Developments on neuronal probes and the levels of complexity in silicon ICs make it possible to map the synaptic connectivity of mammalian brains and paste the map into solid-state memory, the researchers claim saying it is a next path for semiconductor research.
Neuromorphic semiconductor research started decades ago with the idea of creating a close analogue of synaptic activity that could eventually be scaled up. However, practical computation moved to generalized digital representations of synaptic connectivity that could achieve useful scale in software simulations. These were neural networks run on computers in the 1990s that have more recently been implemented in hardware accelerators. Other developments that are loosely aligned to neuromorphic architectures include asynchronous, event-driven and in-memory processing.
Now engineers and scholars from Samsung and Harvard University have published a paper titled ‘Neuromorphic electronics based on copying and pasting the brain’ in Nature Electronics.
In the paper the engineers describe a return to more closely mimicking biological brains with the idea of creating a functional synaptic connectivity map of a mammalian neuronal network using neuroscience tools and then pasting this map onto a high-density three-dimensional network of solid-state memories. This would likely be an array of flash memories or resistive RAMs for reduced power consumption.
The researchers argue that the copy-and-paste approach would lead to silicon ICs that better mimic the brain and in terms of low power, facile learning, adaptation, and even autonomy and cognition.
A key part of the approach is the use of a nanoelectrode array, for mapping synaptic connections, that has been developed by Donhee Ham, Fellow of Samsung Advanced Institute of Technology (SAIT) and Professor of Harvard University, Professor Hongkun Park of Harvard University, two of the authors. Sungwoo Hwang, President and CEO of Samsung SDS and former Head of SAIT, and Kinam Kim, Vice Chairman and CEO of Samsung Electronics are the other co-corresponding authors.
The nanoelectrode array can monitor a large number of neurons so it can record their electrical signals with high sensitivity. These massively parallel intracellular recordings inform the neuronal wiring map, indicating where neurons connect with one another and how strong these connections are.
Typically, these neuronal probes have 1,028 spikes that act as sensors. Even with scaling up it may take a very large number of re-positionings of the probe to capture significant parts of the brain structure and it may not be practical to access some deeper parts of the brain.
Once captured the neuronal map can then be programmed into memory cells so that the conductance represents the strength of each neuronal connection in the copied map.
The authors propose that a network of specially-engineered non-volatile memories could be attached to the back end of the nanoelectrode array and could be directly driven by the recorded signals to create a direct download of parts of the brain’s neuronal connection map onto the memory chip.
The human brain has approximately 100 billion neurons and about 100 trillion synaptic connections. However three-dimensional memories are starting to approach that level of complexity and this could open up a new area of research for the semiconductor memory industry.
“The vision we present is highly ambitious,” said Ham, in a statement. “But working toward such a heroic goal will push the boundaries of machine intelligence, neuroscience, and semiconductor technology.”
Related links and articles: