Meta-Knowledge Transfer @ ECML/PKDD 2022

7 papers
The 2022 edition of the Meta-Knowledge Transfer workshop at ECML/PKDD featured seven papers focusing on the advancement of metalearning and optimization techniques. A significant portion of the research addressed challenges in few-shot learning, particularly under conditions of extreme domain shift and cross-domain image classification. Other central themes included the development of robust performance estimation methods, such as interpretable meta-scores and embedding proximity scores for Neural Architecture Search (NAS). The workshop also highlighted progress in automated machine learning (AutoML) and optimization strategies. Key topics included online hyperparameter optimization for non-stationary environments, autonomous tuning of exploration parameters, and the use of pretrained models for zero-shot generalization. Additionally, researchers explored neuro-symbolic integration for verifiable reasoning and trust region methods for policy optimization. Notable contributors to this year’s proceedings included Henry Gouk, Felix Mohr, and Jan N. van Rijn, whose work collectively aimed to improve the efficiency and adaptability of machine learning systems across diverse tasks and environments.

Top Research Directions

The 23 research directions with Meta-Knowledge Transfer @ ECML/PKDD 2022 papers, ranked by paper count. Share is the fraction of the edition’s papers in the direction.

#DirectionPapersShare
1Accelerating Neural Architecture Search Without Full Model Training228.6%
2Bridging the Domain Shift in Few Shot Learning114.3%
3Dynamic Hyperparameter Optimization for Reinforcement Learning114.3%
4Feature Scaling for Stable Neural Network Training114.3%
5Hierarchical Space Partitioning for Black Box Optimization114.3%
6Integrating Neural Perception with Symbolic Reasoning114.3%
7Layer Wise Feature Over Specialization in Deep Networks114.3%
8Learned Performance Predictors for Automated Optimization114.3%
9Mitigating Overfitting with Frozen Feature Extractors114.3%
10Performance Saturation in Few-Shot Learning114.3%
11Probabilistic Rank Aggregation from Pairwise Comparisons114.3%
12Probabilistic Treatment Ranking with Cumulative Ranking Curves114.3%
13Quantifying Statistical Uncertainty in Model Rankings114.3%
14Reducing Annotation Costs through Active Sample Selection114.3%
15Relative Performance Ranking for Efficient Model Selection114.3%
16Scaling Gaussian Processes Beyond Cubic Complexity114.3%
17Simulating Distribution Shifts with Episodic Meta-Learning114.3%
18The Computational Bottleneck of Meta-Training114.3%
19The Natural to Medical Image Feature Mismatch114.3%
20The Representational Bottleneck of Frozen Foundation Models114.3%
21Unified Evaluation Frameworks for Reproducible Machine Learning114.3%
22Why Hyperparameters Fail to Transfer Across Architectures114.3%
23Zero-Shot Transfer of Meta-Learned Optimization Strategies114.3%

Download all directions as CSV

Institutions

Institutions credited by Meta-Knowledge Transfer @ ECML/PKDD 2022 papers: a paper credits every institution named in its affiliations, once. 21 institutions share 7 credited papers.

#InstitutionPapersShare
1University of Edinburgh 228.6%
204z08gp61 114.3%
3Centre National de la Recherche Scientifique 114.3%
4ChaLearn 114.3%
5Computer Vision Center 114.3%
6Eindhoven University of Technology 114.3%
7Institut national de recherche en sciences et technologies du numérique 114.3%
8Laboratoire Interdisciplinaire des Sciences du Numérique 114.3%
9Leibniz University Hannover 114.3%
10Leiden University 114.3%
11Mila - Quebec Artificial Intelligence Institute 114.3%
12Politecnico di Milano 114.3%
13Polytechnique Montréal 114.3%
14Tsinghua University 114.3%
15Universidad de La Sabana 114.3%
16Universidade do Porto 114.3%
17Universitat de Barcelona 114.3%
18University of Waikato 114.3%
19University of Warsaw 114.3%
20Université Paris-Saclay 114.3%
21Warsaw University of Technology 114.3%

Download all institutions as CSV

About this data

Papers. Meta-Knowledge Transfer @ ECML/PKDD 2022 lists 7 papers on Lacuna, duplicate versions counted once; rejected and withdrawn submissions are left out.

Directions. Directions are Lacuna’s clustering of the concept elements extracted from each paper, so a paper can belong to several and direction counts add up to more than the paper count. A direction’s share is its papers divided by the edition’s papers.

Institutions. A paper credits each institution named in its affiliations, resolved to a ROR record, once. 7 of 7 papers (100.0%) credit at least one institution and are the base for institution shares; the rest have no resolved affiliation yet.

Data. Counts come from Lacuna’s data build of 2026-11-19 and change as papers are added. Tables as CSV: directions · institutions. Cite as: Lacuna, “Meta-Knowledge Transfer @ ECML/PKDD 2022”, https://lacuna.tiptreesystems.com/venue/meta-knowledge-transfer-ecml-pkdd-xetk/2022.

Top Authors

AuthorPapers
Henry Gouk2
Alberto Maria Metelli1
Bernhard Pfahringer1
Chaoyu Guan1
Difan Deng1
Dustin Carrión-Ojeda1
Eibe Frank1
Hong Chen1
Joaquin Vanschoren1
Marcello Restelli1
Marius Lindauer1
Prabhant Singh1
Przemyslaw Biecek1
Sergio Escalera1
Wenwu Zhu1
Xin Wang1

Impactful Papers

Papers from Meta-Knowledge Transfer @ ECML/PKDD 2022 that shaped research clusters.

Influential

Papers

Title
Advances in Metalearning: ECML/PKDD Workshop on Meta-Knowledge Transfer
Experiments in Cross-domain Few-shot Learning for Image Classification: Extended Abstract
Faster Performance Estimation for NAS with Embedding Proximity Score
Interpretable Meta-Score for Model Performance: Extended Abstract
NeurIPS’22 Cross-Domain MetaDL competition: Design and baseline results
Searching in the Forest for Local Bayesian Optimization
Trust Region Meta Learning for Policy Optimization