"I'm more of an expert in models than in AI."
Emilio Soria is an expert in artificial intelligence and one of the most knowledgeable people in Spain on the matter. He is a professor at the University of Valencia, founder of the research group IDAL (Intelligent Data Analysis Laboratory) of the School of Engineering of Valencia, director of the Master's in Inteligencia Artificial Avanzada y Aplicada: IA^3 (Advanced and Applied Artificial Intelligence: AI^3), and author of works such as "Inteligencia Artificial: casos prácticos con aprendizaje profundo (Artificial Intelligence: practical cases with deep learning)."
He studied physics because he liked "the relationship between mathematics and the real world." When he finished his studies, he remained linked to the academic field and as in the virtual world, something like in Matrix, "I was offered two books. Since I've always liked a challenge, I chose the one about adaptive filters. That decision set the course of my professional life."
Although he began his research on adaptive filters, a device that attempts to model the relationship between signals in real-time iteratively, he later focused on neural networks. The result of this work was his doctoral thesis related to problems applied to electrocardiography. And so, his academic life evolved: "We were applying more complicated models: fuzzy logic, neuro-fuzzy systems, etc., until where we are now." Thanks to all his work in this area, Emilio is recognized as more of an expert in models than an expert in artificial intelligence.
As the founder of the IDAL research group at the University of Valencia, he launched an appeal from Pangeanic's "Pangea AI & Languages" podcast: "Our doors are open for those who want to work with data and there is no shortage of projects."
IDAL's main goal is the study and application of intelligent methods of data analysis for pattern recognition, with applications in the fields of prediction, classification, or trend determination.
"IDAL is a research group in which we make all kinds of applications with all kinds of data." At IDAL "we use structured and unstructured data and, in general, large amounts of data. But it is true that we are starting to forget that with 100 or 200 data patterns you can develop a model that can be the solution to your problem."
Emilio Soria is also director of the Master's in Advanced and Applied Artificial Intelligence: AI^3 from the University of Valencia and we asked him what models are being applied in machine translation. He told us that "in the field of research we have been paralyzed since 2017 when a model related to the famous film the Transformers came out. It is a model that works well in all applications and that takes context into account."
For Emilio Soria there is an aspiration linked to this situation of stagnation in terms of research: "Someone must come along and raise their hand to offer something different to what we already have."
Spain has recently created an Agency for the Supervision of AI and is promoting a project in Europe in the field of testing artificial intelligence involving public and private entities for the regulation of abuse. For Emilio, it is clear that "Spain has become a world reference in terms of ethics and artificial intelligence. However, I think we have put the cart before the horse." In this regard, he says that the production size "is not big enough to require something like this when we do not yet have, for example, an AI application agency for language processing."
Among his lines of research, Emilio Soria is also focused on multimodal models in deep learning: "Maybe it's easier to understand with an example. For instance, in medicine, there are diagnostic systems based on images. If I have a particular patient's medical history and picture, why do I have to split the information? With the multimodal model, you attack the model with images, with text, with a graph. And then you can better decide what is best for a patient."
This kind of model even serves to recognize fraud in an accident: "The graphs can be used to see the relationship level of the people involved in the accident, providing more reliable information and preventing fraud."
The trends in NLP for Emilio Soria are focused on "the different advances of conversion from voice to text, for example, Whisper from OpenAI that allows you to translate any voice to text."
Whisper is an automatic speech recognition (ASR) system trained using 680,000 hours of multilingual, multitasking supervised data collected on the web. Although he also acknowledges the importance of "advances in cloned voice."
For him, "incorporating these key elements into machine translation will generate a number of applications that will have an important impact on short- and medium-term trends in natural language processing."