Birds of a Feather

by Raghu Srinivasan

Birds of a feather flock together. That is partly how we've trained a machine learning model to tell them apart. Over 27500 photos of 200 different bird species from this Kaggle dataset were run through a convolutional neural network a couple dozen times. That's about 125-150 images of each bird species. We then validated with another 1000 images i.e. about 5 of each species. The accuracy of the model is about 75%. That means that 7-8 times out of 10 the model correctly predicts the species of bird when presented an image. A purely random guess would be right 1/200th of the time which is only 0.5% so the 75% is fairly decent.

To get started simply choose a file of an image of a bird from your computer or mobile device and hit upload and a prediction will be made for your image! Remember there are nearly 10,000 species of birds on Earth and this model has learned just 200 of them!