Lenka Hýlová: Giving Machine Learning Algorithms Additional Knowledge About The Real World Can Improve Their Performance
Recently, the usage of machine learning algorithms in various applications has massively increased. Machines are capable of completing tasks that seemed to be impossible a few years ago and often outperform humans. However, each machine learning algorithm can solve only a narrow range of tasks - only those that it has already seen in the training process. The more training data we have, the better the extrapolation from training data is. Unlikely humans, they are not able to adapt themselves to other, previously inexperienced, situations. Moreover, if we have only a little data to train on, the performance suffers.
The world is complex and humans, often unconsciously, use information and experience that are not directly connected to the assigned task. For example, we know that fish are very unlikely to live in a desert, so we recognize an unknown animal on a picture from the desert to be a lizard, even though they look similar. If the model had additional information about the world, its objects and their relations, it would be able to better handle unknown situations. Such information can be supplied for example in the form of knowledge graphs, that are already being used in search engines, social networks and other applications.
One of the possible uses of machine learning is an automated image‐based inspection of grains used to assess their quality. There already exists a system EyeFoss™ that uses feature‐based models to sort wheat and barley grains. However, the grains differ from year to year, from one field to another and also the customers’ needs vary a lot. The existing model can handle only a limited variation of all the possible circumstances and, therefore, needs to be completely retrained every time depending on the current needs. The automatization would save energy and human resources needed for frequent retraining and improve performance. Moreover, it would reduce the labour needed to apply a similar procedure to other types of seeds.
This project has two goals. First, improve machine learning algorithms for image analysis by giving them additional knowledge about general concepts in the world. Second, these theoretical results will be in collaboration with the project partner FOSS applied to the problem of assessing the quality of grains and seeds. This will help to increase production and reduce food waste by improving the classifier. Furthermore, a similar approach may be applied to other types of grains, like coffee or maize, even though the training data for these grains is limited.