Giovani in un'ora - Ciclo di seminari - Seconda parte

Day - Time: 09 October 2024, h.11:00

Place: Area della Ricerca CNR di Pisa - Room: C-29

Speakers

• Saira Bano
• Giovanni Mauro
• Giovanna Broccia

Referent

Fabio Carrara

Abstract

Abstract: Traditional knowledge distillation (KD) typically uses a large, complex teacher model, often trained to a single modality, to transfer knowledge to a smaller, more efficient student model. However, these methods overlook the explanatory knowledge embedded in the teacher's internal representations. This lack of transparency hinders insight into the decision-making process and raises concerns about confidence and interpretability, especially in cross-modality contexts where the teacher's and student's models handle different data modalities. To address this gap, we propose an innovative method for cross-modal knowledge distillation that focuses on improving model efficiency and interpretability. The core of our approach modifies the standard KD process by using a multimodal teacher ensemble, where each teacher is trained for a specific modality. The results of these teachers are then weighted using Shapley values to ensure that the most relevant features contribute to the final prediction. We compute these values by projecting both the teacher ensemble results and the student model inputs into a common latent space, allowing the student model to learn extensive multimodal representations while requiring only its own modality during inference. This method retains the predictive power of the multimodal ensemble while enhancing model efficiency, making it well-suited for real-world applications across various domains such as healthcare and wellbeing, as well as in scenarios that involve human in the loop, such as autonomous driving, online gaming, and remote piloting.

Giovanni Mauro - "Dynamic models of gentrification"

Abstract: The phenomenon of gentrificaiton of an urban area, characterized by the displacement of lower-income residents due to rising living costs and an influx of wealthier individuals, is influenced by factors like increased amenities and infrastructures. This study introduces an agent-based model that simulates gentrification through the relocation of agents belonging to three different income groups, whose movements are driven by living costs. The model incorporates key aspects of economic and sociological theories of gentrification to generate realistic patterns of neighborhood transition. To quantify gentrification, we use a temporal network-based approach to devise a measure that tracks the outflow of low-income and the simultaneous inflow of middle- and high-income residents over time. 

Numerical experiments show that the role of High-Income residents is key in order for gentrification events to occur. Furthermore, our network-based measure of gentrification is shown to consistently predict such events, making it a potential candidate as an early warning of neighbourhood gentrification in real-world scenarios. Finally, our analyses highlight how city density favours the arising of gentrification patterns.

Altogether, in this work we show how this framework provides useful tools for understanding gentrification dynamics and informing urban planning and policy decisions.

Abstract: The representation of requirements plays a critical role in the accuracy of requirements inspection. While visual representations, such as UML diagrams, are widely used to enhance comprehension and identification of issues in requirements their effectiveness is still debated. Cognitive abilities, such as working memory and mental rotation, may also influence inspection accuracy, particularly depending on how the requirements are represented.

This study aims to evaluate whether the use of UML sequence diagrams improves the accuracy of requirements inspection and how cognitive abilities interact with the type of representation (UML vs. text) to influence inspection performance.

We conducted a crossover experiment with 38 participants to assess the accuracy of finding issues under two treatments (text and UML support). A linear mixed-effects model was used to analyse treatment, period, sequence, and cognitive ability effects.

The results indicate no significant advantage of using UML diagrams over text-based requirements, with the treatment effect showing a slight, non-significant decrease in accuracy for UML. However, the analysis revealed a marginally significant interaction between UML treatment and mental rotation ability, suggesting that participants with higher mental rotation skills tended to perform better with UML representations. Working memory capacity was found to have a significant positive effect on accuracy, although no significant interaction was observed between working memory capacity and treatment type.

The findings suggest that while UML diagrams may not universally improve requirements inspection accuracy, cognitive abilities can influence inspection performance. These results highlight the importance of considering individual cognitive profiles when designing and assigning requirements inspection tasks.