Biologically-Inspired Machine Intelligence Technique for Activity Classification in Smart Home Environments

With the widespread adoption of Internet-connected devices and the prevalence of applications in the Internet of things (IoT), devices in smart homes can generate enormous amounts of data. There is a requirement for machine-learning techniques to learn from historical patterns and predict future activities. There is an increased interest in machine-learning techniques that can provide useful and interesting services in the smart home domain. The areas that machine-learning techniques can help advance are varied and ever-evolving. Predicting and classifying the Activities of Daily Living (ADLs) of inhabitants in a smart home environment are key modules to automate smart home devices. Some prominent examples include uses for entertainment, elderly care, healthcare, and security. The abilities of the machine-learning technique to find meaningful spatio-temporal relations of high-dimensional data and to learn from streaming datasets are important requirements.

Recently, the Hierarchical Temporal Memory (HTM) theory has been presented as a biologically-inspired machine-learning theory that attempts to mimic the neocortex, the front part of the human brain. The main features of the HTM theory include the ability to learn and predict temporal patterns. The HTM theory and its computational implementation, Cortical Learning Algorithms (CLA), present a potential alternative to traditional machine intelligence. This research aims to apply a new biologically inspired machine intelligence technique based on the HTM theory and its CLA implementation to classify ADLs of inhabitants by analysing data captured from different sensors in a smart home scenario.

This research started by reviewing existing research in classification and prediction of ADLs. A comprehensive evaluation of state-of-the-art machine learning techniques and their application in the context of smart homes has been carried out as a primary research. HTM theory and its implementation the CLA have been studied, and experiments were conducted to identify the weaknesses and limitations of applying the HTM theory and the CLA for activity classification in the smart home environment.

To test and evaluate the performance of a proposed machine intelligence technique, there is a need for a dataset that represents the ADLs in a smart home scenario. Due to the excessive cost of building real smart home datasets and the lack of real datasets from smart homes, to tackle this issue, as a secondary contribution to knowledge, this research used OpenSHS (Open Smart Home Simulator), an open-source, cross-platform 3D smart home simulator. In addition, forty-two ADL datasets, Simulated Activities of Daily Living Dataset (SIMADL) were extracted from the OpenSHS tool and made available publicly.

This research proposes multi-region CLA techniques to learn short- and long-term patterns. Two novel multi-region CLA techniques featuring a multiple spatial pooler and temporal memory regions that incorporate a hash encoder and a Multi-Layer Perceptron (MLP) classifier were proposed and applied. While, the hash encoder can deal with multi-dimensional datasets, because of the existing encoders of the standard NuPIC encoders are prepared to deal with a single column or a small number of columns and the MLP classifier was used rather than the classifiers used in the current implementation of CLA to produce meaningful predictions. Additionally, the two novel multi-region CLAs, Parallel Spatio-Temporal Memory Stream (CLA2) and Cascaded Temporal Memories Stream (CLA3) were developed to learn short- and long-term patterns from, streaming datasets. To remove the limitation of memory management of the original CLA that contains one spatial pooler and one temporal memory, the original CLA learns using one memory level, such model learns either short-term or long-term patterns, not both of them. A novel CLA2 was proposed and developed, to cope with learning both shortterm and long-term patterns. CLA2 can learn both short-term and long-term patterns in parallel. CLA2 includes two spatial poolers and two temporal memories to simulate short-term and long-term memories. The number of cells per column was decreased in the first region to learn short-term patterns. In the second region, the number of cells per column was increased to learn long-term patterns and the outputs of both regions were concatenated into one vector. The second proposed algorithm (CLA3) has three regions, one spatial pooler, and three cascaded temporal Memories (TM) regions, where the first region learns smaller features, and the second and third regions are more abstract in order to learn and recognise patterns and concatenates the outputs from all three regions into one vector.

An evaluation and comparison of the proposed algorithms against state-of-the-art supervised machine-learning techniques and the standard CLA for classification was conducted using both the simulated smart home SIMADL dataset, generated using the OpenSHS, and the ARAS dataset, that comprises data captured from real-world activities of residents residing in two houses, the obtained results for the real datasets offered less performance than the synthetic datasets. Because the inhabitants are asked to record their activities manually, it was prone to human errors. The real-world dataset ARAS House A is inconsistent because one of the inhabitants left the house for long period of time, which impacts the performance results. The results of the proposed algorithms for the classification of ADLs show that its performance are promising. For the CLA2, the average overall F-measure for all the synthetic datasets, SIMADL and the real-world datasets, ARAS is 84.88%, while the highest F-measure (86.87%) was achieved by the Convolutional Neural Network (CNN) model. The (CLA2) has achieved an F-measure of 92.63% for House B of the real-world ARAS dataset, which outperforms state-of-the-art classification models. For the CLA3, the average overall F-measure for all the synthetic and real datasets is 84.50% . The proposed algorithms improve the best performance of base-line (standard) CLAs by an average F-measure of 51.81% overall.