Thread on Improvements

[dmklee]

I am opening a thread to list and discuss improvements. It will be helpful down the road to have our impressions recorded while they are still fresh

First Coding Day (MNIST digit recognition)

• The students' school Gmail accounts could not run Colab due to access. Mr. Burges (I believe) said that this is not fixable since Google prevents certain services for accounts of minors.
• We solved this issue by having students sign in with their personal Gmail accounts. Almost all students already had one, but we should make this clear that students need to create one beforehand. We instructed them to open Private tabs and then sign in with their personal account during class. In hindsight, I do not think Private is necessary; instead, students can choose to switch accounts in their normal browser.
• Adding questions to the notebook was helpful, as it gave students a clear task to work on. It also made it easier to start discussions with students while walking around.
• The notebook should be changed so there is more control over the final accuracy. We told them to play around with model architectures, but the provided model already achieved near-perfect accuracy. Thus, the feedback loop was not great. Also, making it harder to get a high accuracy should create more engagement in the form of competition between groups.

[dmklee]

Second Coding Day (Fetch Housing Data)

• I think the exercise led to great discussions, and it was nice to be able to adjust the input features
• The notebook can be cleaned up a bit (the final two plots were barely used and failed if additional features were used). I think there were some advanced techniques that could be ignored (e.g. parameters of Adam, learning rate schedule), and the code could be modified to make it easier to modify the network (e.g. create variables for input_size, hidden_size and output_size). I also think the names of all variables should be provided in the "Input Variable Selection" section
• The students seemed more engaged because their changes resulted in noticeable changes in model performance. (There was a decent amount of noise due to random seed, which we could fix if we wanted)
• It could be a cool plot to show the predictions based on longitude and latitude overlayed on a map

[dmklee]

Visit Day

• Itinerary: walk to Northeastern (20 min); student presentations on what they learned (15 min); wrap up lecture on overfitting (10 min); grad student talks about applications of ML (15 min); snacks (~20 min); tours of ISEC Robotics lab with demos (30 min); walk back
• The walk can be cut shorter by going thru the back of OB and it is wheelchair accessible
• Having visit on last day of program seems better because the students have more understanding of ML so the grad student presentations are more accessible
• The paperwork process still seems annoying. I think it may be best to try to get the admin folks in a face-to-face meeting/call; I get the impression that some of it is a formality and could be streamlined if you talk to the right person.
• Overall the visit went really well! The students were very impressed with ISEC