How does the neural network learn?
The neural network, NN, is actively taught by a deep learning professional who has access to a good amount of training data. In short, an image of a varroa infested bee is shown to the neural network and we let the network guess. If the network guesses wrong, we let it know. It will then adjust itself very slightly to make a more educated guess next time. By doing this several thousand times the neural network will learn very abstract features and eventually become better than a human at visually detecting mites.
However this is a very simplified way of looking at it. In reality, an artificla neural network is a very complex beast that is difficult to tame and comes in many different forms and shapes. One of the first milestones for the development team was to find the way of training that is optimal for this particular task. Finding very small objects in a huge amount of information.
What if you don’t find enough training images for the neural network?
A crowdsourcing campaign secured resources to develop Tagger. Our tool to structure images and manually label regions to train the NN. We have collected about 4000 images and will contnue to use the data gathered from the app as a tool for further ongoing training of the neural network.
Is the development team actually capable of solving this task?
Yes! The image analysing team members have over 15 years of experience with developing complex systems. The team is lead by mobile and tech industry veteran Emil Romanus. Romanus has previously been involved in mobile world-wide successes, among others the complex 3D and physics simulation game “Apparatus” in 2011. Romanus and his team has been focusing on deep learning for the past few years and have made several large investments in the hardware required for the training of neural networks.
Does the development team have experience with similar solutions?
Romanus’ team has worked with several other companies and solved similar issues. Most importantly the team has worked with a british company that identifies and reports the occurence of company logos and brand marks within photos, a task as complex detecting varroa mites on bees.
Building the app, backend and other software that is required for the project is a trivial task for the team.
Why hasn’t this been done before?
The hardware wasn’t available and the idea was probably too far-fetched to be taken seriously. Teorem has the hardware and the team has solved similar tasks. The innovator Björn Lagerman has a long record of combining knowledge in new and creative ways as well as the experience to lead complex projects.
Where can I upload images?
Please collect images. Especially from frames where brood is visible and with varroa on bees. Upload on http://tagger.beescanning.com and contribute by labeling regions.
How is the project financed?
Initially we were funded by backers on Kickstarter. Then we received financing from the European Innovation Program, EIP. We are also supported by Almi Örebro and Handelsbanken in Lindesberg. Please read more at http://fribi.se/sponsorer/