SHARON CHEBET

HB100/G/5550/18

DATA SCIENCE 2 CAT 2

Pattern Recognition and Machine Learning

Pattern recognition is a scientific discipline concerned with the automatic discovery of

regularities in data through the use of computer algorithms and with the use of these

regularities to take actions such as classifying the data into different categories. In statistics,

algorithms for pattern recognition are sometimes referred to as learning algorithms and can be

used for inference or prediction.

Algorithms for pattern recognition can be grouped into two major classes, the class

of supervised learning methods and the class of unsupervised learning methods. In supervised

learning algorithms the algorithm learns how the features relate to the labels. The idea is that

once the algorithm has been trained, meaning that an estimate has been obtained, the

algorithm can be used to predict the labels for previously unseen objects. Unsupervised

learning algorithms, in contrast, do not depend on the availability of training data with known

labels but instead, the algorithm itself determines the combination of features that best

represent a class and that maximize the probability of assigning the correct label to a new unit.

The Case of Firearm Identification

Firearm identification is one area of forensics in which learning algorithms have been

used to address questions of source. Firearm models, types or even specific firearms may be
automatically detected from captured audio files or continuous audio streams e.g. recording

microphones, microphone arrays and many others using machine learning techniques. The

detection may also be based on captured still images or video files/streams.

Sound records of firearms recorded at manufacturing stage, for example, may be

queried to compare with the captured data in identifying individual firearms. For example,

procedures may be instigated for recording the unique audio profile of each firearm at

purchase or before sale with the results stored in law enforcement databases. This may allow

for linkage of the record for a firearm, including registration number and owner details, to a

crime.

Firearm sounds may be clustered to group firearms by their unique audio fingerprint.

Specific firearm types may be associated with specific crime types using machine learning.

Similarly, wound types and locations may be determined from the sound of the impact, for

instance, likely to have hit or whether the hit is to a bullet-proof vest, etc. Wound types and

locations may also be associated with specific firearms allowing predictive analysis and input

for protection mechanisms. Similar techniques may also be used to detect and classify types

and quantity of explosive material.

Data Capturing

This involves acquisition or mining of data from open data sources. In firearm

investigations, devices that may assist in capturing crime scene data are still image camera,

access entry system, microphone closed circuit television (CCTV) systems, security system

devices, bystander smartphones and video camera may be used to capture audio, image, video
and other data associated with the crime scene. The captured data may be used in a machine

learning model along with existing data to analyze the crime scene, determine number, type,

model of firearms used, type or amount of explosive material used, etc. Results of the analysis

may also be employed in updating existing information databases, enhancing accuracy of

machine learning models and generating various types of reports.

Data Maintenance

Data from the data input devices may be received by one or more servers of the

system directly or via one or more data collection applications executed on various computing

devices. For example, a crime scene technician may download captured audio or video data

from a data capture device to their tablet computer, enter notes about the crime scene and then

upload the entire information to the server. In another example, the one or more servers may

query data capture devices at the crime scene and download captured data automatically. 

Data Processing

The one or more servers may process the received or captured data using one or more

machine learning based approaches such as supervised learning systems, Bayesian networks,

neural networks, unsupervised learning systems and cluster analyses. The system may retrieve

information such as catalogued data (audio, video, etc.) associated with different firearms,

explosive devices and crime scene investigations. Comparing the catalogued data to the

received data and analyzing for similarities and distinctions, the system may determine a

number, a type and a model of the firearms used in the crime or a type and amount of

explosive material used in the crime. Catalogued information may be received or retrieved
from a number of individual sources or from a central database. Servers of the system and

data sources may communicate over public or private networks, may be cloud based or may

follow other topologic configurations.

Data Analysis

A machine learning model may receive captured input data and manual input data,

then evaluate the received data in light of catalogued data. The machine learning model may

be used to identify firearms and explosives. In addition, effects of the firearms or explosives

such as wounded or killed victims, property damage, etc. may also be analyzed by the model.

The machine learning model may also provide a predictive analysis, for example,

linking crime types to specific firearms models and wounds to specific firearms, allowing law

enforcement and first responders to adjust their strategies based on available information.

Furthermore, results of the analyses may be used to improve the machine learning process

through self-feedback and revisions or updates may be made to the cataloged data based on

the process. For example, new audio captures of firearm shots may be added to a fingerprint

database thus accuracy of existing records may be improved.

Data Communication

This includes reporting your data analysis results using the various data visualization

tools such as, summary tables, scatter plots, histograms, bar graphs, pie charts and box and

whisker plots. Data reporting may involve presenting your work to an audience, interested

stakeholders, peers in a lecture, seminar, workshop, conference or in a journal publication and

may also involve presentation of your date in a web-based applications such as a web page or

a website.

Decision Making of the Data

Forensic scientists are able to detect if a specific firearm fired a specific bullet casing

due to wear on the barrel that results in a unique deformation of each casing. The unique wear

may also lead to differences in the sound of a bullet discharge, which may be inaudible to a

human ear but detectable by a computer given sufficient audio quality. Such an individual

firearm audio fingerprint may be used to differentiate between and count various guns of the

same model used in a crime and to find the instances of previously recorded firearm crimes

using these firearms.

In an example scenario, a specific firearm may be determined from the audio

fingerprint. The owner's registration information may be looked up once the firearm is

determined so that the likely perpetrator can be found. As the likely perpetrator is determined

from the license information, the process of detecting the actual perpetrator through these

other data sources may be substantially simplified. The data associated with a previous owner

or user of the firearm may also be retrieved. Data about the user in this case may be used to

inform law enforcement response. Perpetrators who are inferred to have killed many people

may be met with an elevated response.