Subwork_Aicha

• 12/4

Deepened my knowledge about algorithm design more spceifically FedAvg. I looked for papers about this algorithm and chose to start with this one: https://arxiv.org/pdf/1602.05629.pdf

Currently trying to understand the math behind this algorithm by reading lectures about it. For example: https://www.stat.cmu.edu/~ryantibs/convexopt-F18/scribes/Lecture_9.pdf; https://openlearninglibrary.mit.edu/courses/course-v1:MITx+6.036+1T2019/courseware/Week4/gradient_descent/?activate_block_id=block-v1%3AMITx%2B6.036%2B1T2019%2Btype%40sequential%2Bblock%40gradient_descent

Summer

ToDo Week of 22/5

• refine my knowledge in Deep Learning: finish this class http://introtodeeplearning.com/

• test and run simulations on Setven's work

ToDo Week of 29/5

• FL tasks : Select sleep tracking and implement it on FL simulation.

• details to be added later as I read papers and learn about this task -> chosed paper: FedMCRNN: Federated Learning using Multiple Convolutional Recurrent Neural Networks for Sleep Quality Prediction https://dl.acm.org/doi/10.1145/3512731.3534207 used data: https://datasets.simula.no/pmdata/; paper: https://dl.acm.org/doi/10.1145/3339825.3394926

ToDo Week 6/5

• Finish the implemention of sleep tracking. Simulate the paper mentioned above.

• Use other algorithms.

Implementation details: Data Preparation: Accodring to the paper, only six important activity categories were used to assess sleep quality effectively: calories, distance, sedentary minutes, lightly activity minutes, moderately active minutes, and very active minutes. I still need to research the reason behind this choice The used files are .json. So far, I've completed 70% of the cleaning of the data (remove duplicates, modify empty values) of one participant out of 16. The cleaning of the data of all participants are similar. Two participants have incomplete data, so they'll not be considered in the study. After cleaning all the data, the originally sepearte json data files will be merged into a single file for easier manipulation later.

ToDo Week 6/12

• Timeseries analysis of the data set/work on data windowing

• Application of MLP, CNN, and LSTM

• Data Windowing: function that can generate any form of a window by selecting the input width, shift, and output width

• Data preparation: (70%, 20%, 10%) split for the training, validation, and test sets data is not randomly shuffled before training to ensure that chopping the data into windows of consecutive samples is still possible, and to ensure that the validation/test results are more realistic, being evaluated on the data collected after the model was trained.

• Normalization: subtract the mean and divide by the standard deviation of each feature

• Applying MLP, CNN, and LSTM, however, the accuracies are very low.

ToDo Week 6/18

• Achieve the best accuracy for training

June Self-evaluation report

• Goal Attainment:

The original plan for June was to finish implementing a model for the sleep efficiency task for the app by the end of mid-June, and then work on other tasks. However, as my work progressed, this goal seemed far to reach. I ended up spending most of the month working only on sleep efficiency. The biggest challenge I faced was the unavailability of open-source papers that tackle the same problem: "sleep efficiency", so I just learned how to implement a model by myself from the ground up. Also, as I was progressing, I wrote down weekly plans but most of the time I didn't finish them because more challenges came up that I wasn't expecting.

• Skills Development:

First, I started by looking for a dataset and doing data preprocessing. I chose the PMData dataset which was used in a research paper I read. Surprisingly, data preprocessing was time-consuming. It took me a while to finish it because there were different files that needed to be merged together based on the date_time, and making decisions on which data to keep and how to fill out empty cells was also challenging. During this process, I refined my programming skills and mastered some libraries like Pandas and sklearn. Next, I read this book "Deep Learning with Python" and other online articles to learn more about deep learning. I can say I learned a great deal about DL which helped me work with DL models more confidently. Also, I learned about time-series forecasting which was used in our sleep efficiency task. Later, I started writing models for the previously described task and used KerasTuner for automatic hyperparameter search. KerasTuner was also a new framework I learned about. Overall, I believe that although I didn't produce many concrete results this past month, I learned a ton which can hopefully speed up my work in the future.

• Collaboration and teamwork:

As each member had a different task to work on, I didn't feel like there was much collaboration in our team, at least for the algorithm team. Most of my work was done individually. However, I liked the weekly group meetings because they gave me a boost of energy and motivation, and I enjoyed learning about other team members' progress.