Artificial intelligence has advanced to the point where machines can learn – but in simple ways. To date, the technology has learned patterns in their algorithm which has allowed it to expand their ability to be more sufficient. Recognizing people's faces, identifying vocal patterns associated with people, reading medical documents and they can even carry out their own conversation.
However, their ability to carry out conversations had to be worked on a bit further due to exploits in the design of the technology. Users had found out that, during the early stages of the technology's release, as a chat-bot, it would only repeat anything that users sent back to them. This would eventually become a problem because the text received from the users would be offensive, racist and have some crude language associated with it. Which went against its intended use, so the creators had to shut it down before it got any worse and in turn, they decided to improve on it instead.
These are all done through something called neural networks - a complex computer algorithm that analyzes a large amount of data which in turn, allows it to learn about those things it has received. However, there always seems to be a problem with this implementation as sometimes the technology does not come up with an accurate answer and researchers have trouble finding out why or how this happens.
Image representing how the neural network works. (Photo credit: The Building Blocks of Interpretability)
Google's team took the initiative to address and research this issue more promptly to find out why the system behaves the way it does. New research supports their findings and gives in-depth analysis of each feature/visualization that ideally supports the way this technology works and why/how it comes up with their conclusions.
More researchers seem to have developed a high interest in the technology and are in the process of developing/implementing ways to better understand their networks. It is also of high importance to try to understand their methods as they make more decisions made by humans.
Ideally, this technology is only meant to imitate the neurons in the human brain and correspond them with the ability to find patterns in image processing, such as recognizing an image and the objects. Each neuron associated with the system represents a mathematical algorithm, and that allows the system to come up with a complete choice in the end.
How the neurons work to arrive at a conclusion of an image by identifying parts of the image. (Image credit: The Building Blocks of Interpretability)
The way the neuron works is by identifying certain attributes in a photo, such as the curve of a shape or anything that gives measurement to what a shape is in the photo. The end result is that Google wants to provide a way to identify those objects in the photo, giving it an effect of what it actually represents. Those objects can be combined and compared with other photos to give it a fully responsive and accurate answer to what it actually is identifying. Objects can be anything as long as the algorithm matches the way it's visualized.
This type of work is still in the early stages of development, but researchers are hopeful that it'll help to provide answers as to why the system makes mistakes the way it does and how often. Perhaps there will also be some idea as to how they can improve the technology and enhance it in ways that allow them to prevent the system from making the same mistakes.
In comparison, the more complex the system gets, the more difficult it will be to understand why and how the system makes those mistakes. It would almost be like understanding how the human mind works in most instances.
Have a story tip? Message me at: cabe(at)element14(dot)com