Machine Learning and Biometric Systems
Machine learning is the systematic study of scientific algorithms that provide the system with the ability to simulate human learning activities without being explicitly programmed. Machine learning also studies the biometric topographies to simulate an individual’s identification learning activities.
Machine learning has made the functioning of biometrics identification possible and has also made much advancement in biometric pattern recognition. Machine learning approaches are further divided into three types: Unsupervised Learning, Supervised Learning, and Reinforcement Learning. These approaches help in the identification, classification, clustering, dimensionality reduction and recognition tasks which are needed to develop biometric systems.
Now, we will learn about the three machine learning approaches more deeply:
Unsupervised Learning
What is unsupervised learning?
Unsupervised learning is a part of machine learning that studies from test data that has not been classified, labeled or characterized. Instead of responding to the feedback, unsupervised machine learning identifies the commonalities in the data, and it responds by the presence or absence of such unities in every new piece of information.
Working of Unsupervised Learning.
- Let us consider a system which receives some series of inputs unlike 𝑋1, 𝑋2, 𝑋3, …, Where 𝑋 𝑋 represents an input and the set 𝑋 = 𝑋𝑋 is called the sample set that corresponds to a common database or dataset. The input data can be anything including the total number of fingerprint edges, eyes distance, hand survey of biometric pattern recognition, color or image on the retina and vein graph representation, etc.
- In this technology, the computer or the machine receives an inputs 𝑋1, 𝑋2, 𝑋3, …, and based on that input it builds the representations that are used in decision making, efficiently communicating the results to another system, predicting future inputs, etc.
- The unsupervised learning is mainly focused on clustering and dimensionality reduction tasks. Numerous algorithms have been developed to achieve this goal, but conventional approaches help in decision making which are as follows:
- Expectation-maximization algorithm
- Hebbian Learning approaches
- Convolutional Neural Networks
- Gaussian Mixture Models
Biometrics and Unsupervised Learning
The unsupervised scientific algorithms are designed for biometric applications which are mainly focused on specific data protection by encrypting biometric information, biometric data extraction, feature level fusion, behavioral pattern detection among others. Besides, biometric systems which have been implemented by using unsupervised learning ensures better learning policies and registration, successively allowing better classification and exact proof localization of biometric features.
- Unsupervised Learning in finger vein pattern- It can be used for full automatic finger vein pattern extraction. This method is considered to be an excellent approach to achieve biometric pattern recognition. However, it only usually serves as a preliminary stage for data analysis, better learning policies definition, features fusion (clustering tasks), etc. It can be considered as an initial data issues dealing approach to improve classification labors. Unsupervised Learning can also be used for full automatic finger vein pattern extraction.
- Unsupervised Learning in Fingerprint recognition- The fingerprint recognition was improved with an iterative Expectation-Maximization algorithm for collusion strategy adaptation.
- Retinal Pattern Matching- Vlachos and Dermatas propose a novel unsupervised clustering algorithm named nearest neighbor clustering algorithm (NNCA), which has been used successfully for retinal vessel segmentation.
- Voice Detection- Scientist Bahari had proposed the Voice activity detection method. He introduced a distributed energy signal unmixing method to locate the nodes around each source for wireless acoustic sensor networks (WASNs). Unsupervised learning algorithms unlike K-Means, K-medians, and K-medoids algorithms were used for voice activity source detection. Later on, the clustering algorithm has been used to extract the biometric voice features from the obtained energy signals.
Supervised Learning
What is supervised Learning?
Supervised learning is a modern technology of learning a function that infers a role from labeled training data which maps an input to an output based on example input-output pairs. In supervised learning, the data consists of a set of training examples where each example is a pair which includes an input vector and the desired output value or the supervisory signal. Supervised learning algorithm studies the training data and produces an inferred function, which is used for mapping new examples.
Working of Supervised Learning
- Let us consider a system which receives some series of inputs unlike 𝑋1, 𝑋2, 𝑋3, One can differentiate the supervised learning from unsupervised learning, as in supervised learning, the desired sequence of outputs unlike 𝑋1, 𝑋2, 𝑋3, …, is also given and the target of the system is to learn to construct the correct output from the given new input.
- According to the study of a different survey of biometric pattern recognition, it states that unlike unsupervised learning, the supervised learning also serves primarily in the final stages of the identification which is based on the biometrics. The unsupervised algorithms are used for discovering latent factors, discovering clusters, discovering matrix completion, graph structure, while supervised learning is focused more on the classification and regression of biometric identification.
- Supervised learning has proved itself useful for many biometric modalities fusion, biometric data classification and regression for reliable, prosperous and secure multi-biometric systems. With the help of supervised learning, many successful results have been obtained in the field of biometric identification and recognition. Recently a method of with accuracy up to 97.35% was developed by using a useful deep neural net (DNN) architecture and learning method that leverage an extensive labeled dataset of faces to obtain a face representation that generalizes well to other datasets.
Biometric and Supervised Learning
Supervised learning has been serving for numerous biometric applications by using a large number of algorithms. In contradistinction to unsupervised learning, which only uses mainly K-means algorithm for biometric applications, supervised education offers a variety of approaches for biometric pattern classification principally. Few algorithms of supervised learning are given below:
- Convolutional Neural Nets (CNN)
- Kernel Methods (SVM, Kernel Perceptron)
- Decision Trees
- Logistic Regression
Face Recognition- The ‘Decision Trees’ algorithm of Supervised Learning is applied for the exact face recognition Biometrics. According to the latest survey, this algorithm results shows a maximum accuracy of 100% on the FERET dataset and 99% on the CAS-PEALR1 dataset.
Speech Emotion Classification- For independent speaker verification, the ‘Support Vector Machines (SVM)’ algorithm is used. The baseline accuracy for speech emotion recognition was around 50% to 90% depending on the selected technique.
Facial Emotion Recognition- The ‘Kernel Perceptron’ Learning approach is used for Facial emotion biometric recognition. The classifier recognizes the 6 different Emotions with 98.6% efficiency on the JAFFE dataset.
Reinforcement Learning
Reinforcement learning is a type of dynamic machine learning programming which systematically learns to perform a new task and trains the algorithms using a system of reward and punishment. This learning technology is concerned with the software agents who take necessary actions in a real-time environment to maximize some notion of accumulative reward.
The reinforcement learning algorithm is built on the same concept that a child uses to learn a new task. Similarly, this algorithm learns by interacting with its environment. The software agent automatically receives rewards by performing correctly and penalties for performing inaccurately. The agent is programmed to determine without the intervention of a human only by maximizing its compensation and minimizing its penalty.
Working of Reinforcement Learning
- In reinforcement learning, the machine interacts with its environment by producing actions such as 𝑋1, 𝑋2, 𝑋3, …., an. These actions affect the state of the situation, and results in the machine receiving some scalar rewards 𝑋1, 𝑋2, …,rn or punishments p1, p2, p3…..,pn.
- As a learning problem, it refers to learning to control a system to maximize some numerical value, which represents a long-term objective.
- Reinforcement learning is based on the notion that if an action is followed by improvement, then the tendency to produce that action is strengthened else that action is not taken into practice.
Reinforcement Learning and Biometrics
Reinforcement learning seems to be more versatile than supervised and unsupervised learning. It is useful for both unsupervised labors and supervised labors. However, reinforcement learning is limited to reasonably low dimensional problems. But Deep Reinforcement Learning (DRL) has proven to be useful to solve this problem. Despite the successes of DRL, many issues need to be addressed before these techniques are applied to a broad range of complicated real-world issues.