Source-linked AI summary
Experimental demonstration of associative memory with memristive neural networks
Yuriy V. Pershin, Massimiliano Di Ventra
TL;DR
Electronic associative-memory networks lack a satisfactory synapse representation because synapses must be history-dependent, analog, and plastic. The paper builds inexpensive programmable memristor-emulator synapses and electronic neurons, then experimentally demonstrates associative memory in a three-neuron network. After simultaneous training, either input causes the output neuron to fire.
Problem
Electronic associative-memory networks have lacked synapses that are plastic, history-dependent, and capable of storing continuous analog states.
Method
The paper develops flexible electronic neurons and a programmable memristor emulator, then combines them in a three-neuron network with two synapses.
Results
The network experimentally develops an association: after simultaneous input stimulation, either input causes the output neuron to fire.
Takeaways & Limitations
Inexpensive, flexible electronic neurons and memristive synapses can reproduce associative memory in a simple neural network.
Abstract
from arXiv · showhide
When someone mentions the name of a known person we immediately recall her face and possibly many other traits. This is because we possess the so-called associative memory, that is the ability to correlate different memories to the same fact or event. Associative memory is such a fundamental and encompassing human ability (and not just human) that the network of neurons in our brain must perform it quite easily. The question is then whether electronic neural networks (electronic schemes that act somewhat similarly to human brains) can be built to perform this type of function. Although the field of neural networks has developed for many years, a key element, namely the synapses between adjacent neurons, has been lacking a satisfactory electronic representation. The reason for this is that a passive circuit element able to reproduce the synapse behaviour needs to remember its past dynamical history, store a continuous set of states, and be "plastic" according to the pre-synaptic and post-synaptic neuronal activity. Here we show that all this can be accomplished by a memory-resistor (memristor for short). In particular, by using simple and inexpensive off-the-shelf components we have built a memristor emulator which realizes all required synaptic properties. Most importantly, we have demonstrated experimentally the formation of associative memory in a simple neural network consisting of three electronic neurons connected by two memristor-emulator synapses. This experimental demonstration opens up new possibilities in the understanding of neural processes using memory devices, an important step forward to reproduce complex learning, adaptive and spontaneous behaviour with electronic neural networks.
I. INTRODUCTION
Associative memory is difficult to realize electronically because synapses must be plastic, history-dependent, and capable of storing continuous analog states. The paper addresses this gap with memristive synapses and demonstrates associative memory in a simple three-neuron network.
- Electronic associative-memory networks require synapses whose strength changes with signals, depends on dynamical history, and stores continuous values.
- Existing electronic synapse implementations often use many circuit elements, occupying substantial VLSI area and limiting synapse density.
- Memristors offer a promising synapse realization because their response depends on dynamical history and their size can be as small as 30×30×2 nm3.
- The paper develops inexpensive, flexible electronic neurons and memristor-based synapses that can be tuned to reproduce biological-cell functions.
- A three-neuron network uses two synapses to associate initially separate input events through simultaneous stimulation and Hebbian learning.
A. Electronic neuron
The electronic neuron converts thresholded input voltages into forward- and backward-propagating pulses. Above threshold, stronger inputs produce more closely spaced pulses.
- The neuron continuously monitors input voltage and fires only when it exceeds threshold VT.In the experiments, VT = 1.5V.
- The implementation uses an analog-to-digital converter and microcontroller, requiring only two additional 10k resistors in the reported realization.
- Each firing event generates forward and backward pulses with constant amplitude, while pulse separation varies with input amplitude.The implemented pulse amplitude was 2.5V.
- When input voltage exceeds threshold, pulse separation decreases as input amplitude increases.
B. Electronic synapse
The electronic synapse is a programmable memristor emulator that continuously updates resistance from measured voltage and internal-state dynamics. Its measured current–voltage response exhibits characteristic memristive behavior.
- The emulator can reproduce voltage- or current-controlled memristive systems through programmable state and response functions.
- Its digital potentiometer, ADC, and microcontroller continuously measure voltage and update the programmed resistance.
- The threshold model changes memristance at different rates below and above threshold voltage, while limiting resistance between R1 and R2.
- The emulator produces pinched hysteresis I–V loops through the origin, a characteristic used to demonstrate memristive behavior.The reported loops also show frequency dependence.
C. Associative memory
The three-neuron circuit learns an association by changing the initially weak synapse during simultaneous stimulation. After learning, either input triggers the output neuron, demonstrating associative memory.
- The network initially has one strong low-resistance synapse and one weak high-resistance synapse, so only the first input triggers the output.
- During simultaneous stimulation, overlapping forward and backward pulses create a high voltage across the weak synapse and drive its resistance downward.
- After learning, stimulation of either the original or paired input causes the output neuron to fire.
- In the demonstrated network, synapse resistance can only decrease because the applied memristor voltage is nonpositive.The paper notes that depression could be added with a positive offset or bias and nonzero α.
III. CONCLUSION
The authors show that electronic memristive neurons and synapses can reproduce associative memory when combined in a neural network. Their inexpensive, flexible schemes could support studies of more complex adaptive neural networks and learning processes.
- Electronic memristive synapses and neurons reproduce important functionalities of their biological counterparts.
- A network combining these components gives rise to associative memory, an important brain function.
- Off-the-shelf inexpensive components make the proposed electronic neurons and synapses accessible for laboratory construction.
- Their flexibility supports investigations of more complex neural networks that adapt to incoming signals and make decisions based on correlations between memories.
- The platform opens possibilities for reproducing different types of learning and more complex neural processes.