The Ikerbasque researchers Fadi Dornaika and Ignacio Arganda of the Department of Computer Sciences and Artificial Intelligence at the UPV/EHU are proposing ways forward in computer vision and machine learning. Specifically, they have achieved very good outcomes by using semi-supervised techniques in the field of facial beauty prediction, and they have also explored which error functions help to train the networks better so as to improve the estimation of facial age.
Machines are being taught to understand what they see
The Computer Vision and Pattern Discovery group of the UPV/EHU-University of the Basque Country has provided ways forward in predicting facial beauty and estimating facial age
First publication date: 30/12/2019
Research in artificial intelligence covers a whole host of fields with the aim of achieving what could be the ideal intelligent machine that can perceive its environment and undertake actions that maximise its chances of success in some goal or task. The UPV/EHU’s Computer Vision and Pattern Discovery group is involved in computer-assisted vision and machine learning. “Basically, we use modern techniques in artificial intelligence to resolve a whole range of image problems, in all kinds of images: 2D, 3D, videos, etc.,” explained Ignacio Arganda, Ikerbasque researcher in the UPV/EHU’s Department of Computer Sciences and Artificial Intelligence.
The research group does in fact have expertise in a wide range of subjects, such as biomedical images (detection of cells, tissue, tumours, etc.), facial images (perception of beauty, age estimation) and street images (location of vehicles, pedestrians, etc.), which they tackle in collaboration with researchers from other organisations and research centres. “On the whole, these are machine learning techniques because we tend to base ourselves on a set of labelled data, images or videos (in which the location of objects or their category is known); we then use them to teach or train our statistical or artificial intelligence models to assign these same labels to examples they had not seen previously,” explained Arganda.
What does the network focus on to deduce a person’s age or give a beauty score?
In pieces of research each linked to facial images the researchers put forward improvements in connection with predicting beauty and estimating age. “In the research into beauty prediction, we endeavoured to replicate the beauty scores given in different databases using semi-supervised techniques (in which not all the images were labelled),” explained Dr Arganda. “To do this, we used networks in which different features employed in training the models used to predict beauty are extracted.” In this line, the team members showed that “semi-supervised learning, never before used in this type of problem, gives results that are as good as or even better than in supervised learning (in which all the images are labelled)”, he pointed out.
“Convolutional neural networks (CNNs) are used to estimate age: what you have is an initial image, then a series of filters gradually extract features that help to reach a final decision, in other words, a number, in this case age,” he added. In this line “we conducted an empirical study to see which error functions help to train the networks in this field best, because the error in the estimations can be minimized in various ways", explained the researcher. The experimental results obtained point to the way in which age estimation can be improved.
Deep neural networks are used for predictions and estimations of this kind: “Networks with numerous connections, numerous filters, millions of data, etc. But it is important to understand what the network is focussing on to predict a person’s age, or to make any other kind of prediction. Right now, there is another field of research in which we are immersed and which is known as explainable or interpretable artificial intelligence; it explores techniques designed to clarify what the network focusses its attention on,” explained Arganda.
At the same time, the researcher warned that “machine learning techniques affect our lives more than we think.<0} A huge quantity of data is being generated and high-level decisions are being made on the basis of these systems. It is very important to take the ethical factor into consideration. In fact, gigantic databases are used in machine learning to train the models, and all the biases contained in these data are replicated in the predictions and estimations made by the models and this could be really harmful. In this respect, there are open pieces of research that are exploring how to remove the different biases from the data”. In Arganda’s view, “we are finding ourselves at an exciting moment in this field”.
Fadi Dornaika, an Ikerbasque research professor in the UPV/EHU’s Department of Computer Sciences and Artificial Intelligence, leads the Computer Vision and Pattern Discovery research group and specialises in computer vision, image processing, pattern recognition and machine learning. Besides Dornaika and Arganda, the following people participated in the research: the group’s researchers Anne Elorza and Abdelmalik Moujahid, the cotutelle doctoral student Kunwei Wang, and S.E. Bekhouche of the University of (Finland) and the University of Djelfa (Algeria).
- Toward graph-based semi-supervised face beauty prediction Expert Systems With Applications (2019) DOI: 10.1016/j.eswa.2019.112990