A Bristol based tech-startup, Graphcore, has come up with a variety of “AI brain scans” using their development chip and AI software to produce images that clearly indicate how AI systems think and perform tasks. These images of their Intelligent Processing Unit, while functioning, look just like that of a scan of a normal human brain.
Machine learning systems go through a series of computational graphs from the start to the end of a particular process. This means that a machine learning program starts with a phase “Construction phase” when to perform a task. Here the graph indicates the computations required for that task. Next is the execution phase where through computations, the machine acquires resources that are provided at the construction phase to run a task. During execution, the machine works through “passes” which are performed through backward and forward movements on the data. Graphcore’s images indicate these movements and their connections with different colors, showing a similarity to the compilations of the brain scans.
During an interview with wired Graphcore’s CEO, Nigel Toon, mentioned that there is little or no difference between the highlights from the scans of an actual brain and the AI brain. This indicates that both the human brain and the intelligence processing unit do the same thing. A human brain has neurons which form connections through synapses. This kind of connection is very similar to machine learning.
Graphcore said in a blog post that the chip that was used in creating the images will be fully available in 2017 and they have also developed an IPU which is currently the most successful AI system that exists, as per their claim. They also said they are in the process of developing a much superior GPU than that of Google and NVIDIA.