Please enable JavaScript.
Coggle requires JavaScript to display documents.
ToDos - Coggle Diagram
ToDos
Knowledge Extension Project
Calibration
Test its working as intended
Throw it on the cluster
Add loading of calibration to evaluation
Test split in calibrated & uncalibrated
Robustness
Augmentation robustness check
Check the
Add Camelyon17 Dataset
Think about artificial dataset
Create various shortcuts
Design potential shortcut settings
ADP - Baseline
CIFAR10
CIFAR100
ResNet20
Run 1 - baseline
Run 2 - baseline
Run 3 - baseline
Run 1 - 2,0.5
Run 2 - 2,0.5
Run 3 - 2,0.5
ResNet20
Run 1 - baseline
Run 2 - baseline
Run 3 - baseline
Run 1 - 2,0.5
Run 2 - 2,0.5
Run 3 - 2,0.5
Run 1 - 2,0.0
Run 2 - 2,0.0
Run 3 - 2,0.0
ResNet32
Run 1 - baseline
Run 2 - baseline
Run 1 - 2,0.5
Run 2 - 2,0.5
Run 1 - 2,0.0
Run 2 - 2,0.0
Own reimplementation
CIFAR10
CIFAR100
ResNet20
3x baseline
3x 2, 0.5
3x 2,0.0
ResNet20
3x baseline
3x 2, 0.5
3x 2,0.0
ResNet32
3x baseline
3x 2, 0.5
3x 2,0.0
Reach out to people
Create short summarization video of what I am working on (2 Minutes Elevator pitch)
How much performance can
all correlated models reach that are
created by random chance?
Sort the (unregularized) models
that are trained through random chance and "greedily" ensemble them --> If it improves use it.
Is it worth it to create high-performing models and ensemble them or are lower performing models okay as well?
Highlight hypothesis "dominance"
Times Model A performs better in a Class
than Model B
Times Model B performs better in a Class than model A
Relation shows if it "focusses"
on a different hypothesis
Metric: Ensemble Efficiency
Ratio of Ensemble performance relative
to mean model performance
Dataset Extensions
INATURALIST-1K
Flowers 102
Food 101
Papers
Reading
Random New Paper
Clinical Benjamin Paper
Reiterate paper
Make sure key-points are worked out
Decide on Journal with ppl
Paper Reviewing
BVM Paper
Do Reformatting to meet guidelines
Create File to upload
Upload to server
Gitlab - Repo
Use gitlab token
Immatrikulation
Hand stuff to Office (for signature)
Forward it to the graduate office-.