Perceptrons from memristors.

Area of Science:

• Neuromorphic Engineering
• Computational Neuroscience
• Materials Science

Background:

• Memristors are memory resistors used in neuromorphic architectures for synapses and non-volatile memory.
• Existing models typically use memristors for only one component (synapse or neuron), not both.
• A unified memristor-based approach for both neurons and synapses is lacking.

Purpose of the Study:

• To propose and model single and multilayer perceptrons built exclusively from memristors.
• To adapt existing learning algorithms for these novel memristor-based neural networks.
• To explore the potential of memristors as universal function approximators in neural networks.

Main Methods:

• Development of theoretical models for memristor-based single and multilayer perceptrons.
• Adaptation of the delta rule for single-layer perceptrons.
• Adaptation of the backpropagation algorithm for multilayer perceptrons.
• Validation of network performance against established perceptron criteria and theorems.

Main Results:

• Successfully modeled single and multilayer perceptrons using only memristors.
• Demonstrated that these memristor-based perceptrons perform as expected, including satisfying Minsky-Papert's theorem.
• Confirmed memristors' capability as universal function approximators, a consequence of the Universal Approximation Theorem.
• Showcased the potential for energy-efficient neural network architectures.

Conclusions:

• The proposed models enable the creation of neural networks entirely from memristors.
• These networks offer a pathway to highly energy-efficient neuromorphic computing.
• The findings open possibilities for adapting classical and quantum learning systems to a memristor-based paradigm.