Road Signs & Machine Learning: Day 2

By: Abhi Mora

Implementing data into Teachable Machine

I have just uploaded my images from my phone into Teachable Machine. This process took a very long time, as it seems as though Teachable Machine doesn't support HEIC files. Is this something related to the artificial intelligence not being able to use these files? Or is it just the website? I hypothesize that it might just be the website because HEIC files are usually unpopular and may not be able to be read. When I researched a bit into this, there wasn't much that came up, so I believe that it was a problem with the website, and these types of problems aren't limitations for Artificial Intelligence.

Once I uploaded the images into Teachable Machine, my camera turned on and it said that I looked more like a stop sign rather than a no outlet sign, I tried it on my friends and it was pretty fun. However, once I got back on track, I went on Google and searched for these signs and put those images in front of my camera, and the AI was one hundred percent right.

After I had fun playing around with this model, I researched more about what Teachable Machine does with the AI and found that only 85% percent of the images are used to train the model, while the other 15 percent are used for testing. These test samples help the AI understand if it is accurate or not based on never-before-seen data. This confuses me, as I believe that it doesn't need to test the images on images I give, but it should go on the internet itself and look for information itself.

For the last thing of the day, I added some images from the internet, and something that shocked me was that the AI was only 93% sure that one of the clearest stop sign images was a stop sign. Although it is a high number, it isn't nearly safe enough for an AI to be implemented on the street and have a possibility of misinterpreting a sign. When I tried to put Peter Griffin from Family Guy, the AI said that he was 100 percent stop sign. This makes me realize that AI isn't as reliable as it is being used today, and for it to be safe, trillions of bytes of data need to be put into place before it can be used.