- This event has passed.
Machine Learning for Forensic Practitioners Short Course – Session 1
Thursday, September 1, 2022 at 1:00 pm - 3:00 pm UTC-5
This event is scheduled to take place on Thursday, September 1, 2022. It is the first session of a three-part short course. A registration form can be found below.
About the Short Course:
The use of learning algorithms will increase as measurement of features in various types of evidence improve. This is particularly true in the case of pattern evidence. Forensic scientists will greatly benefit from understanding the basic ideas that underpin statistical learning since these types of methods have already been proposed for firearms examination, fingerprints, glass comparison, and shoe print evidence. Most quantitative training for forensic scientists emphasize classical statistical ideas, so a workshop in which forensic practitioners are exposed to learning algorithms is novel and timely.
When a task consists of deciding whether two items are similar enough to suggest that they could have a common source, an alternative approach is to use statistical or machine learning. Machine learning is the term used to refer to a family of statistical methods and computer algorithms that find patterns in data and has been around for decades. There are many different types of algorithms, but a basic taxonomy is to distinguish between supervised learning algorithms and unsupervised learning algorithms. The purpose of this short course is on supervised learning methods.
Supervised algorithms rely on training data, for which ground truth is known, and on test data, on which the performance of the algorithm can be tested. In a simple example, several bullets are fired from a large number of guns. To train an algorithm to recognize whether a pair of bullets was fired from the same or from a different gun, one might begin by creating all possible pairs of bullets, and compute, for example, the difference in the average striation depth for each pair. This difference is a feature, and perhaps it takes on low values when two bullets were fired from the same gun and high values otherwise. Presented with the value of the feature for pairs of bullets known to have been fired from the same or from different guns, the algorithm then “learns” that same-gun bullets tend to exhibit values of the feature in a certain range that is different for different-gun bullets. With this knowledge, the algorithm can then classify other pairs of bullets for which it does not know in advance whether the bullets were shot from the same or from a different gun.
In real applications, the number of features can be very large, and the number of classes can also be large. In classification examples, the response variable—or class—is discrete, but algorithms can also be used when the response is continuous; in this case, the problem is to predict the value of a variable given information on a large number of features.
Each session builds upon the previous one(s) and recordings will be available in the event registrants are unable to attend one or more of the live sessions. Researchers, collaborators, and members of the broader forensics and statistics communities are encouraged to attend. Short course registrants who attend all sessions will receive a certificate of completion.
Kingland Data Analytics Faculty Fellow and Professor, Iowa State University
Distinguished Professor of Liberal Arts and Sciences and Professor of Statistics, Iowa State University
Research Scientist, CSAFE