Machine Learning Meetups

Date: Friday, 9 January 2026
Time: 11:00–12:00 (EET)
Format: Online (Zoom)
Location: https://aalto.zoom.us/j/62824564124

Speaker: Prof. Lasse Leskelä
Title: Consistent Spectral Clustering in Sparse Tensor Block Models

Abstract:
High-order clustering aims to classify objects in multiway datasets arising in fields such as bioinformatics, social network analysis, and recommendation systems. These datasets are often sparse and high-dimensional, posing significant statistical and computational challenges.

In this talk, we introduce a tensor block model for sparse integer-valued data tensors and propose a simple spectral clustering algorithm augmented with a trimming step to mitigate noise fluctuations. We identify a density threshold guaranteeing consistency and model sparsity using a sub-Poisson concentration framework that accommodates heavier-than-sub-Gaussian tails. Notably, this class of tensor block models is closed under aggregation across arbitrary modes, enabling a principled analysis of the tradeoff between signal loss and noise reduction. The theoretical results are illustrated through simulation experiments.

The talk is based on joint work with Ian Välimaa: https://arxiv.org/abs/2501.13820

*************************

Date: Friday, 23 January 2026
Time: 11:00–12:00 (EET)
Format: Online (Zoom)
Location: https://aalto.zoom.us/j/62824564124

Speaker: Dr. Matilde Tristany Farinha
Title: Bio-Plausible Learning for Artificial Neural Networks

Abstract:
This talk explores biologically plausible learning algorithms for artificial neural networks, motivated by the slowing of Moore’s Law and the emergence of analogue neuromorphic hardware for edge AI. We introduce Deep Feedback Control (DFC), a bio-inspired learning framework that enables learning to be fully local in time and space while supporting effective credit assignment across diverse feedback connectivity patterns.

We discuss how DFC can strongly influence neural activity, increasing biological plausibility compared to backpropagation, and how differential Hebbian learning can be integrated into feedback-based deep learning frameworks without requiring neurons to compare compartmental signals. Finally, we show how these bio-plausible learning algorithms can enhance robustness against adversarial attacks.

Speaker Bio: 
Matilde Tristany Farinha is currently a Data Scientist at UBS. She holds a PhD in Computer Science from ETH Zurich, where her research focused on biologically plausible learning algorithms for artificial neural networks. She also holds a BSc and MSc in Mathematics.

Fun facts: 
* She enjoys spending her free time outdoors, especially in the mountains.
* Her Paul Erdős number is 4.
* She once considered attending art school—and still enjoys painting as a hobby.

For more information, please contact professor Alexander Jung.