AI is extremely good at identifying and using patterns. However, when it comes to relational reasoning, artificial intelligence systems badly fail at it.
Relational reasoning is an essential part of higher thought and logic. It allows humans to make connections and compare places, sequences, entities and other variables.
Researchers at Google’s DeepMind have tackled this problem and developed a simple relational reasoning algorithm to help machines use logic. The new system relies on AI neural networks that mimic the way neurons are connected. More specifically, dedicated neural networks called relation networks work together to identify the connections and relations between entities.
Relation networks are AI architectures that focus explicitly on relational reasoning. The network compares pair of objects individually in order to identify the relationships that exist between the objects.
This is the first major step in building smarter AI systems capable of reasoning about the relations between entities and their properties. Unlike other reasoning networks, RNs are simple, plug-and-play, and flexible, allowing researchers to use them in various fields.
Researchers used RN-augmented networks to a variety of tasks that heavily rely on relational reasoning, including visual QA, text-based QA, and dynamic physical systems. Thanks to the versatility of these networks, researchers were able to achieve state-of-the-art, super-human performance.
The visual QA test
Visual QA tests challenge AI systems to answer questions about an image. This task requires high-level scene understanding, and complex relational reasoning – spatial and otherwise. To make this task even more difficult, oftentimes vast knowledge of the world is not available in the training data. Moreover, questions contain ambiguities and linguistic biases.
DeepMind researchers used two versions of the CLEVR dataset for this particular test: the pixel version, in which images were represented in standard 2D pixel form, and a state description version, where images were explicitly represented by state description matrices with factored object descriptions.
Here are the results:
Our model achieved state-of-the-art performance on CLEVR at 95.5%, exceeding the best model trained only on the pixel images and questions at the time of the dataset’s publication by 27%, and surpassing human performance in the task (see Table 1 and Figure 3). These results – in particular, those obtained in the compare attribute and count categories – are a testament to the ability of our model to do relational reasoning. In fact, it is in these categories that state-of-the-art models struggle most. Furthermore, the relative simplicity of the network components used in our model suggests that the diﬃculty of the CLEVR task lies in its relational reasoning demands, not on the language or the visual processing.
Using relation networks to improve AI reasoning
Relation networks are very useful for identifying inter-entity relations. By embedding relation networks modules into deep learning architectures, researchers can improve AI performance on tasks that involve relational reasoning.
Experiments revealed that RNs are very flexible and are capable of structured reasoning even with unstructured inputs and outputs.
Relation networks can be used to enhance reinforcement learning agents’ capacity to understand rich scenes, model social networks, and solve abstract problems.
For more information, you can check out the research paper here.
YOU MAY ALSO LIKE: Chinese companies may soon overtake Google in AI researchFollow The AI Center on social media:
Join me as I track the latest progress in AI research.
Latest posts by Mary Blaut (see all)
- AI powers staff-free stores: Shop without any human interaction - January 18, 2018
- AI helps researchers improve gene editing results - January 13, 2018
- True Emoji is an AI app that uses your expressions to create animated emojis - November 21, 2017