3IA PhD/Postdoc Seminar #50

Flash presentation - Temporal Action Segmentation for Robot-Assisted Surgery Analysis

Abstract: Temporal action segmentation in robot-assisted surgery videos involves automatically identifying and temporally localizing surgical actions throughout a procedure. This task is fundamental for developing intelligent surgical assistance systems that can understand surgical workflow in real-time. Our research focuses on analyzing endoscopic video recordings from robotic surgeries to recognize fine-grained surgical gestures and actions, such as suturing maneuvers, tissue manipulation, and instrument movements.

Contextualizing Toxicity: An Annotation Framework for Unveiling Pragmatics in Conversations of Online Discussion Forums

Abstract: The role of context has attracted increasing attention in research on toxicity detection. Interpreting toxic language remains a complex and multifaceted challenge, shaped by numerous linguistic, contextual, and social factors. However, current approaches often define “context” narrowly, focusing primarily on surface lexical cues such as hate lexicons, profanity markers, or sentiment polarity. These features, while useful, are insufficient to capture the interactional dynamics, user behaviors, and intentionality that shape such phenomena. To address this gap, this paper introduces a novel and systematic annotation framework, grounded in Speech Act Theory (Austin, 1962), aimed at deciphering the illocutionary and perlocutionary dimensions of conversation, which are unexplored in existing studies. We apply this framework to a new dataset of complete Reddit conversation threads, sampled to include discussions that turn toxic (124 conversations, 1990 messages). We evaluate the performance of GPT models (GPT-3, GPT-4, and GPT-5) on this challenging annotation task, providing insights into how large language models capture pragmatic and contextual dimensions of online toxicity.  

Importance weighted directed graph variational auto-encoder for block modelling of complex networks

Abstract: We addresses the fundamental challenges of jointly performing node  clustering and representation learning in directed and valued graphs, which need both  global and local network structures to be captured. While these two tasks are  highly interdependent, they are often treated separately in existing works. We  propose the deep zero-inflated latent position block model (Deep-ZLPBM) in the  context of directed and valued networks characterized by non-symmetric adjacency  matrices with positive integer entries. Our approach leverages a variational              autoencoder (VAE) framework, combining a directed graph neural network (DirGNN)  encoder designed to handle directed edges and a zero-inflated Poisson (ZIP) block  modelling decoder to model sparse, integer-weighted interactions. Recognizing  the limitations of the standard evidence lower bound (ELBO) in VAEs, we explore  the importance weighted ELBO (iw-ELBO), a tighter bound on the marginal  log-likelihood optimized via gradient ascent, to enhance inference. Extensive  experiments on synthetic datasets demonstrate that iw-ELBO optimization yields  significant performance gains. Moreover, our results validate that Deep-ZLPBM  effectively models complex network structures, providing interpretable partial  memberships and insightful visualizations for directed, valued graphs.  

Open discussion about all the contributions