Did you know that eButterfly has an AI-powered Image Recognition Tool (eButterflAI) that can help you identify a butterfly species from any picture?
Here’s a quick infographic showing you how to use it
Did you know that eButterfly has an AI-powered Image Recognition Tool (eButterflAI) that can help you identify a butterfly species from any picture?
Here’s a quick infographic showing you how to use it
Curious to know if this AI draws on the ever growing database of photos to generalize a species phenotype? If not, how is the reference image chosen?
That is an excellent question @GilesR !
Our AI team used a massive database of images (more than 100 GB) to train our North American algorithm, allowing it to identify around 2,600 different species.
The algorithm analyzes the images and “learns” how each species looks; that way, when faced with a new photo, it draws from its learning with the training images and proposes a species.
You can see in the image below a heatmap created by the algorithm, where the “hottest” colours represent the areas of the image that the AI considered more critical as telltale elements to identify that species.
We are currently working on a global model using an even larger dataset of roughly 2 million images, allowing it to identify more than 15,000 species worldwide. This model will be released very soon!
This is great! And very useful. I am very impressed by developments at e-butterfly. So, as I understand it, photos we submit to e-butterfly are used to characterise species. But I wonder if the AI is alert to mimicry, polymorphism, gynandromorphy, and can it recommend when naming a new subspecies is appropriate? Perhaps next year?
That is very interesting question @GilesR . Indeed, each photo submitted to the platform is fed to the AI algorithm so it will keep increasing accuracy overtime.
The model is totally capable of dealing with mimicry and polymorphism as long as it has enough examples of it in its training database. And even though it might not be able to directly recommend naming a new subspecies, it can flag “weird” findings so we can share them with experts to be identified manually ( and adding it to the training database) or, if appropriate, recommend a new subspecies.