Our team wants to build a machine learning tool to empower people, particularly non-technical domain experts and users who like to explore. We came up with DashML, a web-based machine learning dashboard to provide them simple solutions without coding.
What it does
At DashML, we are building a machine learning dashboard that benefits both end-users and the app developers/domain experts. While users can upload an image and receive a classification result, app developers could utilize the insights presented on our dashboard to improve their models further.
The current pandemic may boom the telehealth programs, and we see DashML could fit into this. For example, a user needs diagnosis for mild skin conditions but is not able to see a dermatologist, such as eczema, psoriasis, and rosacea, which requires domain expertise to distinguish. In this case, a user could upload an image of her/his skin condition to DashMl. Our model would provide the user with a real-time diagnosis through our image classifier model trained on this subject matter. On the other hand, the domain experts could use the DashML dashboard to understand the model result better through understanding the actual and predicted class of the image and the probabilities
How I built it
We trained the MNIST model in Python using
scikit-learn. We used
bentoML in order to serve the model as an API. The dashboard frontend was built using React in conjunction with simple HTML/CSS/JS. We built a JS proxy server to convert POST requests to the
application/octet-stream format to work with bentoML.
Challenges I ran into
bentoML really requires a lot of specific things—it struggles with
keras (rather than
tf.keras), and it requires POST requests be in the format
application/octet-stream. It also does not have a lot of useful documentation at the moment. All challenges we had broadly fall into making the different parts of our project interoperable.
What I learned
We learned that time management and communication are key. It was exciting to work with other pod mates during this first week of fellowship as we not only learned a lot from working on the project itself but also learned and got to know each other through this Hackathon.
What's next for DashML
Given more time, DashML would like to 1) train more domain-specific model 2) create a Grad-Cam heatmap including actual/predict/losses/probabilities to provide more insights to domain experts.