Mia built an artificial brain that knows a barn owl from a stork
As a hobby ornithologist, Mia Pontoppidan wanted a tool for identifying bird species. So in her thesis project from software development at ITU, she investigated how well artificial intelligence (AI) could classify Danish birds – and whether you can build a neural network with only a year and a half’s worth of coding experience.
Why did you decide to write about bird classification?
As a birdwatcher, I know that identifying species based on images in a book can sometimes be difficult. For example, seagulls can live for up to 25-30 years and go through several plumages that look very different.
So I thought it would be fun to see if AI technology, or deep learning more specifically, is advanced enough to be able to classify pictures of different Danish bird species.
»I built a Convolutional Neural Network (CNN), which is a neural network used for image recognition. It’s called a neural network because it’s actually a mathematical modeling of the brain’s visual cortex.
I thought it would be fun to see if AI technology, or deep learning more specifically, is advanced enough to be able to classify pictures of different Danish bird species.
What was your method?
First, I compiled my own dataset with hundreds of pictures of five different Danish bird species. The goal was to train a neural network to identify and correctly classify these species.
The more pictures you train your neural network on, the better it will become, so I trained my network on different datasets to see which method gave the highest accuracy.
I got the best results from training the network on ImageNet, which is an open data set with over a million images of everything from houses and cars to pictures of birds and other animals. By finding patterns in this vast amount of image data, the network learned to identify if it was a bird in the picture, or something else, like a cat or a house.
Then I trained the last part of the network on my data set with Danish birds. In the end, it classified these bird species with an accuracy of over 72 percent, which is very good for such a small data set.
What did you learn along the way?
My bachelor was in business economics and communication, so I only had a year and a half’s coding experience and no knowledge about neural networks before starting the project.
This meant that a large part of the task was to understand the underlying theory and learn to actually build the network. I also learned a lot about how intelligence works - both the artificial and the biological kind.
»For me, the thesis was also about finding out how accessible the technology is – how well I could get the network to identify the birds with a limited technical knowledge and only a laptop to work on.
It was really exciting and motivating to be able use technology to solve an everyday problem.
It was challenging, but also fun to see that you can actually get quite far in a short period of time. I learned a lot from the thesis, and it was really exciting and motivating to be able to use technology to solve an everyday problem.
Can we expect to see a bird-identifying app on the market anytime soon?
I think it will soon be possible to an app that can classify birds using images. Neural networks can’t very accurately distinguish species of seagulls that are difficult even for humans to identify. But this will definitely become possible.
Researchers at for instance Cornell University in the US are working with image classification of bird species, and of course they have far more resources and computing power available than I had.
What are you working with today?
I work as a technology consultant at Accenture, where I can combine the technical IT understanding I got at ITU, and the business understanding and communication skills I got from my bachelor’s degree.
Vibeke Arildsen, Press Officer, phone 2555 0447, email email@example.com