Team D5: Autonomath

Here is a link to the YouTube video: https://youtu.be/oNiYizB6CVg

https://docs.google.com/document/d/18jZyMWeDjq650P1P0c9-Uf5SmIs8uBe4CLD360bi_II/edit?usp=sharing

Above is the link to our final report as a Google Docs!

David Diao (dyd)

This week I worked on preparing for the final presentation. We made slides and created a presentation in preparation for next week’s presentations. I also worked on debugging our AWS issues with database integration.

 

Jessie Li (jessieli)

This week I worked on preparing for the final presentation. Our group made slides and created a presentation in preparation for next week’s presentations. Our project is still having bugs with AWS, but it is generally coming along.

David Diao (dyd)

This week I worked primarily on integrating the local web application onto AWS. I also worked more on the frontend user interface and worked with Jessie to conduct some user studies on the workflow of our application.

 

Jessie Li (jessieli)

This week I worked primarily on helping David integrate the local web application onto AWS. In addition, we developed on with the frontend user interface, doing a brief user study to test the flow of data through the system and how the user can interact with it. We made some changes to the frontend based on the user study to make the application more intuitive.

David Diao (dyd)

This week I worked on debugging the integration of Tushar and Jessie’s code in my backend. I also worked on the frontend so the user can check if the backend recognized their digits correctly. After that, the user can edit or submit their answer in order to see their result and move onto a new question.

 

Jessie Li

This week I worked on switching the dataset to MNIST and developed on and trained my CNN to work on the MNIST test set. The resulting accuracy is 98%, but while doing user testing, the accuracy seems somewhere around maybe 70%. This is perhaps related to how we are processing the image to classify; more testing is needed. I also helped Tushar work on his segmentation algorithm, writing a connected components finder in order to aggregate the labels found using the segmentation algorithm.

David Diao (dyd)

This week, I worked on our team’s demo and getting my component ready for the demo. I finished integrating Jessie’s CNN and Tushar’s segmentation portions into the backend, and now have much of the basic use cases finished.

Jessie Li (jessieli)

This week, I worked on our preparing for out team’s demo and getting my CNN component ready for the demo. I trained a few versions of the CNN and tested them for accuracy; the accuracy is still a bit awkward and not as high as I’d like, so further testing is required.

David Diao (dyd)

This week, I worked a lot on the web app to get it ready for the demo. I finished building out the basic frontend interfaces that we will need, and also finished up the backend endpoints that will be crucial for the demo. In addition, I worked on integrating Jessie and Tushar’s code into the web app.

Jessie Li (jessieli)

This week, I worked on building a CNN for the HASYv2 dataset. I trained a few networks with different epoch and batch numbers, but the highest I could get a random test set to was 79%. Part of this is because the reduced HASYv2 dataset is fairly small; much of the huge dataset we did not actually need. This needs some discussion; for the demo, at least, we can use the MNIST dataset of just digits (no characters or symbols), which my CNN has trained to a 98% accuracy rate.

 

Tushar Chetal (tchetal)

Had to change the segmentation algorithm based on the changes we discussed. Finished writing the segmentation algorithm.

Jessie Li

This week, I continued to develop on the convolutional neural network. I researched a few different structures for the convolutional neural network, and implemented a from-scratch version of a basic network network with 2 convolutional layers, 2 max-pooling layer, and a multi-layer perceptron using Python and numpy. For the rest of this week and next week, I will be testing the network on MNIST dataset first, and then a modified (slimmed down) version of HASYv2.

David Diao

This week, I continued to develop the web application backend. I created all our backend endpoints and endpoints to communicate with DynamoDB. However, currently, I don’t think we’ll be using DynamoDB as we want to get the more important parts of our project working first. I began writing the endpoint handlers but did not get too far as I got sick for the latter half of the week.

Tushar Chetal

I was out for interviews this week and did not have much time to work on the code. Worked on the ethics assignment.

Jessie Li (jessieli)

Unfortunately this week, I was out of town due to some medical appointments. I made some progress on a CNN for the HASYv2 dataset, drafting up a rough version of the CNN. For the week after spring break, I plan to do some preliminary testing of the dataset on my CNN and iterate based on accuracy.

David Diao (dyd)

This week, I drafted and designed how our frontend interface would look. In addition, I worked more on the backend and database design that I started on last week. I ended up learning more about how to integrate our Go backend with DynamoDB, and also how to use the python code we’re running our CV and CNN on through Golang.

Tushar Chetal (tchetal)

This week I started writing out the code for the different classes that I designed last week and started writing some code for the segmentation.

David Diao (dyd)

This week, I worked on the backend design of our web application. I also did research into how the web app would work on a Windows Surface Pro desktop app vs a online web app, and we decided that we should stick with our web app stack. For the backend, I drafted up multiple endpoints that we could possibly use, and drafted up schemas of what our tables would look like in DynamoDB. I also obtained AWS credits from Professor Nace and will be looking into how to work AWS into our project and how to link everything together.

Jessie Li (jessieli)

This week I redid much of the work I lost when I lost my computer. I downloaded the dataset, manually cleaned out the data, and worked through the steps of creating a CNN to classify the data, manually picking classes to work through. In addition, I began to research and develop a convolutional neural net on the cleaned version of the HASYv2 dataset. To start, I tried training a CNN using sklearn (from python) to see what kinds of results I could achieve on the cleaned dataset and to gain experience manipulating the dataset. I also began work on the design report due next Monday.

Tushar Chetal (tchetal)

This week I couldn’t get much work done since I was traveling for interviews and did not spend much time on school work. I did start writing designing the classes and description of objects theoretically.