Mehmet Supervision Meeting - 8th December

Mehmet Supervision Meeting - 8th December === ###### tags: `Progression of the study` `MSc projects` `Projects` `ICPE2023` :::info - **Location:** Zoom - **Date:** 8th December 2022 - **Attendees:** - **Apologies:** None ::: # Updates since the last meeting. - I designed a script to test MLPs in various shapes and configurations. - As there is an infinite number of configurations of MPLs, I seek to find a proper configuration interval in terms of the number of hidden layers and the number of neurons in each hidden layer. - Until now, Raspberry Pi 4b is not able to run an MLP with 4 hidden layers and 4096 neurons in each. - 4 hidden layers with 1024 neurons in each works. - Continual learning is not stable for now. I have been following ContinualAI Community. The current techniques/strategies more focus on classification. Then I was thinking about how to achieve and found some ideas about the progression of the study. - It would be more efficient if we could not recommend a specific name. Instead, we could recommend a configuration. By doing this, we can provide contuniality in sense. - The study can be turned into a NAS-related topic. With the help of consumption results, we can offer architectural configurations such as the use of between X and Y number of convolutional layers and M and N number of fully connected layers with O optimizer etc is recommended. - If any time remains after the main chunk, I can put, somehow, continual learning for types of layers that emerge in future. - I could not find a more automated way for collecting consumption results. The people who work on a similar topic do the same way as I do. - This morning, the comments of reviewers of ICPE2023 are announced. - There 2 accepts, 1 weak accept and 1 reject. I assume that the paper is accepted. - Comments are confusing. Especially the second reviewer commented positively but rejected the paper. - The first one suggested reasonable corrections. The 3rd and the 4th reviewer made constructive criticisms, and these motivated me. - The rebuttal is optional, however, I would like to improve the paper by adherıng their feedback. # *"The Worry List"* - I do not think that the use of the SPEC dataset we do now is useful. Because the dataset contains brand names. I assume if we could only stick with hardware configuration we would recommend hardware configuration either. - For this scenario, I offer to change the previous study a bit by eliminating brand-related columns of the dataset. We already designed deep networks and we can train a new deep network. - As a result, we can recommend 'Use this power of hardware X and hardware Y for this number of FC layers and Conv layers' to users. - The time interval for rebutting is short. # Short-term plans for next steps - Rebutting the reviewer comments - After collecting some MLP results, I will start to work on them. - EDA, ?, ? # Deadlines on the horizon - Rebuttal: December 7-9, 2022 - It is planned to close the rebuttal window on Monday, December 12. # Meeting notes :::info **Please note:** We will populate this section collaboratively during the meeting. ::: - - - - - # Actions :::info All of our actions should have the following things: - An owner, the person leading on the activity. - Sufficient detail so we can remember what we are supposed to do. - A deadline for completion. ::: :::info These actions will appear in the next meeting notes either as an update since last time (if completed), on the worry list (if we have run into issues), or in the short-term plans (if we've had to push this item forward). ::: - [ ] - [ ] - [ ]