Understanding data: How statistics and machine learning learn from samples

There is an analogy in data science between inferential statistics and machine learning. Both fields aim to extract knowledge from limited data.

Imagine wanting to understand the health of an entire forest. It’s impossible to examine every single tree. Instead, a researcher might take a random sample of trees and measure their height, trunk diameter, and signs of disease. Using inferential statistics, they can estimate the health of the forest based on that sample.

Likewise, a machine learning model could be trained to predict forest health from aerial images. The model would analyze a dataset of images labeled as “healthy” or “unhealthy”. It would learn to identify subtle patterns in color and texture that distinguish between the two states. Once trained, the model could assess the health of vast areas of forest much faster than a human observer.

A key similarity between the two approaches is the concept of learning from samples. Both fields involve moving from the specific (a sample) to the general (a population or a broader pattern). And both fields acknowledge uncertainty. Statistics relies on concepts like confidence intervals to quantify uncertainty, while machine learning models always carry the risk of prediction errors.

The notion of using a limited set of data to gain insight into a larger population or phenomenon is not unique to statistics and machine learning. It is fundamental to many areas of research and decision-making, from scientific research to manufacturing quality control, to market research.

Image above: example of different defoliation levels, from left to right: 0%, 10%, 25% and 55% (source: https://medium.com/ecovisioneth/automating-tree-health-monitoring-from-images-with-machine-learning-1df21c44c81a)



Cite this blog post
Stéphane Condorelli (2024, March 19). Understanding data: How statistics and machine learning learn from samples. Exploring Large Language Models. Retrieved June 22, 2024, from https://doi.org/10.58079/w76l

Leave a comment

Your email address will not be published. Required fields are marked *

This site uses Akismet to reduce spam. Learn how your comment data is processed.

Search OpenEdition Search

You will be redirected to OpenEdition Search