A research review about reasoning in Artificial Neural Networks [Link]. The paper briefly reviews recent approaches in building artificial neural networks that can learn to reason, such as attention mechanism, memory augmented neural networks, neural-symbolic AI and graph neural networks. This is written as part of project preparation prior to my MSc dissertation.
Over the past decade, deep learning systems have enjoyed tremendous success in the area of computer vision and natural language understanding. However, these systems still struggle in tasks which require deliberate thinking and reasoning process. Recently, many approaches to extend the capability of neural networks have emerged. This paper reviews several recent approaches which take inspiration from human’s cognitive process and symbolic view of artificial intelligence to solve reasoning tasks, and outlines future research directions in this challenging problem.