Artificial intelligence research

Breakthrough research in image texture and video descriptors including local binary patterns (LBP), world-renowned facial image analysis, 3D vision and emotion AI are some of the cornerstones of the Core AI research at the University of Oulu. The Center for Machine Vision and Signal Analysis (CMVS) invents and develops novel approaches for computer vision, affective computing and machine learning for a wide range of problems and applications.

The Machine Vision research has resulted in state-of-the-art facial image analysis, for example in the field of facial expression and micro-expression recognition and detection of spoofing attacks in biometric systems. The group has also investigated novel methods for lip reading and measuring heart rate from facial videos.

Visual 3D perception is another aspect of the CMVS’ research, where the main area of application is augmented reality (AR), where 3D vision is a prerequisite for the alignment of virtual and real worlds. In the field of 3D vision, the Center has aimed at fusing conventional geometric approaches with modern learning-based techniques.

Emotion AI is where emotional intelligence converges with human computer interaction (HCI). Emotion AI is intended to make HCI more like actual human interaction by taking emotion analysis into account. The Center is a pioneer in facial micro-expression analysis for hidden emotion reading and among the first ones in dynamic facial expression recognition. The Center has proposed new methods to measure physiological signals from remotely captured videos–ie. to measure a person’s heart rate without any contact. This opens lots of potential not only in emotion AI but also in health-related applications.

Underlying the research is machine learning, the CMVS’ image texture and video descriptor research has been widely used as a basis in a large number of applications around the world, e.g. biometrics, medical image analysis, affective computing and industrial inspection. The goal is to invent computationally light but discriminative image and video descriptors that can be efficiently used in wearable devices or smart sensors. One of the focuses in methods for this is to reduce or remove entirely the limitations of current deep learning approaches, which require enormous quantities of well annotated data and high-performance computers, making many real world applications useless. Humans can learn many things just from a few examples without needing massive amounts of training data.

To break the current dominant model of dumb devices with always-on cloud-connectivity in the upcoming 5G mobile networks and future 6G technologies, EdgeAI is pulling computations in the cloud closer to the devices.

This will alleviate many issues connected to the future development on devices, networks and applications. Modern architectures are compute-, space-, and power-hungry. Edge computing promises to tackle these problems by bringing the computations closer to the devices, where the data is generated and applications are used. Current Artificial Intelligence (AI) and Machine Learning (ML) methods mostly assume that all computations are conducted in a homogenous cloud with ample resources available.

Edge computing needs to be brought closer together with AI/ML methods. Edge computing is by design distributed and leans towards decentralization; however, modern AI/ML methods are only just beginning to allow distributed and decentralized computations. This is where EdgeAI comes in.

Developed at the University of Oulu, EdgeAI is the interplay of AI and edge computing. The interplay is further divided into edge computing for AI (under the moniker Edge4AI) and AI for edge computing (AI4Edge).

• The ICON group at the University of Oulu investigates the theoretical and algorithmic underpinning of distributed low-latency, and reliable training, inference and control in dynamic environments, focusing on both machine learning for communication (ML4C) and communication for machine learning (C4ML) research. The overarching goal is to enable intelligent and autonomous devices and high-stake applications.
• NSOFT (Network security and softwarization) group, which belongs to CWC Networks and Systems research unit, studies and develops secure, scalable and resource-efficient techniques for 5G and beyond 5G systems.
• AI provides new capabilities for the control and management of edge-based computational platforms: for example, predictive or decentralized control and system orchestration as well as predictive maintenance for the network infrastructure. In turn, distributed edge platforms with IoT connectivity provide massive amounts of local data for AI models.
• EdgeAI will have a major impact on future security issues. AI improves Edge security with personalized security systems that are adapted to each individual location or user particulars. AI methods will improve user privacy by decentralizing trust, personal privacy guards and fine-grained control on consent management and data ownership. Privacy concerns, on the other hand, will enhance AI methods by providing novel privacy-preserving regularization and data-generating Machine Learning models.

To harness the full potential of the data that surrounds us, we need ways and means for collecting, pre-processing, handling and utilizing it. This is what researchers of Applied AI are developing.

The data is not valuable in itself. Rather, it has to be processed carefully in order to find the information that can then be used in smart advisory and decision support systems. Real-time applications and big data means that the process has to be automated by using intelligent methods that can adjust to seasonality or personal changes.

