SSB Debate: Model-based Inference vs. Machine Learning

Transcription

1 SSB Debate: Model-based nference vs. Machine Learning June 3, 2018 SSB 2018 June 3, / 20

2 Machine learning in the biological sciences SSB 2018 June 3, / 20

3 Machine learning in the biological sciences and in our meeting! SSB 2018 June 3, / 20

4 ntroduction to topics and terminology Machine learning: Collection of algorithmic methods for pattern recognition, classification, and prediction, that may be based on models derived from existing data Model-based inference: Techniques in which probability models are defined and then fit to a data set, with the goals of evaluating model fit, estimating parameters, or testing hypotheses SSB 2018 June 3, / 20

5 Example Motivating problem: Given a set of measurements on a sample of lizards, determine species-level memberships for the individuals Data: For each lizard sampled, collect two continuous measurements: snout-to-vent length (SVL) and hind-leg length (HLL) Questions of interest: How many species are represented in the sample? To what species does each individual belong? Given a new lizard, how can we classify it into a species given what we ve learned from our initial sample? SSB 2018 June 3, / 20

6 Types of machine learning algorithms Supervised learning: Lizards in the sample are associated with species labels Hind-leg Length Snout-to-Vent Length SSB 2018 June 3, / 20

7 Types of machine learning algorithms Supervised learning example 1: disciminant analysis Basic idea: Find a discriminant function for each group that can be evaluated at the observed variables and used to classify individuals into groups Example: For these data, the linear discriminant function is SVL HLL Evaluating this for our two groups, we find Green, long-tailed lizards Blue, short-tailed lizards SSB 2018 June 3, / 20

8 Example Hind-leg Length Snout-to-Vent Length SSB 2018 June 3, / 20

9 Types of machine learning algorithms Supervised learning example 2: decision trees SSB 2018 June 3, / 20

10 Example Hind-leg Length Snout-to-Vent Length SSB 2018 June 3, / 20

11 Types of machine learning algorithms Supervised learning example 3: Support Vector Machines (SVM) Basic idea: Find a line (or plane) that produces maximum separation between two groups of data Hind-leg Length Snout-to-Vent Length SSB 2018 June 3, / 20

12 Types of machine learning algorithms Supervised learning: many, many other algorithms/methods exist Examples: Neural networks k-nearest Neighbor (k-nn) classifier Random forests SSB 2018 June 3, / 20

13 Types of machine learning algorithms Unsupervised learning: Lizards in the sample are NOT associated with species labels species must be learned from the data Example data: Hind-leg Length Snout-to-Vent Length SSB 2018 June 3, / 20

14 Types of machine learning algorithms Unsupervised learning example 1: k-means clustering Basic idea: partition the data into k groups such that the sum of squares from the data points to the assigned cluster centers is minimized Example: For our data with k = 2, we have Green, long-tailed lizards Blue, short-tailed lizards SSB 2018 June 3, / 20

15 Example Hind-leg Length Snout-to-Vent Length SSB 2018 June 3, / 20

16 Types of machine learning algorithms Unsupervised learning example 2: Principal components analysis Basic idea: use an orthogonal transformation to convert a set of observations of possibly correlated variables into a set of values of linearly uncorrelated variables called principal components PCA2 (45%) PCA1 (55%) SSB 2018 June 3, / 20

17 Types of machine learning algorithms Unsupervised learning: many, many other algorithms/methods exist Examples: Hierarchical clustering Self-organizing map (SOM) SSB 2018 June 3, / 20

18 Model-based methods Basic idea: Write a formal model and fit the model to the data Assumption is that the model describes the data su ciently well that the desired insights can be made through parameter estimation (Schrider and Kern 2017) Examples: Maximum likelihood Bayesian methods Discriminant analysis?? Logistic regression?? n many cases, a method has elements of both machine learning and model-based inference, and methods are di cult to classify! SSB 2018 June 3, / 20

19 Machine learning vs. model-based methods Now that we ve got the basics down, we can make some broad comparisons Machine learning methods Most often used for classification and prediction Useful when there are LOTS of data (training set, test set, pattern extraction) Useful when the number of predictor variables is much larger than the number of observations Model-based methods Most often used for inference - parameter estimation and hypothesis testing Model fitting may become di cult for large data sets / complex models Methods perform best when the number of observations is larger than the number of predictors SSB 2018 June 3, / 20

20 ntroducing our panelists Cécile Ané Professor, Departments of Statistics and Botany University of Wisconsin-Madison Machine learning Joe Rusinko Associate Professor, Department of Mathematics and Computer Science Hobart and William Smith Colleges Model-based methods SSB 2018 June 3, / 20