Page 18 - Security Today, March 2019

Supervised and Unsupervised Machine Learning
The two most prevalent types of machine learning are Unsupervised and Supervised: Unsupervised machine learning, also called Data Mining, tackles very narrow problems by analyzing unstructured data—data that has not been organized or labeled in advance—in order to find patterns. With unsupervised machine learning, the com- puter is looking for discernable patterns in the data and searching for an unknown output or “ground truth.” One of the main focuses in unsupervised machine learning is anomaly detection. In this case, the computer identifies points in a dataset or stream that are outside the normal range without this range being pre-defined.
In supervised machine learning, computers are “trained” to prop- erly classify inputs. This training occurs by providing the computer with structured datasets—data that has been organized or labeled in a predefined manner—that correlate thousands of possible inputs with corresponding labels that the computer understands.
The computer learns these correlations, through training, in or- der to be able to apply its understanding to new inputs. Once the computer has ingested and classified a new input, programmers must evaluate the “truthfulness” or accuracy of the output that the com- puter generates.
The programmer must tell the computer how accurate its clas- sification is in order to train the computer to improve its ability to recognize new inputs. For example, if you label and input millions of images of roses and petunias into the computer with their associ- ated labels, through supervised machine learning, the computer will ideally be able to differentiate between future images of roses and petunias at a tolerable rate.
Deep Learning and Working with Structured and Unstructured Data
One of the sub-disciplines under AI includes research in neural networks. Working with structured data, this research analyzes the relationship between inputs and outputs to gain new insights. Deep learning, also known as deep neural networks, is a specific formula- tion of neural networks that also works with structured data. What is exciting about deep learning is that the accuracies gained lately have often even exceeded what humans can do with specific tasks.
What’s Achievable Today with Deep Learning
In the physical security industry, we are achieving increased accuracy using deep learning to solve structured problems—problems that in- volve knowing what the output of the data should generally be. For example, automatic license plate recognition (ALPR) is a structured problem because, when we train our algorithms, we work with a data set of raw ALPR images, including letters, numbers, and symbols, to arrive at a classified output. In this case, the output is an image of license plate XYZ123.
At Genetec, we are actively using deep learning for purpose-built solutions that rely on identifying trends and dependencies between features present in the data itself. We are currently using deep learn- ing in AutoVu, our ALPR system, to increase the accuracy and ve- racity rates of license plate tag reads. By applying computer vision algorithms, we have greatly reduced false positive reads for law en- forcement officers when they identify and stop a vehicle of interest. Similarly, KiwiVision Privacy Protector has also been working with deep learning to improve the accuracy of its anonymization tool.
16
0319 | SECURITY TODAY
DEEP LEARNING