ObML: It would be nice if the tutorial covered categories that aren't already in the standard big generic models. For example, instead of dog and cat, use something new that doesn't exist in the model, and show how to train for that.
Also as feedback, the terminology "submit your images" clashes with the statement that data doesn't leave your own system. Does the latter statement mean "in normal use after training is completed"? As opposed to during training, where the images do leave your system? There is wiggle room for either meaning here, so more clarity would be helpful.
Base64 uses a 64 character alphabet (hence the name), so you can only represent 64 different values within each byte. 64 == 2^6, so basically you are using 6 bits out of every byte and losing the other 2 bits.
HTTP can use compression to reduce the size of the Base64 representation, but assuming the images were already using a compressed format, all this is going to do is mitigate the inefficiency, it's not going to be more efficient overall.
This mainly applies when you have to transmit binary data using a text-based format like JSON or XML. I've also used this before to build a shell-script-based installer, where a compressed archive is base64'd in a variable.
ConvNets for image classification (like what the article is about) often achieve better-than-average-human performance
There is false, except under some contrived tests. NNs still aren't very good at handling rotations, can be confused by changing single pixels and often recognise shapes out of complete static.
But not entirely clear when looking at the site, what would be the pricing to classify say 20,000 images from 20 different folders?
Classifying events is often most easily done by looking at the metadata of the image, so converting to base64 just the image is probably a bad idea. (Specifically, I'd expect the date/time and possibly location data to be of help.)
Classifying the people is a helluva challenge. As someone in a family of 6, I don't think machine classification has much of a chance on getting pictures of my kids right. Especially if it doesn't have the dates on the pictures. I've seen some pictures of my oldest daughter that I would swear are my youngest son.
Many AI classes talk about the fact that metadata are less often relevant than we think and are often embedded in the image in a way the classifier easily learns.
In an event, the classifier will probably spot a few architectural elements that will make useless to rely on the time of the picture.
On my first attempt at image classification with deep learning I was surprised at how a very badly design and not custom-tailored network managed to outperform a classifier that had several days of work into it.
I'd say give it a try anyway, you could be surprised.
None of this was to say this isn't impressive. It truly is. I don't know why my bar for the magic of image recognition is so high.
ready to clone & run...