It’s easier than it sounds.
Training a Neural Net to use computer vision, identifying different sets of objects in images.
At [REDACTED], we use neural nets to help us identify customer objects within frames from marketing videos. This net was connected to our web app, to aid in the outlining of objects.
Example of how the neural net would identify objects. In this case, being trained on “kitchens”.
While working on this project, I helped collect and organize training data, had the data annotated and verified annotations. With the help of some interns under my wing, we were able to use a python script to scrape the web for image search results, run another script to clean up the data to leave us with useable images, and then upload to an annotation service (Labelbox).
I would then iteratively train the net, verify results, adjust data for desired results, and try again. I worked on several different categories, including “Humans”, “Clothing”, “Jewelry”, “Makeup”, and more. Each set would include 10+ labels. For example, “Clothing” was trained with test data and labels identifying “pants”, “skirt”, “shorts”, “shirt”, “shoes”, etc.
Anyone who tells you they can make a truly unbiased neural net or AI is lying.