Meta-Knowledge Transfer @ ECML/PKDD

7 papers23 directions
Meta-Knowledge Transfer is a research workshop held in conjunction with the European Conference on Machine Learning and Principles and Practice of Knowledge Discovery in Databases (ECML/PKDD). The venue focuses on the advancement of meta-learning techniques, specifically addressing how knowledge acquired from previous tasks can be transferred to improve performance on new, unseen tasks. Its primary objective is to facilitate technical discussions on the mechanisms of knowledge transfer, model generalization, and the optimization of learning processes across diverse domains. Research presented at this venue covers a variety of technical directions within the meta-learning landscape. Key areas of focus include cross-domain few-shot learning, where models are trained to adapt to extreme domain shifts with minimal data, and the development of pretrained optimization models for zero-shot generalization. Other significant topics include neuro-symbolic integration for verifiable reasoning, active data selection for label-efficient learning, and online hyperparameter optimization for non-stationary environments. The venue also features work on performance estimation for neural architecture search (NAS) and the design of meta-learning competitions, such as the NeurIPS Cross-Domain MetaDL challenge. The workshop serves as a platform for researchers such as Henry Gouk, Felix Mohr, Jan N. van Rijn, and Pavel Brazdil to explore methods for improving the efficiency and robustness of machine learning systems. Contributions often involve autonomous tuning of exploration parameters, trust region meta-learning for policy optimization, and the creation of interpretable meta-scores for model performance. By addressing challenges in cross-task performance normalization and local Bayesian optimization, the venue contributes to the development of more adaptable AI models capable of operating in complex environments.

Papers by Year

2022 7

Top Research Directions

Research directions most represented by papers from this venue.

Accelerating Neural Architecture Search Without Full Model Training 2
Probabilistic Treatment Ranking with Cumulative Ranking Curves 1
Performance Saturation in Few-Shot Learning 1
Probabilistic Rank Aggregation from Pairwise Comparisons 1
Hierarchical Space Partitioning for Black Box Optimization 1
Why Hyperparameters Fail to Transfer Across Architectures 1
Feature Scaling for Stable Neural Network Training 1
Layer Wise Feature Over Specialization in Deep Networks 1
Zero-Shot Transfer of Meta-Learned Optimization Strategies 1
Quantifying Statistical Uncertainty in Model Rankings 1
The Representational Bottleneck of Frozen Foundation Models 1
Relative Performance Ranking for Efficient Model Selection 1
Simulating Distribution Shifts with Episodic Meta-Learning 1
Learned Performance Predictors for Automated Optimization 1
Mitigating Overfitting with Frozen Feature Extractors 1

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 this venue that shaped research clusters.

Influential

Papers

TitleYear
Advances in Metalearning: ECML/PKDD Workshop on Meta-Knowledge Transfer2022
Experiments in Cross-domain Few-shot Learning for Image Classification: Extended Abstract2022
Faster Performance Estimation for NAS with Embedding Proximity Score2022
Interpretable Meta-Score for Model Performance: Extended Abstract2022
NeurIPS’22 Cross-Domain MetaDL competition: Design and baseline results2022
Searching in the Forest for Local Bayesian Optimization2022
Trust Region Meta Learning for Policy Optimization2022