ELLIS MeetUp Nijmegen - September 2024 ( ELLIS Fellows Final Presentations )

Dear All,

We are excited to announce the next ELLIS MeetUp Nijmegen. In this talk, we have several short presentations by ELLIS Excellence Fellows at Nijmegen unit on how to push the boundaries in various domains in AI and Machine Learning.

Join ELLIS Unit LinkedIn group if you are not yet a member:
https://www.linkedin.com/company/ellis-unit-ijmegen/?viewAsMember=true

Meetings will be held monthly on the first Wednesday of each month throughout the academic year 2024/25.

This Month:
Time: 4 September 2024, 15:00-17:00,
Location: Erasmus Building, Room 18.02A/18.03, Erasmusplein 1, 6525 HT Nijmegen.

The meeting is in-person, however still the there is possibility to join online: Microsoft Teams Link

Title of talks and presenters:

• Title: FrODO: Fractional Order Distributed Optimization,

ELLIS Presenter: Andrei Lixandru

abstract: Distributed optimization (DO) is critical in machine learning, especially for federated learning and multi-agent coordination applications. DO involves networked agents, each with a local function, collaboratively optimizing a shared variable to minimize their collective functions. However, when these functions have ill-conditioned Hessians, DO struggles with optimal convergence rates. To address this, we introduce fractional calculus in DO to incorporate long-term memory, enhancing stability and convergence in scenarios with challenging objective functions. We assess the efficacy of this approach in several settings and compare it with other state-of-the-art methods.

• Title: State-space Wishart Processes for Multivariate Count Data Time Series Analysis,

ELLIS Presenter: Benedetta Felici

Abstract: Time series analysis of multivariate count data holds significance in diverse application areas. In this work, we present IntGWP, a state-space model tailored for fitting multivariate and integer valued data. Our scheme assumes Poisson distributed observations and allows for a dynamic model of intensity rates through a Wishart Process prior. We conduct simulations on both real-world and synthetic data, demonstrating competitive performance in terms of model fit and prediction. However, our experiments reveal some limitations: IntGWP does not provide an intuitive model of covariance, lacks the capability for multi-task learning, and its model fit does not significantly benefit from modeling observations with discrete random variables.

• Title: Event-based Near-eye Gaze Tracking using a Spiking Neural Network,
  ELLIS Presenter: Stijn Groenen

Abstract: We investigate Dynamic Vision Sensor (DVS) cameras and spiking neural networks (SNN) for the purpose of fast and efficient near-eye gaze tracking. The event data contains both spatial and temporal dimensions, so our current model uses a combination of convolutional and recurrent layers to predict the gaze vector of the user. We use adaptive leaky-integrate-and-fire neurons with liquid time constants to allow the spiking behaviour to adapt to the input. The model is trained on pre-existing datasets using a forward-propagation-through-time (FPTT) method, where we only backpropagate over the last few time steps, rather than the entire event stream. We are still working on improving our SNN, but our preliminary results show a mean error of ~7.7°, with a mean adaptable update rate of 191Hz.

• Title: TBA
  ELLIS Presenter: Pelle Kools