In the University of Oulu, we have focused on applications that are a part of the everyday life, such as health applications and activity recognition, traffic and environmental sensor networks and also industry applications in high quality manufacturing.

Applied AI researchers have used movement and bio-signal data collected by wearable sensors for wellness and medical applications. For instance, they have studied human activity recognition and early detection of migraine attacks. The collected data is always unique to a given person as everyone has their own style of moving, for instance. Therefore, they have developed a dynamic machine learning models for wearable sensor data, adapting to a user’s individual data and any changes in it, which might be caused by illness, for example. The models can also adapt to environmental changes like seasonal changes.

Buildings get smarter with services that are based on data streams from sensor networks. Environmental sensor networks provide huge amounts of data, which the tools of applied AI are capable of visualising and processing automatically in real-time. The data is used to provide useful information about the facilities to the users.

Applied AI is also used in various areas of the manufacturing industry. A company’s competitive edge depends on the guaranteed high quality of the product and ensures availability of the processes. Applied AI researchers have developed tools for online quality monitoring. The tools enable data-driven manufacturing, product design and digital twinning. Applied AI researchers have also developed process optimization tools and tools for smart decision support for timely reaction during manufacturing. The core of the tools is in machine learning based prediction models and they also build information visualization based on the user’s needs.

Characteristics of applied AI solutions are the dynamicity of the models and the capability for continuous learning. Powerful machine learning methods form the base for applied AI’s proactive systems. Instead of full black-box methods, the emphasis is on methods that provide information about the role of the variables in the model. This information can then be utilized in automated model interpretation that enable the development of AI-assisted decision support.

Unleashing the full potential of Artificial Intelligence in the field of medical sciences means more accurate diagnoses, faster, and with greater efficiency. The use of AI and advanced data analytics will greatly benefit many areas of medical research. Elderly care, mental healthcare, preventive dental care and telemedicine applications are all set to gain from the new methodology. But the benefits of developing new technology reach further than improving treatments for specific diseases and conditions, like Parkinson’s. Advances in AI mean also great strides for the healthcare industry in general.

The University of Oulu is doing state-of-the-art research in utilizing AI in analyzing knee X-rays in order to provide more systematic and structured reporting on osteoarthritis that is easy to interpret. The AI method is robust and already performs at a similar level or better than a single radiologist. For the patient, this has multiple benefits. Ultimately, the patient will receive a more accurate diagnosis much faster than what is possible at the moment. Not only will they be able to plan for recovery armed with better information, but the method allows the prediction of disease progression for patients. This is not possible with current clinical techniques. As the AI-based method allows the prediction of the progression of osteoarthritis, the population at large will benefit greatly. This will have a large impact on society in general and the pharmaceutical industry specifically.

Another example of the benefits of AI methodology in examining and interpreting X-rays is the detection of abnormalities in chest X-rays. Chest radiography is the most common X-ray study and therefore in a position to be greatly improved by the increased efficiency, accuracy and speed that AI-based methods provide.

The use of AI in the field of medicine requires large amounts of well-annotated data. With the collection and utilization of patient data to be used in the training of such AI-based methods, privacy concerns are at the very heart of developing AI methodology. Currently there is a newer generation of methodology that is being developed with these concerns in mind.

Air pollution and climate change are in the heart of research done by the ATMOS group at the University of Oulu. The atmosphere and its processes are studied with computational methods utilizing measurement data and theoretical knowledge of physics, chemistry and meteorology. We use Machine Learning methods to leverage and improve the performance of traditional physical models, to enable more accurate predictions, and to provide us with new insights to the large pool of digital data that is continuously produced with measurements around the world.

The atmosphere is not an isolated system, but in constant interaction with us humans. What we do in our everyday life has implications to the composition of air, to air quality. Air pollution, whether solid particles or gaseous compounds, is harmful to human health and affects atmospheric chemistry and climate. Human action, such as traffic and internet usage, produces emissions of greenhouse gases and other air pollution. As we aim to understand ways to reduce atmospheric pollutant emissions, we need to have ways to quantify people, their actions and underlying motivations.

Modelling people and their behaviour with approaches leaning on Artificial Intelligence enables us to see the impacts that emerge from a large number of individuals choosing to do this or that. Both humans and AI can learn and adapt, which is why AI is a perfect approach to investigate ways to nudge or incentivize individuals to change their behaviour towards more climate-friendly habits.