Summary
Memory and learning are human central cognitive abilities. The importance of understanding human memory functioning is evident from its central role in our cognitive health as well as its role as the main inspiration behind developments in artificial intelligence, in particular artificial deep neural networks (DNN). Despite considerable progress in the recent years in the area of DNNs, robustness of these networks is an important open issue. In particular, noise robustness, i.e., DNNs are fragile in maintaining the correct predictions if their input is perturbed. In contrast, a healthy human’s memory system maintains performance despite perturbed inputs. This motivates us to learn from the biological neuronal networks of human memory for a more robust DNN. The human memory is composed of several modules responsible for processing, learning, and recalling the received information. Among the memory modules is the working memory (WM) which is responsible for holding and processing information in a temporary fashion and in service of higher order cognitive tasks, e.g. decision making. The short-term nature of the WM makes it a great example for designing dynamic DNNs, which are useful in safety critical applications in uncertain environments. The aim of this proposal is to build a combined model-based and data-driven mathematical framework for understanding Recall dynamics of human Working Memory Networks (ReWoMeN) for realization of a robust DNN as well as contributing to the mechanistic understanding of the human WM. ReWoMeN address three main challenges including derivation of a biologically plausible system-level model to account for the measured data of human experience of WM recalling, analysis of such a complex model for explaining and predicting WM behavior, and comparing the robustness of our WM model with a recurrent DNN in an image recognition application.
Unfold all
/
Fold all
More information & hyperlinks
| Web resources: | https://cordis.europa.eu/project/id/101030415 |
| Start date: | 01-06-2021 |
| End date: | 31-05-2023 |
| Total budget - Public funding: | 175 572,48 Euro - 175 572,00 Euro |
Cordis data
Original description
Memory and learning are human central cognitive abilities. The importance of understanding human memory functioning is evident from its central role in our cognitive health as well as its role as the main inspiration behind developments in artificial intelligence, in particular artificial deep neural networks (DNN). Despite considerable progress in the recent years in the area of DNNs, robustness of these networks is an important open issue. In particular, noise robustness, i.e., DNNs are fragile in maintaining the correct predictions if their input is perturbed. In contrast, a healthy human’s memory system maintains performance despite perturbed inputs. This motivates us to learn from the biological neuronal networks of human memory for a more robust DNN. The human memory is composed of several modules responsible for processing, learning, and recalling the received information. Among the memory modules is the working memory (WM) which is responsible for holding and processing information in a temporary fashion and in service of higher order cognitive tasks, e.g. decision making. The short-term nature of the WM makes it a great example for designing dynamic DNNs, which are useful in safety critical applications in uncertain environments. The aim of this proposal is to build a combined model-based and data-driven mathematical framework for understanding Recall dynamics of human Working Memory Networks (ReWoMeN) for realization of a robust DNN as well as contributing to the mechanistic understanding of the human WM. ReWoMeN address three main challenges including derivation of a biologically plausible system-level model to account for the measured data of human experience of WM recalling, analysis of such a complex model for explaining and predicting WM behavior, and comparing the robustness of our WM model with a recurrent DNN in an image recognition application.Status
CLOSEDCall topic
MSCA-IF-2020Update Date
28-04-2024
Geographical location(s)
Structured mapping
