RAJASEKAR VENKATESAN
I am a Data Scientist working at Singapore Airlines. My research expertise is in the area of time-series prediction, sequence modelling, natural language processing, graph analytics and recommendation systems integrating neural networks, deep learning and representation learning techniques. I completed my PhD in School of EEE, Nanyang Technological University, Singapore specialising in Machine Learning. My current work primarily revolves around developing hybrid models combining static data in combination with time-varying data using MLP and LSTM concurrently and are trained to complement each other. During my PhD, my research work is on Human-inspired Progressive Learning Technique (PLT) for Classification Problems during which I developed and introduced PLT for machine learning classification problems. It is capable of dynamically learning multiple new classes on the run without requiring any retraining. I've also developed Label Independent (aka) Universal Classifier capable of performing binary, multi-class and multi-label classification problems. More information about my codes are available in 'Code' page.
RESEARCH HIGHLIGHTS
Progressive Learning - Multi-label Classifier
Development of online multi-label classifier integrated with progressive learning technique
Progressive Learning - Universal Classifier
Integration of progressive learning technique (PLT) with universal classifier
PROGRESSIVE LEARNING TECHNIQUE
This newly developed learning technique is independent of the number of class constraints and it can learn new classes while still retaining the knowledge of previous classes. Whenever a new class (non-native to the knowledge learnt thus far) is encountered, the neural network structure gets remodeled automatically by facilitating new neurons and interconnections, and the parameters are calculated in such a way that it retains the knowledge learnt thus far. This technique is suitable for real-world applications where the number of classes is often unknown and online learning from real-time data is required.
The key objective of the PLT is that it can dynamically learn new classes on the run. Suppose the network is initially trained to classify ‘m’ number of classes. Consider the network encounters ‘c’ number of new classes which are alien to the previously learnt class, the PLT will adapt automatically and starts to learn the new class by maintaining the knowledge of previously learnt classes.
The introduction of new class(es) to the network, results in changes in the dimension of the output vector and the output weight matrix. Also the newly formed matrices with increased dimension should be evaluated in such a way that it still retains the knowledge learnt thus far and also facilitates the learning of the newly introduced class(es). The proposed algorithm can not only learn sequential introduction of single new class, but also simultaneous and sequential introduction of multiple new classes. The proposed algorithm is also independent of the time of introduction of the new class(es).
UNIVERSAL CLASSIFIER
Classification involves the learning of the mapping function that associates input samples to corresponding target label. There are two major categories of classification problems: Single-label classification and Multi-label classification. Traditional binary and multi-class classifications are sub-categories of single-label classification. Several classifiers are developed for binary, multi-class and multi-label classification problems, but there are no classifiers available in the literature capable of performing all three types of classification. I've developed a novel online universal classifier capable of performing all the aforementioned types of classification problems. Being a high speed online classifier, the proposed technique can be applied to streaming data applications.
The developed classifier will need to identify both the classification type and the target labels associated with the input samples. This results in increased complexity in achieving universal classification technique. There are three key challenges to be addressed to achieve universal classifier.
1. Identification of classification type
2. Estimating the number of target labels corresponding to each input sample
3. Identifying each of the associated target labels.
The proposed approach falls under the category of an algorithm adaptation method in which the base algorithm is extended to adapt to the requirements of the universal classification.