Delivering excellence in our work requires us to not only stay informed about upcoming developments in technology, but also be part of shaping it. Research in the field of artificial intelligence, machine learning and deep learning enables us to deliver state of the art solutions for our clients.
At craftworks we are currently mainly focusing on these four areas:
Meta Learning is a subfield of machine learning and is often also referred to as “learning to learn”. In Meta Learning, machine learning algorithms are trained to make predictions based on metadata instead of raw data. For example, a Meta Learning algorithm learns to predict the accuracy of a type of neural network applied to a specific supervised learning problem. The algorithm does this by using the properties of the neural network, such as number of neurons and the architecture, and characteristics of the dataset specifying the learning problem, such as the number of variables, as input. Therefore, Meta Learning allows for generalization of learning problems and may lead to drastically increased efficiency when substituting otherwise expensive experimentation procedures. At craftworks, we research and use Meta Learning techniques for exactly that.
Applying machine learning to real world problems is not only time consuming and resource intensive, but also quite challenging. This is where Automated Machine Learning steps in: it frames the machine learning process as an optimization problem that can be solved efficiently while reaching a solution that is close to optimal. The capabilities of Automated Machine Learning range from data cleaning and feature engineering all the way to model selection and hyperparameter tuning. This frees up time of data scientists and, additionally, enables people not familiar with machine learning to leverage its power. Therefore, several engineers are working on developing cutting-edge Automated Machine Learning software at craftworks.
The field of Representation Learning, or Feature Learning, is engaged with the automatic discovery of optimum representations of data given a specific learning problem. Learning the representation of data instead of crafting it manually using traditional feature engineering techniques has three main advantages. First, it is time saving as it replaces manual feature engineering. Second, due to the fact that the representation is learned in the process of model optimization, the resulting features are often significantly superior to hand-crafted ones. Third, representation learning often has the useful byproduct of exposing previously unknown structures in the data. At craftworks, we use Representation Learning techniques in a variety of problems, ranging from Natural Language Processing to Computer Vision.
Reinforcement Learning is an iterative optimization approach that is motivated by the way humans learn: trial-and-error. In combination with deep learning techniques, Reinforcement Learning is used for teaching computers highly complex games, make robots learn arbitrary tasks, optimize control systems of power plants and many more. Also craftworks develops novel Reinforcement Learning approaches for solving multi-agent problems with incomplete information.