Please enable JavaScript.
Coggle requires JavaScript to display documents.
Potential research topics - Coggle Diagram
Potential research topics
Continual Learning
Integration of new samples without
training from the very beginning again
Extension of new Tasks to be
trained in succession
CL/LLL: Tassilo
Robustness
Lifelong Learning + Robustness
Cohort from one Dataset +
new Data coming form other Center?
Datasets
Multiorgan-CT Dataset
Selber Task unterschiedliche
Scanner/Zentren
Issues
Catastrophic Forgetting
Incremental Learning
Self-Sup. Pretraining + Transfer in low label regime has overlap with lifelong+robustness
How to define benchmark in medical setting? (e.g. in natural image processing ImageNetC)
Michael
High Imbalance
of Pathologies
Training with some classes that are highly
imbalanced w.r.t. some other class (Long Tail Distribution)
Few-Shot
Approaches
Oversampling of the
MIchael
Simulation to Reality
wie GTA V in autonomous driving?
Check if the framework is usable
How to paste into the Images?
Probably messy simulations and hard to get out when starting collaboration
Transfer Learning
Self-supervised Pre-training
lotteries generalize in medical as well?
Logical Reasoning
a lá Dreamcoder
Finding logical programs that
allow one to find basic structural paradigms
Dreamcoder works on already known rules
Hard to do in images?
Create basic "Building blocks"
maybe some Gabor filters or take some already trained models?
How to get "easy" and known programs from some known basic examples?
Since we are looking for some reasoning
maybe create a class-like hierarchical structure?
Simulated Datasets?
Object Based reasoning
CLEVR: A Diagnostic Dataset for
Compositional Language and Elementary Visual Reasoning
Extract some "radiomics" features of the objects
and start random forest
Meta-learning
Bio-bank
Open-World Object Detection (
https://arxiv.org/abs/2103.02603
)
Medical Setting? Construct from densely annotated data?
RACOON? (when?)
Gib gib
Task 17 stlye? (many labels in Abdominal CT)
Michael & Gregor
Human Machine differences in Medical Imaging (Bethge Lab style?)
How understandable are current
medical ML methods?
Pendant to:
http://arxiv.org/abs/2106.12447
Look for tasks in which unskilled Mechanical Turk guys could perform equally well
Has to be based on 2D slices?
Some classification
between Better/Worse
(Maybe with regression for a follow-up paper?)
Unskilled worker vs Expert vs ML methods