Choose an application

• Reference Manager
• EndNote
• RefWorks (Direct export to RefWorks)

During the last decade, deep learning played a leading role in both research and practical applications of artificial intelligence. Its ability to learn from massive datasets makes this approach able to perform tasks that typically require human intelligence such as recognizing images, understanding natural language, predicting customer behavior, etc. Nevertheless, it is becoming clear that deep learning struggles with reasoning and generalization beyond the dataset, which is limiting further improvements in the field of artificial intelligence. On the other hand, logic-based methods (also known as symbolic artificial intelligence) excel at reasoning and generalization. Even with a limited amount of data, those kinds of methods are able to reason about the problem and deduce new information. However, if there is noise in the provided dataset or if this dataset gets too large, these methods lose these abilities. The idea arose that a combination of deep learning and logic-based methods could be successful on certain tasks because the strengths and weaknesses of both approaches are complementary. This concept is explored in the field of neural-symbolic artificial intelligence (NeSy). This field is quickly becoming one of the most popular fields in artificial intelligence research and experiences a fast growth because many methods that provide a combination of logic and deep learning are proposed. This leads to the problem that methods have only been evaluated on a limited number of tasks and few experimental comparisons are done between the different methods. The results of this work are two-fold. First, tasks for which it has been shown that neural-symbolic methods can be successful are grouped into categories. This creates a clearer view of the capabilities of these methods and highlights the use cases of neural-symbolic artificial intelligence. Secondly, the requirements for methods to perform well on two of these tasks (distantly supervised classification and structured prediction) are discussed in detail. With this information, an assessment is made to state which neural-symbolic methods are suited best for these tasks. The actual performance of the methods on these two tasks are then verified experimentally. These analyses allow for the initiation of a discussion about the general applicability of the proposed neural-symbolic methods.

Choose an application

• Reference Manager
• EndNote
• RefWorks (Direct export to RefWorks)

In dit artikel zal je meer lezen over de STEM-fiets. Wij hebben een praktisch praktijkonderzoek gedaan en willen graag dit artikel met jullie delen. We formuleren een antwoord op ‘Hoe kunnen we educatief speelmateriaal ontwerpen voor de speelplaats?’ . Achtereenvolgens bespreken we hoe we ons ontwerp duurzamer kunnen maken, hoe we zorgen voor een goede verankering op een speelplaats en hoe we voor een breed, kwalitatief, zinvol en toegankelijke aanbod kunnen geven aan de leerlingen. We hebben hiervoor samengewerkt met experten en leerkrachten, om zo extra informatie te krijgen over het technische gedeelte van de fiets. De grootste samenwerkingen zijn verlopen met WeWatt en de CEO hiervan, Patricia Ceysens en met UCLL eXPerience Lab. We hebben twee ontwerpen uitgewerkt, een op één low-tech niveau en één op een hightech niveau.