The Hacker Within Berkeley

scikit-learn - Ross Barnowski and Shannon McCurdy

November 18, 2015 at 5-6:30pm in BIDS, 190 Doe Library

Attending

Meeting Info

Ross Barnowski

Ross is a Nuclear Engineering PhD student in Kai Vetter’s group.

Shannon McCurdy

Shannon is a postdoc in computational biology.

Discussion: scikit-learn

Ross walked us through a demo notebook which can be found here. You can clone it from github.com/thehackerwithin/berkeley.

Shannon walked us through some useful resources. The documentation for sklearn seems to parallel a book called The Elements of Statistical Learning, and Shannon recommends this as a resource.

Linear Regression

If y is nx1 and x i nxp, we have an unknown coefficient matrix W, which is px1. The error term is then nx1. The assumption is that x and y are linearly related. The fit, W, minimizes the vertical error. The least squares cost function, which comes up in regression in this way, is a model for the error.

Note that in this example, when p>n, we enter a danger zone for validity of this model. Shannon wanted us to note, in this context, scikit-learn doesn’t necessarily warn you when this happens. So, don’t trust that scikit-learn will always warn you if you aren’t using the models in the appropriate regime.

Shrinkage Models

A bunch of different shrinkage models are included in scikit-learn. One that Shannon uses in her work is Lasso.

The idea, functionally, is that we add a penalty to the least squares cost function. The penalty is related to the magnitude of each coefficient. That is, if you are going to add some nonzero element in the matrix, it must contribute well to the fit with y. This is a parsimony metric which enforces sparsity in the solution vector. This helps with interpretability because it emphasizes the most important coefficients.

In the Wild

Shannon has encountered least squares and lasso in two different problems in her work.

Example: In her research she looks into event times, where only a subset (half) of the events are recorded. Using an exponential probability and an indicator (whether or not an event was recorded), she can describe the probability of an event happening. Given this, she can separate the probability into a maximum likelihood problem which can be minimized (using exponential regression) to determine the least squares soluation and she can reframe the Newton-Raphson step into an ordinary least squares lasso situation. If you didn’t follow this completely, check out Tibshirani’s website on the general topic of lasso models.

Lightning Talks

Finally, there will be a time for a couple of Lightning Talks, which are 5-10 minute blasts of information about a particular topic or question of interest to the group. This topic can be anything useful, new, or interesting to scientists who compute. It may be some new skill you have recently picked up in your research, a productivity tool you have recently learned to love, a quick demo of a useful library, or anything you feel we would enjoy learning.
Note that the lightning talk time is a good way to bring a question to the group. If you have a bug you need help with, here’s the place to ask many ears about it at once.

Name : Topic

Notes and links

Name : Topic

Notes and links

Share