Commit 2d4cc39c authored by Per Rutquist's avatar Per Rutquist Committed by overleaf

Update on Overleaf.

parent c5a900aa
......@@ -247,8 +247,10 @@ In the machine learning community, there is a recognised need for more theoretic
Many algorithms, that - at the core - rely on optimization techniques, have been successfully applied to solve various difficult real-world problems. However, the question of which specific optimization technique to choose to solve a given task often is both extremely important and difficult to answer, which is owed to the fact that often, without prior information about the problem, there is no specific optimization technique that will clearly outperform others on the given task \footnote{This is often referred to as the \textit{No free lunch theorem} in search and optimization \TODO[Torsten]{cite}.}. For example, the performance of a deep learning classifier often crucially depends on the choice of the stochastic optimizer during training and there is no clear way to choose the underlying nonlinear optimizer for a model predictive controller. Hence, many optimization-based techniques that are supposedly intelligent, e.g. in machine learning and optimization-based control, require extensive manual expert knowledge to choose the specific optimization algorithm and -parameters. \newline
In Meta Learning \TODO[Hutter, Bödecker]{Distinguish Meta learning from Meta optimization here?}, we aim to use the information that we have, namely task-specific data, to automatically choose among, or even \textit{learn} novel optimization techniques that can solve a given task without expert supervision. Even though there already exist data-driven approaches that can choose among a pre-specified set of optimization algorithms \TODO[Torsten]{cite}, and even design new optimizers \TODO[Torsten]{cite}, they are often computationally demanding and require vast amounts of data. \newline
We will investigate data-driven methods for improving efficient learning and optimisation approaches and generating them from scratch. We will, e.g., draw on methods from reinforcement learning, optimisation, and genetic algorithms to generate neural networks that control the neural architectures and hyperparameters of various learning and optimisation methods. We will investigate data-efficient problems such as gradient-based and gradient-free optimisers as well as data inefficient problems such as reinforcement learning.
The Machine learning group led by Prof. Hutter has carried out extensive prior research on this topic \TODO[Hutter]{List references} and organizes workshops in conferences on a regular basis \footnote{ E.g. \TODO[Hutter]{List Franks workshops on AutoML/Meta Learning}}.
\TODO[Hutter, Bödecker]{Formulate the phd topics in written paragraphs (see 2.1)}
\begin{itemize}
\item Learning Better Gradient-Based Optimisers
\item Learning Better Gradient-Free Optimisers
......@@ -258,16 +260,12 @@ We will investigate data-driven methods for improving efficient learning and opt
\end{itemize}
\textit{People: Hutter, Bödecker, Lütkebohmert, Diehl?}
\textbf{Further titles, to be modified:}
\begin{itemize}
\item Uncertainty and Risk
\item (Future / Model) Prediction
\item Time Series and Sequential Learning (Continual Learning)
\item Data and Computational Efficiency
\end{itemize}
\subsection{ Uncertainty and Risk } \TODO{Decide on title}
\subsection{ (Future / Model) Prediction } \TODO{Decide on title}
\subsection{ Time Series and Sequential Learning (Continual ) Learning) } \TODO{Decide on title}
\subsection{ Data and Computational Efficiency } \TODO{Decide on title}
\subsection{Applications}
\paragraph{Robotics} \TODO[Diehl, Bödecker]{write 2-3 sentences on an application, possibly also mentioning autonomous driving.}
......@@ -308,8 +306,7 @@ Finally the yearly gatherings will provide room for discussion between the LECO
%Process to ensure progress of work
Each PhD student will be assigned an advisory board consisting of two PIs from the LECO consortium plus a mentor. The mentor will be a either a Post-Doc from the consortium or an affiliated partner. The PhD student and advisory board will meet twice a year (at least once per year in person) to discuss progress since the last meeting, developments in the field, and ideas for future research. Based on the discussion results the PhD student will write a yearly progress report, that is shared with the entire consortium.
All LECO members will have access to the lectures given by other LECO members. \TODO[Tobi]{add selection of classes already offered}
There already exists a number of experimental setups and systems that are used for both research and teaching purposes.
All LECO members will have access to the lectures, seminars, and lab courses given by other LECO members. These classes include (but are not limited to): Mathematics of deep learning, Reinforcement Learning, Automated Machine Learning, Deep Learning, Modelling and System Identification, Statistics and Machine Learning, Numerical Optimal Control, Numerical Optimisation. The lectures are constantly updated to represent the state-of-the-art, new classes are added as required.
\priority[2]{ %
Existing and planned (*) graduate level lectures and seminars relevant for LECO doctoral students include: %
......@@ -363,18 +360,15 @@ Existing and planned (*) graduate level lectures and seminars relevant for LECO
\item Robotics
\end{enumerate}}
In addition to these lectures and applications number of soft-skill courses will be offered to LECO members:
\begin{itemize}
\item Good scientific practice
\item Scientific writing (paper and application)
\item Presentation of scientific results
\item Ethical aspects of LECO / (Artificial Intelligence, Machine Learning, and Control)
\end{itemize}
In addition to these lectures and applications number of soft-skill courses will be offered to LECO members. Topics of these courses include: Good scientific practice, Scientific writing (paper and application), Presentation of scientific results, Ethical aspects of LECO / (Artificial Intelligence, Machine Learning, and Control)
There already exists a number of experimental setups and systems that are used for both research and teaching purposes. Members of the LECO consortium will have access to these setups for experimental evaluation of developed methods and receive the best possible support by the respective group.
% Ethical aspects & Gender equality!
The LECO consortium strives to attract good students and researches regardless of their gender and ethnicity. To ensure gender equality and to prevent racism or discrimination of any form, there will be two elected confidants (at least one female) that members can turn two when required.
The LECO consortium strives to attract good students and researches regardless of their gender and ethnicity. To ensure gender equality and to prevent racism or discrimination of any form, there will be two elected confidants (at least one female) that members can turn two when required. These confidants report to the LECO chair and all work together to reduce discrimination and individual hardship to an absolute minimum.
To nurse the collaboration between LECO members from different disciplines there will be a co-working space that all members will have access to and are encouraged to use.
To aid the collaboration between LECO members from different disciplines there will be a co-working space that all members will have access to and are encouraged to use.
Joint lunch, coffee breaks, \ldots
......
Markdown is supported
0% or
You are about to add 0 people to the discussion. Proceed with caution.
Finish editing this message first!
Please register or to comment