Neural Networks in Robotics
https://eandt.theiet.org/content/articles/2016/11/robots-learning-to-adapt-and-thrive-when-faced-with-new-challenges/
A lot of research in computer architecture and related areas is tending towards higher levels of abstractions – that is, making use of tools and automation techniques in order for engineers to tackle larger problems. The above article, published today, explains why neural networks have emerged as a powerful tool in making these types of abstractions. Neural networks represent a way to design an AI which parallels the human brain. Each node represents a neuron, while the connections between neurons (the edges in the neural network graph) represent the links between neurons. That being said, the article elaborates on the fact that it can be extremely difficult to understand why neural networks evolve in the ways they do. Especially for applications in robotics, understanding how the results of the implementation correspond to the decisions on part of the neural network is critical. Perhaps new network structures of neural networks (besides simply adding more stages of neurons) can help establish this connection more strongly.
The article ties in directly with ideas presented in class. Larger structures in networks often provide additional insight, but at the cost of increased complexity. For instance, looking at the Internet as a network in the year 2016 is much more difficult to analyze than it was in its infancy, but now provides hugely different insights. Similarly, applying deeper neural networks through the recent concept of deep learning will certainly enable computer systems to process more difficult tasks, but can only come from a more complete understanding of neural network theory. The article mentions how GPU’s can be reprogrammed to implement neural networks, and it is incredibly exciting to see where these more advanced neural networks will begin to appear.