Minutes – Hacker Noon

In this article, you’ll learn what the functional paradigm is as well as how to use functional programming in Python. You’ll also learn…

While I’m doing this: 

Here’s a really good article on functional programming in Python. If you are permanently baffled by lambda, map, filter, reduce then this is for you. The examples are not terribly practicably applicable though, instead they demonstrate the behaviour of the code. 

[graphicteadatascientist]

Hundreds of digital flashcards on machine learning topics in DRM-less print-quality png, web-quality png, PDF, Anki, and SVG.

Chris Albon made these very cute flashcards for machine learning. I think $12 is a quite reasonable price but if you don’t they could also inspire you to make your own. Idk your life.

[graphicteadatascientist]

Relationship between SVD and PCA. How to use SVD to perform PCA?

Principal component analysis (PCA) is usually explained via an eigen-decomposition of the covariance matrix. However, it can also be performed via singular value decomposition (SVD) of the data mat...

A super-post about single value decomposition (SVD) and principal component analysis (PCA) with a ton of links and info. 

[graphicteadatascientist]

'I was shocked it was so easy': meet the professor who says facial recognition can tell if you're gay

Psychologist Michal Kosinski says artificial intelligence can detect your sexuality and politics just by looking at your face. What if he’s right?

Obviously Michal here is a piss-poor data scientist and and there’s already heaps of articles and blog posts out there ripping his research methodology apart (he’s a bit like Svante Pääbo in that sense).

But he can be used a sort of ethical yard stick. Is this something he would have done? Then it is most likely unethical. His defence boils down to “if I hadn’t done it then someone else would” which supposes that all possible research will be pursued - as if time, resources, and choice are irrelevant. And he lies several time in this interview. The “gaydar” paper was never about to get scrapped. He paid those Amazon mechanical turk workers I don’t know how much money for their half-assed work.  

[graphicteadatascientist]

Look Ma, No For-Loops: Array Programming With NumPy – Real Python

How to take advantage of vectorization and broadcasting so you can use NumPy to its full capacity. In this tutorial you'll see step-by-step how these advanced features in NumPy help you writer faster code.

A frequent complaint about Python is that its calculations are slow due to dynamic typing and probably a bunch of other factors. Most intensive calculations in Python is however relegated to the library NumPy which is implemented in C. NumPy is extremely popular and the basis for most Python data analysis libraries. E.g. pandas is based on NumPy. 

This article is not an easy read but it is a very good walk-through of how NumPy works and how to make the most of it. 

[graphicteadatascientist]

Author summary Computers are now essential in all branches of science, but most researchers are never taught the equivalent of basic lab skills for research computing. As a result, data can get lost, analyses can take much longer than necessary, and researchers are limited in how effectively they can work with software and data. Computing workflows need to follow the same practices as lab projects and notebooks, with organized data, documented steps, and the project structured for reproducibility, but researchers new to computing often don't know where to start. This paper presents a set of good computing practices that every researcher can adopt, regardless of their current level of computational skill. These practices, which encompass data management, programming, collaborating with colleagues, organizing projects, tracking work, and writing manuscripts, are drawn from a wide variety of published sources from our daily lives and from our work with volunteer organizations that have delivered workshops to over 11,000 people since 2010.

This is such a good paper and pretty much any scientist that wants to do anything with code should read it. Maybe print out some key parts and hang them in your office.