Introduction and Minutes

[namiyousef]

Hi guys,

When committing, please use the correct issue number b adding #i where i is the relevant issue number. For example, if I wanted to refer to the issue pertaining to Useful Links, I would say #1.

Please make sure you 'watch' the repository to receive emails / updates when changes are made. You can do this by clicking the eye icon, at the top of the repo.

Also feel free to edit / update / add issues as you see fit.

Project directory and structure:

• FILL THIS, MAKE SURE REQS MEET THAT OF YIPENG
• ADD NOTES ABOUT BRANCHES

[namiyousef]

Meeting date: 17.12.2021 Attendees: (add attendees) Decisions:

• Agreed to use Zoom for meetings, weekly meetings at 5 PM London local time. Meeting link
• Agreed to use Telegram for communication
• Agreed to use Overleaf for report writing (NEED TO ADD LINK) High level decisions to make next week:
• [x] Choose MTL formulation based and justify it
• [ ] Split project tasks and create milestones

[namiyousef]

Meeting date: 20.12.2021 Attendees: (add attendees) Discussion:

• Transfer learning has simple implementations
• Simultaneous has difficult implementations but many papers Decisions:

[yakovsushenok]

Meeting date: 28.12.2021 Discussion:

• Jacob and Connor arrived to the same architecture design for a MTL with segmentation and binary classification and we decided that we will proceed with Connor's MTL as it will be way more convenient for further tasks we need to do.
• By next meeting we would like to expand (and maybe finalize) the MTL network and start writing up our analysis.
• In the meantime people can start researching/reading papers about the OEQ (I will text in the GC to clarify who can do that) because we may want to split up our work on a group who is doing research on OEQ and the code.
• Our next meeting is 03.01.2022 4 pm GMT

[namiyousef]

Meeting date: 07.01.2022

Attendees: @yakovsushenok @namiyousef @NicolasBelissent @ValerieF412 @Aaron-S98 @ConnorWatts @MingqinChen @Sherlyn-Zhang

Discussion:

• MTL takes a long time to run, has some problems with RAM. Mostly ready for testing
• MTL does not perform well (negative flow)
• OEQ idea 1: perform a study that compares using different weights on the losses. That is, for the best performing MTL configuration (e.g. all 3 tasks, seg + bin, seg + bounding box), experiment with different weights on the loss functions to see which one performs best.
• OEQ idea 2: perform a study to see how adding randomly induced weights based on a prior (see this paper) affects generalisation
• OEQ idea 1 is simpler, but must be carefully written about in the report. We cannot say "For finding the best performing MTL, we tried different weights" and then for our OEQ we say "so we wanted to try some more weights". It must be explicit that: for the MTL config, we tried different models, and then in the OEQ we are taking a good model, and seeing how adding different weights to it's losses changes it's behaviour
• OEQ idea 2 is more interesting, but a clear hypothesis is harder to come by, since you are comparing two different models as opposed to experimenting with the efficacy of one.

  Decisions:

• Voted to go for OEQ 1
• Added deadlines to milestones
• Assigned tasks for report writing / testing

[namiyousef]

Meeting date: 06.01.2022

Attendees: @ValerieF412 @namiyousef

Discussion:

• OEQ idea 1: cross-stitching and decoder focussed models (PAD-Net). Agreed that such models would fall under the 70% part of the project, and NOT the OEQ. Decided to drop.
• OEQ idea 2: contrastive learning and attention models. Agreed that such models would fall under the 70% part of the project, and NOT the OEQ. Decided to drop.
• OEQ idea 3: find the optimal weights associated with each auxiliary task by adding a hidden layer. This formulation changes the learning process, so it is not actually correct. As confirmed by Yipeng.
• OEQ idea 4: Experiment how manually added weights (or through cross validation) affect the performance of the MTL tasks. As per Yipeng, though this is simple it is valid!
• OEQ idea 5: Experiment how different priors affect this model

  Decisions:

• share ideas with group
• In particular, discuss QEQ idea 5 and see if a reasonably clear hypothesis can be formulated