Ponedeljkov seminar računalništva in informatike - Arhiv

V ponedeljek, 1. septembra 2025, bosta ob 16:00 uri izvedeni dve
predavanji v okviru PONEDELJKOVEGA SEMINARJA RAČUNALNIŠTVA IN INFORMATIKE
Oddelkov za Informacijske znanosti in tehnologije UP FAMNIT in UP IAM.

ČAS/PROSTOR: 1. september 2025 ob 16.00 prek Zoom-a (https://upr-si.zoom.us/j/297328207?pwd=S3Zpdk1VR3pjckNtWkQwKzlvcDR5UT09).

1. predavanje:
============

------------------------------------------------
PREDAVATELJ: Milan MILIVOJČEVIĆ
------------------------------------------------

Milan Milivojčević is a PhD student in D(LT)^2 lab at UP FAMNIT under the supervision of Asst. Prof. Aleksandar Tošić and Assoc. Prof. Jernej Vičič. His research focuses on using digital twins of smart floors to create synthetic datasets for human activity modeling.

----------------------------------------------------------------------------------------------
NASLOV: Synthetic Data Generation Using a Smart Floor Digital Twin
----------------------------------------------------------------------------------------------

POVZETEK:

The creation of reliable fall detection systems requires access to large quantities of labeled sensor data. Collecting such data in real world settings, especially for critical events such as falls, poses practical, ethical, and safety challenges. In this work, we propose a digital twin of a pressure-sensor-based floor system designed to simulate realistic human-floor interactions in a virtual environment. Our virtual replica mirrors the physical floor structure, composed of tightly networked sensor tiles, enabling the simulation of various fall scenarios using rigid-body ragdoll models. Synthetic sensor data is generated through virtual collisions, providing time-series pressure patterns analogous to real-world recordings. This method enables scalable data generation for training machine learning models focused on binary fall detection, with future expansion to multi-class activity recognition. Our approach aims to bridge the gap between the physical constraints of data collection and the need for extensive training data-sets in ambient assisted living systems.

Seminar bo potekal v angleškem jeziku.

=============================================================================================================

2. predavanje:
============

-------------------------------------------------
PREDAVATELJICA: Kaja PRAPROTNIK
-------------------------------------------------

Kaja Praprotnik is a PhD student at UP FAMNIT under the supervision of Assoc. Prof. Jernej Vičič. Her research focuses on natural language processing with a strong focus on artificial intelligence.

-------------------------------------------------------------------------------------------------------------
NASLOV: Influencing Sentiment in LLM Output via Token Selection Manipulation
-------------------------------------------------------------------------------------------------------------

POVZETEK:

In this work we present an algorithm for steering LLM into generating more positive outcome. There are many papers exploring different techniques of approaching the sentiment steering problem in LLMs, but to our knowledge there are no current methods that would approach the problem by manipulating logits. 
Our research focuses on developing an algorithm, that changes the sentiment of the LLM's output by manipulating the most probable tokens. The output could be steered into more positive and into more negative direction.

Seminar bo potekal v angleškem jeziku.

=============================================================================================================

Seminarja bosta izvedena online prek Zoom-a na sledeči povezavi:

https://upr-si.zoom.us/j/297328207?pwd=S3Zpdk1VR3pjckNtWkQwKzlvcDR5UT09

Meeting ID: 297 328 207
Passcode: 123456789

Vabljeni!