Ä¢¹½Ö±²¥

Opponent Docent Tuomo Kauranne (University of Eastern Finland) and Custos Docent Ilkka Pölönen (Ä¢¹½Ö±²¥). The doctoral dissertation is held in English.

The audience can follow the dissertation online.
Link to the Zoom Webinar (Zoom application or Google Chrome web browser recommended):

Phone number to which the audience can present possible additional questions at the end of the event (to the custos): +358 400 248140

Neurocomputing is one of the oldest and perhaps well-studied area in artificial intelligence, which has also spawned other research areas, like computational architectures and (distributed) parallel computation approaches.

The field of computer vision was conceived from the desire of understanding how humans and animals perceive their environments. The main focus was on how neural processes classify and process information.

Neural networks (a subfield of neurocomputing) are currently popular models. This is evident, for example, when considering AlphaGo, GPT-3 and AlphaFold models, which are mainly neural network models. A great deal of promise is around neural networks because of their versatile application domains. However, neural networks suffers a great drawback of needing large, including very diverse, amount of training data. Additionally, large computing resources are needed as the neural network models grow larger. It is good thing to remember from history, that neural networks were the cause of the first, so-called, artificial intelligence winter, due to their hype and overpromising capabilities.

Probabilistic graphical models are at the foundations of neurocomputing, and are the focus of this research work. From an optimization stand-point, probabilistic graphical models are more difficult to solve. However, from an "architectural" stand-point, graphical models are more versatile compared to neural networks. This research work is divided into two parts. The first part deals with how to improve the computation of graphical models from an optimization point of view, and how to accelerate the computation using elementary high-performance computing approaches. The second part deals with how to combine convolutional networks with graphical models to estimate physical properties of the human skin. This combined approach can be used, for example, to estimate physical property dependencies in skin cancer analysis.