Over the past decade, statistical thinking has gained ground among forensic practitioners and researchers. Machine learning methods, one such type of thinking, are quickly evolving as useful tools to answer the question of source: is the suspect the source of the evidence found at the crime scene?
Machine or statistical learning, as the name suggests, encompasses a collection of algorithms where we “learn” about patterns and associations in a dataset without relying on a specific statistical model. Early research suggests that this type of methodology is promising, but since the field is new to the forensic community, opportunities for training are still limited.
The AAFS Annual Scientific Meeting has not had a session on this topic (that we know), but CSAFE Director Alicia Carriquiry, researcher Heike Hofmann and board members Jeff Salyards and Robert Thompson created a specialized workshop for 2020 attendees. Fifty participants from a wide variety of forensic disciplines along with several lawyers joined our team to learn more.
“The concepts of machine learning may be confusing to individuals without specialized training in the field. Our role at CSAFE is to dispel some of the mystery and make the subject accessible to all forensic science practitioners and lawyers. Our philosophy is that there are no silly questions,” Carriquiry said.
The CSAFE workshop introduced attendees to the basics of supervised learning algorithms in the context of forensic applications, including firearms and footwear examination and trace evidence. A closer look at what participants learned is outlined below.
The Basic Principles of Machine Learning Applied to Forensic Science
What are the forensic tasks at which learning algorithms and machine learning excel? An obvious one is classification, or determining a set of attributes that allows us to place an item in one or more previously defined classes. An attribute is simply a characteristic of the item that can be measured or observed. An example of an attribute is the concentration of a certain compound in a sample of white powder or the length of the femur in a human skeleton. If certain values of those attributes are associated with the class of an object but not with others, then we can use the attribute to associate an item with a class. For example, if the concentration of the compound is high, that might indicate that the powder is cocaine.
Learning algorithms are built through determining the values of the attributes or features that are associated with the different classes and then building a classification rule or classifier. A classifier is just a set of rules that allow us to predict the class of an item based on information about its features. Classifiers uncover associations between features and the class to which object belongs.
Supervised learning algorithms learn about those associations from data for which we know ground truth, the labels associated with each item. These data are called training data. For example, we may have float glass fragments from five different (but known) manufacturers (classes) for which we measure elemental concentrations (attributes or features). The training data allow the algorithm to learn the link between the classes and the values of the features. The goal is to build a classifier that makes few mistakes when classifying items it has not seen before.
How exactly might this apply to forensic science? A few examples are listed below.
- Anthropology: classifying skeletal remains.
- Blood spatter analysis: was the victim standing or sitting?
- Chemistry: is this compound A, B or C?
- Pattern: were these two bullets fired by the same gun?
“Pattern evidence analysis is particularly challenging because images of evidence often have thousands of pixels and are highly multivariate and not well-suited for standard statistical analyses. Learning algorithms enable us to calculate the similarity between images with accuracy and objectivity,” said Carriquiry.
CSAFE’s Role in Community Education on Machine Learning
CSAFE team members are among the foremost experts in the application of machine learning methods to forensic problems, in particular to the analysis of pattern and digital evidence. As researchers develop new ways to integrate machine learning into forensic practice, the center continues to create additional educational opportunities for the forensic science community around this topic. As we continue developing methods and learning about their use in forensic problems, practitioners can expect to see more online and in-person training programs offered in this area.
“The use of learning algorithms will only increase as the measurement of features in various types of evidence improves, particularly for pattern evidence. As machine learning becomes more integral in forensic science practice, organizations like CSAFE will be crucial to helping the community successfully implement these methods,” Carriquiry said.
To inquire about hosting a training in machine learning for your organization, please complete our training form on the CSAFE website. This will help our team determine your needs and goals. Look for CSAFE hosted workshops in your area by visiting our events page, and stay up-to-date on upcoming online learning opportunities from CSAFE by signing up for our newsletter.