Neural networks are a broad topic. But this small example demonstrates how to create a primary neural sensor that takes resistive readings from multiple sensors and multiply it by weight factor and then sum the results. Results are compared to a three-level threshold. Without going too deep into neural networks, we can say that neural cell thresholds are similar to natural biological neurons. For instance, pain levels: itch is a low pain level while burning sensations are combined with cold and warm feelings. Neural sensors can operate in the same way.
Let’s take typical neural sensors, which consist of two inputs with some weights and three outputs. Depending on the threshold level that the sum of inputs gives – we have an output signal on three outputs.
Let us build a real-world neuron with two inputs and three outputs. As a sensor, I will use potentiometers, which can be replaced by photocells or temperature sensors. According to these two sensor readings, a neuron can have three states:
This could be like low light, medium light, high light, o low temp, etc. If you decide to put a neuron to your robot, this could be used for solar cell charging. If it is dark, go to sleep mode; if it is medium-light, go to the brightest area; if it is high light level, stop and start charging.
Each decision is indicated by one of three LEDs.
The above schematic is a simple neuron sensor simulator made of Atmega8 MCU couple potentiometers and three LEDs. Let us write a simple routine to simulate this embedded neuron. For simulation, I used the VMLAB simulator.
In the example, you may see that one input has a weight of 2 and the second of 3 units — the exit selected by comparing w value with thresholds.
Simulation results are here:
This way, using simple neural nets, you can build more complex and smarter systems.
Download VMLAB neuron demo project.