The boom of artificial intelligence is strongly connected to the success of deep learning models in recognizing patterns from data sources such as images, video, text or audio.
This Emblematic Project has the objective of developing evolved AI techniques, advancing from systems based on memorizing patterns to one based on the semantic comprehension of these patterns and the learning of abstractions or procedures that promote high level reasoning. Advances in this field will allow for the creation of better AI systems for the generalization and transfer of knowledge, other tasks.
Why does this constitute a challenge? Because an important disadvantage of current AI memorization-based systems is their black-box kind of operation; this means that, although the AI results are effective, it is highly complex if not impossible to understand the processes and motivations behind every inference.
Understanding the processes used by these models in their operation and the types of information they generate is a key concern in the research agenda of this area. Its study has practical and ethical implications, related to the profiling and the bias of AI results, even though the root of the problem is a more fundamental concern.
Our work aims to develop new techniques that allow for the guiding of the AI systems’ training process so they are capable of integrating meta-learning processes: that is, so they can learn procedures instead of memorizing patterns.
This will have a significant impact on the understanding we have of how these models work, but it will mainly affect the future capability to learn from few examples, thus improving their generalization abilities and, therefore, increasing their performance.
-Detection of fake news in social networks such as Twitter.
-Detection of bots in social networks.
-Detection of harassment (bullying) in social networks.
Felipe del Rio