News

2026

2025

2024

2023

2022

Contact

e-mail: (public key) 
        GPG
        pub   rsa4096 2022-09-18 [SC]
              EA8670F4553BA6CD1C8F87A9BC5EC2F8BCE9C8A5
        uid           [ultimate] Andrii Krutsylo
        sub   rsa4096 2022-09-18 [E]
                
    GPG
    pub   rsa4096 2022-09-18 [SC]
        EA8670F4553BA6CD1C8F87A9BC5EC2F8BCE9C8A5
    uid           [ultimate] Andrii Krutsylo
    sub   rsa4096 2022-09-18 [E]

Security

The authenticity of any (or at least main) page on this website can be verified using GPG:

        # import my public key
        gpg --keyserver keyserver.ubuntu.com --recv-keys 910e309c330be661
        # download path_to_page to verify
        wget -q https://krutsylo.neocities.org/index.html
        # download signature path_to_page.asc
        wget -q https://krutsylo.neocities.org/index.html.asc
        # verify
        gpg --verify index.html.asc index.html
with expected output:
        gpg: Good signature from "Andrii Krutsylo"
                
     # import my public key
     gpg --keyserver keyserver.ubuntu.com --recv-keys 910e309c330be661
     # download path_to_page to verify
     wget -q https://krutsylo.neocities.org/index.html
     # download signature path_to_page.asc
     wget -q https://krutsylo.neocities.org/index.html.asc
     # verify
     gpg --verify index.html.asc index.html
with expected output:
     gpg: Good signature from "Andrii Krutsylo"
I am using freeTSA.org time stamp authority.

License

All texts, images and animations on website (toplevel directory) for which a separate license is not specified are licensed under a Creative Commons Attribution-NonCommercial-ShareAlike 4.0 International License.
All source code on website (toplevel directory) is supplied as is, without any warranty and is licensed under The GNU General Public License v3.0 unless another license is specified.

Teaching

Publications

Selected Papers

Scalable Forward-Forward Algorithm ⭐
arXiv 2025
Andrii Krutsylo
Extends Hinton's Forward-Forward Algorithm — a backpropagation-free optimization method — to deep CNNs. Addresses scalability bottlenecks in the original formulation and demonstrates that the approach can match or approach standard backprop performance.
PDF
Continually Learn to Map Visual Concepts to Large Language Models in Resource-Constrained Environments ⭐
Neurocomputing 2025
Clea Rebillard, Julio Hurtado, Andrii Krutsylo, Lucia Passaro, Vincenzo Lomonaco
Addresses the challenge of grounding evolving visual concepts into large language models under tight memory and compute budgets. Introduces a continual learning framework that incrementally aligns visual encoders with LLMs without full retraining, making vision-language adaptation feasible on edge and resource-limited devices.
PDF Short
Batch Sampling for Experience Replay ⭐
CODS-COMAD 2024 (Best Student Paper Award)
Andrii Krutsylo
Investigates how the composition of training batches drawn from the replay memory affects continual learning performance. Demonstrates that deliberate batch sampling strategies — controlling the diversity and balance of replayed examples — can substantially reduce catastrophic forgetting without increasing memory overhead.
PDF Poster

Journal Articles

Continually Learn to Map Visual Concepts to Large Language Models in Resource-Constrained Environments
Neurocomputing 2025
Clea Rebillard, Julio Hurtado, Andrii Krutsylo, Lucia Passaro, Vincenzo Lomonaco
Addresses the challenge of grounding evolving visual concepts into large language models under tight memory and compute budgets. Introduces a continual learning framework that incrementally aligns visual encoders with LLMs without full retraining, making vision-language adaptation feasible on edge and resource-limited devices.
PDF Short

Conference Papers

Hyperspherical Classifier Heads for Continual Learning and Out-of-Distribution Detection
Activity and Behavior Computing 2026
Andrii Krutsylo, Sungho Suh
Proposes replacing standard linear classifier heads with hyperspherical geometry-based alternatives to simultaneously improve resistance to catastrophic forgetting in continual learning and strengthen out-of-distribution detection — two key challenges typically addressed in isolation.
Merging Versus Separating Replay Samples in Continual Learning
ICANN 2025
Andrii Krutsylo
Studies a fundamental but underexplored design choice in experience replay: whether replayed samples should be merged into the same mini-batches as new data or kept in separate batches. The paper provides an empirical and theoretical analysis of how this decision affects gradient dynamics and forgetting.
PDF
Batch Sampling for Experience Replay
CODS-COMAD 2024 (Best Student Paper Award)
Andrii Krutsylo
Investigates how the composition of training batches drawn from the replay memory affects continual learning performance. Demonstrates that deliberate batch sampling strategies — controlling the diversity and balance of replayed examples — can substantially reduce catastrophic forgetting without increasing memory overhead.
PDF Poster
Hebbian Continual Representation Learning
HICSS 2023
Paweł Morawiecki, Andrii Krutsylo, Maciej Wołczyk, Marek Śmieja
Explores biologically-inspired Hebbian learning rules as a mechanism for building representations that naturally resist catastrophic forgetting. By leveraging local, unsupervised synaptic updates, the method learns transferable features across tasks without storing or replaying past data.
PDF Short
Diverse Memory for Experience Replay in Continual Learning
ESANN 2022 (spotlight)
Andrii Krutsylo, Paweł Morawiecki
Argues that the diversity of examples stored in the replay buffer is as important as its size. Proposes selection strategies that maximize feature-space coverage of the memory, leading to more representative replay and improved retention of past knowledge across sequential tasks.
PDF Poster

Short Papers and Abstracts

The Inter-batch Diversity of Samples in Experience Replay for Continual Learning
AAAI 2024
Andrii Krutsylo
Examines how variation in replayed samples across training iterations — rather than within a single batch — influences the stability-plasticity trade-off. Finds that higher inter-batch diversity correlates with better generalization and reduced forgetting in experience replay settings.
PDF
Evaluating Knowledge Retention in Continual Learning
SAC 2024
Andrii Krutsylo
Critically examines the metrics used to measure how much a model retains from previous tasks after learning new ones. Proposes a more nuanced evaluation protocol that disentangles forgetting from interference, offering a clearer picture of a model's true knowledge retention over time.
PDF Poster
Remember More by Recalling Less: Investigating the Role of Batch Size in Continual Learning with Experience Replay (Student Abstract)
AAAI 2021
Maciej Wołczyk, Andrii Krutsylo
An early investigation revealing a counterintuitive finding: using smaller replay batch sizes can actually improve long-term knowledge retention. This challenges common assumptions about experience replay and motivates a deeper study of how batch size interacts with the learning dynamics of continual models.
PDF

Preprints

Scalable Forward-Forward Algorithm
arXiv 2025
Andrii Krutsylo
Extends Hinton's Forward-Forward Algorithm — a backpropagation-free optimization method — to deep CNNs. Addresses scalability bottlenecks in the original formulation and demonstrates that the approach can match or approach standard backprop performance.
PDF
Continual Learning via Ensemble-Based Depth-Wise Masked Autoencoders for Data Quality Monitoring in High-Energy Physics
CERN CMS Note 2026
D. Julson, E.A. Reinhardt, Andrii Krutsylo, R. Sohal, G. Fidalgo, S. Gleyzer, E. Usai
Applies continual learning to the data quality monitoring at CERN's CMS detector. An ensemble of depth-wise masked autoencoders is trained incrementally to detect anomalies and data degradation across evolving experimental conditions, without retraining from scratch.
PDF