Challenge: Developing deep meta-learning algorithms to model and understand learning curve patterns in machine learning.
Impact: Faster, better, more cost-efficient training and tuning of learning
We are looking for a critical and open-minded person to come work with us to deepen our understanding of learning curves. Learning curves in machine learning plot performance versus training set size. By extrapolating learning curves we can predict how many training samples are necessary for a particular performance. Learning curves can also be used to speed up learning algorithms, model selection, and hyperparameter tuning. However, unexpected and strange learning curves make this task difficult. Furthermore, there is much uncertainty about the general shape of learning curves: are they exponential, power law, or do they have other predictable shapes? A better understanding of learning curves can provide deeper insights into how learning algorithms work, and may inspire new machine learning theory and improved learning algorithms.
You will analyze learning curves, develop deeper knowledge about their shape, and exploit that knowledge to improve applications that rely on learning curves (hyperparameter tuning, model selection, predicting the amount of data needed). The plan in this project is to develop meta-learning algorithms for learning about learning curves. By analyzing a database composed of a large number of learning curves, we want to extract data-driven insights and exploit them. We see especially potential for deep learning methods (meta-learning and generative modeling), but there is room to explore and develop other kinds of models. Moreover, we are interested in translating findings into human-understandable insights (explainable AI) which can inspire new theoretical insights and new algorithms.
In this project you will develop new learning algorithms to learn from learning curves, develop new insights into learning curves, develop benchmarks, design, run, and analyze large-scale experiments. There is room in this project for freedom and creativity. On one hand, the project offers empirical challenges but there is also room for theoretical work. This can be balanced and explored according to your interests.
Predicting the amount of data necessary for learning is relevant for real-life settings where data collection is expensive, difficult, or time-consuming so that data collection costs can be minimized. This issue is important for small companies that want to apply machine learning in a cost-efficient manner. Model selection and hyperparameter tuning are compute-intensive tasks. New and more efficient algorithms based on learning curves will lead to more sustainable machine learning.
Our research environment offers a dynamic, stimulating, and diverse atmosphere, providing you with opportunities to collaborate with experts in the field. You will work within the Pattern Recognition and Bioinformatics group within the Computer Science department Intelligent Systems, which includes researchers working on machine learning, pattern recognition, computer vision, and socially perceptive computing. Our research group is very international and socially active. You will be advised by Tom Viering (http://tomviering.nl/).
We are looking for a candidate that meets the following criteria:
- A Master's degree or equivalent, or about to graduate with one, in a relevant field (Machine Learning, Statistics, Mathematics, Artificial Intelligence, Computer Science, Physics, Engineering, etc.)
- Solid background in linear algebra, probability theory, and statistics.
- Ability to program in Python, C++ or other programming language
- Proficiency in spoken and written English
- Can work together well in a team
We encourage you to apply even if you feel you don't meet all the criteria above, as long as you are willing to acquire the complementary skills. Please keep an eye on my Twitter, the vacancy will be officially opened later this week (around 15th of September 2023)