AI Can Describe Photos!

I have just completed the most complex homework in AI that I have ever done. Creating and training a model that can generate text descriptions from photos!

The architecture is based on a CNN encoder and an RNN decoder; more on this can be found here: https://research.googleblog.com/…/a-picture-is-worth-thousa… and https://cs.stanford.edu/people/karpathy/.

The pre-trained CNN network, based on the beloved InceptionV3, learned the descriptions in just 10 minutes. It’s not perfect, but the results still take my breath away! After all, judge for yourself.

Finally, a big thank you to Andrej Karpathy for his amazing research and to the lecturers from the Russian school for preparing it so beautifully.

For your interest, the network learned the descriptions from many, many examples that people submitted using Mechanical Turk. Have you heard of it?