Scholarship for the National Doctorate in Artificial Intelligence (Cycle XXXVII - a.y. 2021-2022)

Location: ILC-CNR
Title: "Deep Reading"
Contact Person: Vito Pirrelli
Deadline: 23/07/2021
Research Theme
Knowing how to read and understand a text requires the orchestration of a great variety of competing linguistic-cognitive processes. Over the past three decades, research on reading has analyzed these processes from various points of view. Functionally, the (orthotactic, morpho- phonological, semantic-lexical, syntactic and pragmatic) skills required by reading and understanding a text were highlighted. From an algorithmic point of view, the mechanisms underlying these processes and their degree of interaction have been described and analyzed. Finally, neuro-psychological and neuro-functional research has identified the cortical areas and sub-cortical structures that are activated in response to specific aspects of reading. Despite the considerable progress made, however, we are still far from fully understanding how these levels interface with each other.
The recurrent stochastic neural networks distributed on numerous levels of hidden units (also known as networks for "deep" learning) represent not only (in their supervised variant) an extraordinary tool for the development of advanced applications, but also a formidable accelerator of unsupervised neuro-computational research, and a solid test bed for verifying "top-down" cognitive hypotheses on the dynamics of the interaction between linguistic-cognitive levels. At the same time, the intrinsically complex nature of the reading task, which integrates heterogeneous representations in terms of both modalities and levels of structural nesting, offers "bottom-up" neuro-computational research the opportunity to fully understand the specific characteristics of "acquired" and generalized knowledge from deep learning architectures.
The proposed topic is able to achieve the following research objectives:
  • to create a bridge between connectionist emergentism and other machine learning models (such as, for instance, Bayesian models);
  • to shed light on the dynamic interaction between the processes involved in reading, on the relationship between these, cortical areas and sub-cortical structures of the human brain, and on the reasons behind suboptimal reading skills;
  • to understand in depth the processes of hierarchical abstraction involved in the functioning of networks for deep learning, in view of the development of increasingly advanced and efficient neuro-computational architectures;
  • to personalize the teaching of reading, through a constant monitoring activity of the development of reading skills in children, starting from the first years of primary school.