According to the latest papers published on Nature, Samsung and Harvard University researchers seem to have found a better brain-like chip development method, which is to learn from the original structure of the human brain. In the paper, they proposed a set of plans, hoping to “copy and paste” the brain’s neuron circuit diagram on the 3D neuromorphic chip network. This method needs to rely on the nanoelectrode array to enter a large number of neurons in order to record the connection position of the neurons and the strength of these connections. These data can be “copied” and then “pasted” to a 3D network composed of solid-state storage. The components used can be off-the-shelf flash memory or cutting-edge products such as variable resistance memory.
Each memory unit has a conductance corresponding to the strength of the neuron connection in the circuit diagram. In Samsung’s view, this method can effectively return to the direction of “reverse engineering of the brain” originally envisioned by scientists. This solution can be regarded as a “shortcut” for the development of brain-like AI systems. Operations such as flexible learning of new concepts and automatic adaptation to changing conditions are expected to be realized. According to researchers, you might even see purely autonomous machine systems with true cognitive capabilities.
However, considering that the human brain has approximately 100 billion neurons, the number of synaptic connections is a thousand times greater on this basis. In this way, an ideal neuro-type chip requires about 100 trillion memory units. Such a level of complexity is an extremely difficult challenge for any company at the moment (this does not include the resources required for related code development. ). The ideas of Samsung and Harvard can be said to point out a direction for the future development of this type of technology, but the theory may take a long, long time to become a reality.