Focusing on developing distributed algorithms which can be deployed across a network to solve the industry needs.
The AIML:SIG hosts an interactive meeting on the third Wednesday of each month from 10:00 – 10:30 in our Netlabs Boardroom. Members present an introduction to key AIML techniques and showcase cutting-edge applications in their sector. This highlights current AIML involvement in collaborations, projects and proposals and encourages future cross-collaboration among multiple disciplines.
The AIML team across TSSG take inspiration from the cognitive abilities of the human brain and designs algorithms, process and applications that models its behaviour. The AI team are building lightweight AI algorithms that can vary their complexity depending on the urgency of the decision to be made in a similar fashion as to how the human brain can concentrate on complex tasks in short bursts. These algorithms are at the fore of edge-based AI algorithm design where balancing device constraints against decision performance is key.
There are three thematic areas to the AIML:SIG and they include:
- Fog Data Analytics
- Brain activity Pattern Recognition
Over the last number of years, design of various IoT and other Edge devices has improved dramatically. The devices have computational capacity that is sufficient for on-board execution of a rather diverse functionality. In recognition of that, ICT has seen the emergence of the Fog Computing paradigm. Computational resources available at the network Edge are proposed to host a variety of originally cloud-only services. Data analysis is one of the examples of such services. Meanwhile, despite the recent growth computational capacity of edge devices still remains limited. This has to be accounted for by the design of the Data Analytics services. The services need to obtain intelligent insight in a way that poses no disruption to the core device functionality (e.g. sensing the environment). The Fog Data Analytics stream of the AIML SIG focuses on algorithmic processes and techniques for building Data Analytics services operating in highly constrained environments (e.g. on-board an IoT device).
Fig. 1. Fog Data Analytics combines (a) data fusion, (b) Edge mining, (c) bio-inspired analytics and (d) location and user awareness.
Electroencephalography (EEG) is a non-invasive technique for acquiring brain activity data where electrodes attached to the scalp record electrical currents produced by the brain with millisecond temporal resolution. Machine learning techniques such as LDAS (for ERP analysis), support vector machines (SVMs for derived graphs) enable real-time classification of complex patterns distributed across numerous EEG traces simultaneously. Using brain stimulation methods encoded in the AR/VR content to identify ERPs such as P300 responses. Using these auditory and visual stimulations we can identify key brain reactions to this content. These stimulations in turn affect the current brain activity and allow us to determine normal and abnormal reactions to the stimulations.
We investigate both the event and the resulting signals to determine its effects on overall brain activity. A person’s brain functional network can then be modelled with graph theory and artificial intelligence methods such as graph convolutional networks and spectral convolutional graph networks. The activity of the brain is studied by characterizing these EEG patterns and associating common network topologies with correct brain function. This in turn is used to train classifiers which improves the overall prediction model. One application is identification of neurological disorders such as Alzheimers, schizophrenia, depression and Parkinsons which exhibit different patterns and network topologies. This prediction model can be further trained based on the neurofeedback system supplying new stimulations to achieve the desired brain activity reaction and thus the approximated network topology. These stimulations devised by deep learning models are encoded as parts of dynamic content in the neurotherapy feedback to allow the users activity to directly affect the experience. As more subjective data is gathered these predictions gain accuracy. Based on the functionally healthy brain topologies, weightings can be determined based on the subjective reactions to the stimulations that allow a dynamic rehabilitation system to be developed.
With the rapid development of telecommunication, sensing, and computing techniques, traditional transport systems are becoming smarter, safer and more environmental-friendly, and evolving into Intelligent Transport Systems (ITS). Artificial Intelligence and Machine Learning dramatically contribute to this evolution for both individual vehicles and whole transport systems.
Nowadays, onboard sensors and communication devices, e.g., lidar, radar, camera, and vehicle-to-everything (V2X) communications modules can produce vast amounts of raw sensed data, and ML is applied to provide vehicles with more intelligence. An example here is the case of autonomous vehicles where deep learning techniques are used to help vehicles learn self-driving. At transport system level, vehicles, infrastructure, and humans in transport systems interact dynamically and with great complexity and produce big data which needs ML techniques to extract meaningful traffic features. Using data analysis and ML, better design, operation, and management of road networks can be achieved, particularly in the context of integrated systems such as smart cities.
The above illustrates the application of Machine Learning in ITS
Here in TSSG, the current research on AIML in transport lies mainly in two aspects: ITS data analysis (see TransSec project) and communication network management (see AMEND project). For the former, the TransSec team has developed a prototype of intelligent traffic junctions by using real-time traffic information and deep learning algorithms to predict future traffic flow. For the latter, the CONNECT team has developed models for allocating communication and other resources for distributed services having vehicular nodes for sensing and computation.