Activity Five
You’re starting a new venture firm MACHINES ARE SMART, that specializes in machine learning. You are going to pitch a new project to venture capital investors, and you need a proof of concept. Use teachable machine to build a model that does one of the following:
- A special pet door that only lets in cats, for cat lovers.
- A foreign language training app that teaches you to say “hello” in a foreign language and rejects attempts that aren’t correct.
- An app that rejects drinks that aren’t coffee.
- An app that tells if elderly nursing home residents are happy or sad, and alerts loved ones if the nursing home resident is too sad too often.
Reflect on the following question: You’ll have to think about how to train the machine. What kind of data did you include in your training dataset and why? What other kind of data could have been helpful but maybe you couldn’t get in the short-term/for free? Your group may, in some cases, search for photograph sets. One possibility to get large data sets is to convert YouTubes into clips. Did your model work well for what you wanted? In what instances might your model not work very well? Include the link to your project.
Prompt Answers
We included 20 pictures of different coffee drink from black coffee to lattes and mochas in the “coffee” dataset. In the “not coffee” dataset, we put drinks of various colors and different looks. We put some similar in color to coffee in an attempt to get the AI to recognized that not every brown drink was coffee. Data that might have been helpful would have been different videos of coffees and non-coffee drinks or a wider variety of images for both datasets. For the time given, the model did work pretty well. When a picture of coffee was held up, the “coffee” rating was in the 80-100% range. When a non coffee drink was held up, the “coffee” rating was in the 0-25% range. To test if any drink that was brown would register as coffee, we held up images of chocolate milk, and the AI recognized it as “not coffee.” While this is a very simplified image recognizer, it did work pretty well for what we taught it to do. The model might not work very well with more complex images of coffee, or if we held pictures of tea or other drinks that can look very similar to coffee. It may have a hard time recognizing the difference between the two.
Link: