Recognize Player Achievements via Machine Learning 🤖
Closed this issue · 5 comments
Hey team,
We are currently looking into ways that the app can use machine learning algorithms to detect when a player completes a carbon mission. This should make missions easier for players to complete in the future!
Suggestions by Make School alum @RaymondDashWu:
"hmmm so my impressions from that photo are that users will be scored based on tasks they individually do to help them, which might be a bit difficult without a lot of data already since people's tasks and goals are individual... One thought I had would be to use a pretrained model that has been trained on the COCO dataset, which is a large repository of labeled, segmented images covering a lot of categories. That way it can be something like: recognize a recycling bin and give X points"
Suggestions by Make School Data Science Instructor @jcatanza:
"A more direct approach to this problem might be analyzing text (rather than images) in Twitter or Instagram feeds via methods of Natural Language Processing (NLP). Though I am not sure whether this approach meets the goals of your industry partner.
It would be straightforward to train a basic sentiment classifier model to recognize tweets associated with (positive or neutral) environmental value. Though you would need to start by either creating or finding a large set of training data, i.e. ~thousands of labeled example tweets"
Original Question I asked:
"Hey team,
What are your favorite pre-trained models to use for image recognition? Our industry partner for SPD 2.2 is interested in building an image classifier for 'recognizing things people do that are good for the environment.'
Examples may include:
- eating a plant-based meal for lunch
- sorting your trash
- installing LED light bulbs around the home
I’m trying to think of a way to use transfer learning to make building this more efficient. Alternatively, I might need to scope down the feature, since there are probably ways to do this feature w/o machine learning. Any thoughts?"
Ideas on Transfer Learning proposed by @RaymondDashWu:
"Resnet, VGG, or YOLO. If you need a mobile app maybe something like Mobilenet or convert your models to FP16"
The feature has been implemented in #124, so closing for now as the team moves on to the next sprint.