ALETHEIA
In recent years, advances in deep learning have resulted in a quite remarkable increase in performance for several artificial intelligence tasks, such as natural language processing or computer vision. In the specific area of image understanding, in addition to the introduction of novel neural network architectures and training paradigms, deep learning has enormously benefited from exponential growth in the number of labelled samples used for learning. Consequently, more labelled data has allowed the design of deeper models with higher performances.
Unfortunately, labelled data can be very difficult to obtain when solving certain problems, such as anomaly detection. Anomalies, typically defined as statistically not-relevant objects or events detected in images, can be semantically very relevant in certain domains, such as video surveillance, medical imaging, social media, and industrial quality control, for example.
Considering the lack of training data for anomaly detection, one of the main goals of our project will not only be to build deeper neural models which can satisfactorily handle the huge visual variabilities found in training data; we also plan to discover those particular regions, objects and events of interest, not commonly found during training, but which deserve our attention for posterior analysis. The problem of dealing with a reduced number of labelled data (or none at all) is of great interest to the computer vision and machine learning community, more precisely to methodologies known as weakly supervised, semi-supervised, unsupervised, few-shot and zeros-shot learning the ones tackling this actively.
On the other hand, once neural network models are learnt using available labelled data, which is a process that for deeper architectures can take even weeks to be completed, it is quite challenging to update models with new training data, especially when those new samples clash what has been already learnt; in that case, networks should decide which knowledge has to be forgotten and which has to be reused. Therefore, it is not surprising that this training paradigm, called continual learning, has become one of the hot topics in current research.
This project, called ALETHEIA, will focus on dealing with the two aforementioned limitations of deep learning, and its outcomes can constitute a considerable breakthrough in terms of applicability, generalization, and reduction of the computational cost of existing learning paradigms. In essence, this project will concentrate on developing new algorithms for anomaly detection in images when the size of training data is small or even non-existing. Our methodology should also consider that such anomalies might change over time and thus the resulting models should efficiently adapt to different data, by automatically deciding either to discard (forget) old knowledge or to incorporate new one (learn). This balance between learning and forgetting is known as the stability-plasticity dilemma.
To demonstrate the suitability of the methods proposed during this project, specific image classification, object detection, and motion analysis tasks will be addressed in selected problems in which anomalies are of key importance, like medical imaging or video surveillance.
Project PID2020-120611RB-I00 funded by MCIN/ AEI /10.13039/501100011033
Photos & videos
Partners
SHARE THIS EVENT
