- Researchers have created a new type of artificial brain
- It works similarly to the human one
- Can learn “on the go” recognizes numbers or remembers them
Scientists have created a new type of artificial brain that can learn and remember things on the fly, just like humans do. This could lead to more innovative and greener machines that can handle complex and changing real-world data. The research, published in the journal Nature Communications, is a joint work of experts from the University of Sydney and the University of California, Los Angeles.
The artificial brain is made up of nanowire networks, tiny wires billions of times smaller than one meter. These wires create random patterns that look like something from the Mikado game and act like neural networks in our brain.
It works like the human one
They can learn and remember using simple rules that change the electrical resistance where the wires cross. That is called “resistive memory switching,‘ which occurs when cables become more or less conductive depending on the electrical signals they receive. In a similar way, our brain cells, neurons, communicate with each other through synapses.
The researchers tested the nanowire networks on two tasks: recognizing and remembering pictures and recalling sequences of numbers. They used electrical impulses to represent images and numbers, taking inspiration from how our brain processes information.
The leader of the research team, Professor Zdenka Kuncic, explained that memorizing a sequence of numbers is like memorizing a telephone number. The networks also performed well at recognizing images from a database of 70,000 handwritten digits, a common task for machine learning.
It has no problem receiving new data all the time
In her previous research, these networks were only able to remember simple tasks, but their latest version has no problem with dynamic data coming online. According to her, this is a significant advance because previously it was difficult to perform learning with large and changing data sets.
The usual way for machine learning was to first store the data and then train a model on it, but this consumed too much energy for practical use. However, the new approach allowed the nanowire networks to learn and remember “for march,” one data point at a time without storing them or consuming too much power.
According to Ruomin Zhu, a PhD student at the University of Sydney, there are other benefits to online learning. If data were to come in continuously, for example from sensors, machine learning based on artificial neural networks would have to be able to adapt in real time, which it cannot do now.