Machine Learning for Real-Time Applications
The current state-of the-art of machine learning adopts the use of batch learning where the data is available offline, mostly divided into test and training sets. The model is trained for a data set with specific characteristics, any changes on the data characteristics requires re-training and re-modelling; the learning process is not incremental.
The learning environment for streaming applications is totally different; the data is in real time, it is not possible to buffer it, also the algorithms cannot look into the data ahead of time. Moreover, the statistical characteristics of the data is continuously changing at high velocity. This challenging property is known as the concept drift .
This leads to the fact that streaming applications require a different learning paradigms that respects the limitations of the data characteristics.
Online learning is a machine learning paradigm designed to learn from data in real time through the use of Hierarchical Temporal Memory algorithm [2-5]. HTM mimics the learning in the brain and is capable to do high order predictions and it adopts Hebbian like learning. HTM network consists of a number of neurons connected to two different zones; the proximal representing the input and distal representing the lateral connections within the layer. HTM has been applied to predict data in real time for different real world applications the past few years [2,6].
On the other hand, many of the well-established machine learning paradigms either statistical or neural network NN models has recently urged to apply modifications on their algorithms to adapt to the challenges of learning in real time [2,7].
The aim of this thesis is to do a thorough assessment on the machine learning paradigms capable to learn and model data in real time. The assessment should consider both the classical machine learning algorithms modified for real-time data and the online learning paradigm. The comparison should be performed on different data types to evaluate the prediction efficiency.
- Insight into advanced techniques of machine learning
- Working on a real world application
- Collaboration with researchers in the topic of machine learning
- Possibility to implement and research a novel methods
- Good programming skills (Python and tensorflow).
- Good background on Mathematics and signal processing.
- Michael Riegler
- Prof. Ilangko Balasingham
- Dr. Noha El-Ganainy
- “Machine learning with big data: challenges and approaches’ , A. L’Heureux, K. Grolinger, H. El-Yamany, and M. Capretz, IEEE Access, Vol.5, pp.7776-7797, April 2017. Doi: 10.1109/ACCESS.2017.2696365.
- “Unsupervised Real-Time Anomaly Detection for Streaming Data”, S. Ahmed, A. Lavin, S. Purdy, and Z. Agha, Neurocomputing, Vo.262, pp. 134-147, November 2017. Doi: 10.1016/j.neucom.2017.04.070
- “Continuous Online Sequence Learning with an Unsupervised Neural Network Model”, Y. Cui, S. Ahmad, and J. Hawkins, Neural Computation, Vol.28, Issue.11, pp. 2474-2504, November 2016. Doi: 10.1162/NECO_a_00893
- ‘‘Why neurons have thousands of synapses, a theory of sequence memory in neocortex,’’ J. Hawkins and S. Ahmad, Frontiers Neural Circuits, vol. 10, no. 23, pp. 1–13, Mar. 2016. Doi: 10.3389/fncir.2016.00023.
- ‘‘Properties of sparse distributed representations and their application to hierarchical temporal memory.’’ S. Ahmad and J. Hawkins. (2015). [Online]. arxiv.org/abs/1503.07469
- “ A Distributed Anomaly Detection System for In-Vehicle Network Using HTM” C. Wang, Z. Zhao, L. Gong, L. Zhu, Z. Liu, and X. Cheng, IEEE Access, Vol.6, pp. 9091-9098, March 2018. Doi: 10.1109/ACCESS.2018.2799210.
- “An Evaluation of HTM and LSTM for Short-Term Arterial Traffic Flow Prediction.” J. Mackenzie, J. F. Roddick, and R. Zito, IEEE Transactions on Intelligent Transportation Systems, pp. 1-11, August 2018. Doi: 10.1109/TITS.2018.2843349