A stochastic photo-responsive memristive neuron for an in-sensor visual system based on a restricted Boltzmann...

Area of Science:

• Neuromorphic Engineering
• Materials Science
• Computer Science

Background:

• Conventional sensory systems face von Neumann computing bottlenecks.
• In-sensor computing offers a solution by extracting information directly within sensor elements.
• Restricted Boltzmann Machines (RBMs) are effective for feature extraction in complex data.

Purpose of the Study:

• To develop a stochastic photo-responsive neuron for in-sensor Restricted Boltzmann Machines (RBMs).
• To leverage memristor technology for advanced in-sensor computing capabilities.
• To improve image recognition accuracy and efficiency in unpredictable conditions.

Main Methods:

• Development of a stochastic photo-responsive neuron using TiN/In-Ga-Zn-O/TiN and Ag/HfO2/Pt memristors.
• Configuration of the neuron as an input unit within an in-sensor RBM.
• Utilizing semi-empirical simulations to evaluate performance on image datasets.

Main Results:

• The developed neuron exhibits sigmoidal switching probability dependent on light intensity.
• High recognition accuracies of 90.9% (handwritten digits) and 95.5% (face images) were achieved.
• The in-sensor RBM demonstrated effective reconstruction of abnormal face images.

Conclusions:

• Integrating in-sensor computing with probabilistic neural networks (like RBMs) enhances image recognition reliability.
• The stochastic neuron design effectively extracts significant features from complex images.
• This approach shows promise for efficient and robust image recognition in real-world scenarios.