Key parameters to improve artificial synaptic devices

The high power consumption disadvantage of the traditional von Neumann computing method was overcome with the development of neuromorphic computing systems that mimic the human brain.

Concept image of the item. Photo credit: Korea Institute of Science and Technology (KIST).

The implementation of a semiconductor device using a brain information transmission technique requires a high-performance analog artificial synapse device capable of expressing different synaptic connection strengths. This technique exploits signals sent when a neuron generates a spike signal.

Considering the traditional variable resistance memory devices, often used as artificial synapses, the electric field builds up as the filament grows with varying resistance, creating a feedback phenomenon that leads to rapid filament growth.

As a result, it’s difficult to build in much plasticity and maintain analogous (gradual) resistance variations across filament types.

The team of dr. YeonJoo Jeong at the Center for Neuromorphic Engineering at the Korea Institute of Science and Technology has overcome long-standing problems with analog synaptic properties, plasticity, and information retention in memristors and neuromorphic semiconductor devices.

He explained the creation of a synthetic synaptic semiconductor device capable of extremely reliable neuromorphic data processing.

The performance of current neuromorphic semiconductor devices has been hampered by a modest synaptic plasticity, which the KIST research team addressed by tailoring the redox properties of active electrode ions.

Additionally, several transition metals have been doped and used in the synaptic device to alter the probability that active electrode ions are reduced. The high reduction probability of ions has been shown to be a crucial factor in creating high-performance artificial synaptic devices.

The study team therefore added a titanium transition metal with a high ion reduction probability to an already existing artificial synaptic device.

This preserves the analog properties of the synapse and the plasticity of the device at the synapse of the biological brain, which is about five times the difference between high and low resistances.

The team also created a high-performance neuromorphic semiconductor that is about 50 times more effective.

Compared to the current artificial synapse device, information storage has been improved up to 63 times due to the high-level alloying reaction involving the doped titanium transition metal. More accurate simulations of certain brain processes, such as long-term depression and potentiation, could also be performed.

Using the artificial synaptic device they constructed, the team attempted to implement an artificial neural network learning pattern for image recognition. As a result, the error rate was reduced by more than 60% compared to the current artificial synaptic device.

In addition, the accuracy of handwriting image pattern recognition (MNIST) increased by more than 69%. By improving the artificial synaptic device, the research team demonstrated the feasibility of a high-performance neuromorphic computing system.

This study drastically improved synaptic range of motion and information storage, which were the major technical barriers of existing synaptic mimics. In the developed artificial synapse device, the analog operating range of the device has been maximized to express the different connection strengths of the synapse, so that the performance of brain simulation-based artificial intelligence is improved.

dr YeonJoo Jeong, Senior Research Scientist, Center for Neuromorphic Engineering, Korea Institute of Science and Technology

dr Jeong added: “In the follow-up research, we will manufacture a neuromorphic semiconductor chip on the basis of the developed artificial synapse device to realize a high-performance artificial intelligence system, thereby further improving the competitiveness in the field of domestic artificial intelligence systems and semiconductors.

magazine reference

Kan, J. et al. (2022) Cluster analog memristor by redox dynamics engineering for high performance neuromorphic computing. nature communication. doi:10.1038/s41467-022-31804-4.

Source: https://eng.kist.re.kr/eng/index.do

Laisser un commentaire

Votre adresse e-mail ne sera pas publiée.