Evolving notes, images and sounds by Luis Apiolaza

Author: Luis (Page 4 of 66)

Old dog, etc

Yesterday I attended an introduction to “Artificial Intelligence in Forestry”* workshop at the School of Forestry, University of Canterbury. There were plenty of industry people, including some of I taught “a few” years ago.

I am mostly a quantitative genetics/stats person, but I keep my eyes open for anything that could be helpful dealing with my genetics trials, and have dabbled on using spatial things, LiDAR, etc. As I have pointed out in the past: 1. I am a sucker for entertaining problems and 2. old dogs can learn, albeit more slowly than 30 years ago, new tricks.

It was a nice opportunity to dip my toes in applications of machine learning for image analysis (detection and classification), mostly using ArcGis Pro and a short exercise using R. Funnily enough, my biggest barrier in the workshop was dealing with Windows + ArcGis. I am a Mac and Linux person, generally dealing with coding statistical analyses writing commands in text files. Most of my GIS related work is done via R (which I’ve been using since 1997) and any interactive work with maps using QGIS. Finding anything with point and click took me ages, until we got to the R part, which was hard for most people but kind of easy for yours truly.

Most importantly, having a chat with people and getting ideas to try with my data were the coolest part of the day. Thanks to Vega Xu, Justing Morgenroth and Ning Ye for organising the course!

*Confession: I dislike attaching the AI label to everything these days. Perhaps a better name would have been “Machine Learning in Forestry”. Small detail.

Classroom view from the back.

More than heritability 🎶

It is easy to get obsessed with heritabilities when you start working in breeding and genetics. The idea that a fraction of that variability we are observing can be explained by pedigree (or family structure or clonal differences or whatever) is appealing. It gives us an idea of control: there is a whiff of causality in our work. If the trait I care about is heritable THEN I can successfully breed for it.

However, we need to remember that it is a fraction of the variability we care about. If there is little variability, there is little room to select farther to the right of the distribution for the trait.

A typical example in trees would be stem diameter vs basic wood density. The heritability for density is, on average, around 0.6, while for diameter is 0.2 in a lucky day. However, the within-site coefficient of variation for density is about 8% but over 20% for diameter. Much more room to move in diameter.

Nothing groundbreaking, particularly if you have been working for a while. Just a handy reminder if you’re a newbie in breeding things.

Gratuitous earworm: In my head, “more than heritability” sounds like the famous More than a feeling.

Lecture on enshitification

Cory Doctorow’s “My McLuhan lecture on enshittification”. How we got to the current dystopia and how we can fix it. We need more competition, more regulation, more self-help, and more empowered workers.

This applies to any industry; yes, even to your niche job, and to mine.

Python not suitable platform for reproducible research

While [Active Papers] has achieved its mission of demonstrating that unifying computational reproducibility and provenance tracking is doable and useful, it has also demonstrated that Python is not a suitable platform to build on for reproducible research. Breaking changes at all layers of the software stack are too frequent.

Konrad Hinsen in Archiving Active Papers

I started using Python for my PhD around 1997, to control simulations I wrote using Fortran 90. I chose Python based on Konrad Hinsen’s writings at the time in a long-disappeared website. A few years later I moved all my work to R, which I found much more stable. I have some 20-year-old R base code that still runs. 😇

Incidentally, last year I wrote a series of posts on Some love for base R.

« Older posts Newer posts »

© 2024 Palimpsest

Theme by Anders NorenUp ↑