Applications in the medical field for the study of neuronal diseases such as Parkinson's or epilepsy. We use electroencephalographic tests to train classical (CNN) or graphical convolutional models (GCNN). In the first case, EEGs have been transformed into images that represent brain activity according to the different types of brain waves. In the second case EEGs are modeled as graphs so that the nodes represent the electrodes and the edges the connectivity between the different brain areas.
We also use PLN techniques (transformer architectures) to classify stages of Parkinson's disease (PD) progression from patient EEGs.
Other work in this field includes the development of semi-automated machine learning workflows to create models to measure the survival of a patient operated on for Abdominal Aortic Aneurysm (AAA) from clinical data or analysis of breast thermographic images to determine the feasibility of cost-effective breast cancer screening using thermography as a diagnostic technique.
Architecture & Cultural Heritage
Use of GANs for virtual inpainting restoration of artificial landscape images containing archaeological remains. The network identifies key features determined by the internal logic of the architectural style denoted by the ruins and adds the missing architectural elements to obtain an image of the restored building. Unlike other studies, it does not receive any information on which elements should be added or where. We use images of renderings of Greek temples with missing architectural elements that will be added to the image automatically. An improved version uses transformers to define buildings characteristics using natural language.
This approaches have multiple applications in the field of cultural heritage by using interactive apps like guides or to accelerate architectural rehabilitation processes.
The process has been improved with the development of a Text2Image system specific to the field of architecture and heritage that from a descriptive text generates an image with the architectural elements specific to the labeled style that can be used for the virtual reconstruction of images of ruined buildings.
In the audio field, we apply Autoencoders for the reconstruction of damaged voice transmissions. In this case, the model receives voice audios with interferences or cuts making them unintelligible and reconstructs them automatically. The reconstruction is applied to the wave and not separating the different sounds as it happens in most of the applications described up to now. This can be applied in low quality communications where the receiver would have difficulty understanding the message.
We also developed DL models to remove background noise and clean the foreground human voice (denoising) in a process that includes first the classification of the background noise and then its removal, using CNN+LSTM models (classification) and Variational Autoencoders (cleaning), working on spectrograms.
Implementation of a hybrid method to generate new sentences that augment a training dataset for NLP applications. It combines Markov chains and word embeddings to produce high quality data similar to a starting dataset. We have also worked whit more complex models like Transformers to summarize court decisions or the automatic generation of news entries.
A Spanish-Latin translator is also being developed using Transformer models over-trained with a dataset obtained from the works of St. Augustine.
The emergence of attention models has opened up the possibility of applying NLP to areas other than language itself. We also use Transformers (and other derived models developed by us) to analyze biomedical signals (EEGs) and for image generation (Text2Image) in the field of architecture and national historical heritage.
Using data from the EU RASFF portal, we create predictive models using multilayer perceptrons and 1D convolutional networks that can predict the type of product most likely to be contaminated at a given time, as well as the country of origin and the most likely contaminant. These models will allow the optimisation of resources when analysing products at the border, as well as which specific analysis to perform.
By applying SNA (social network analysis) techniques, we carry out quantitative and structural analyses of food import/export routes to/from the EU.