The second prototype is a half shirt with ANN sleeve. This ANN is capable of communicating back propagation values, so the wearer can teach the sleeve.
The video shows the gesture test with the glove. It reacts quite quickly with the glove’s sensor inputs on the input layer. The ANN on this sleeve has 2 hidden layers and 1 output neuron to classify if the gesture is 0 or 1. The classification result comes still with big delay, which I do not know why it reacts so slow….
Here is the sketch of how the neurons are connected.
I tried out smd LED an LDR to reduce the bulkiness of the embroidery. it was not so easy to find the smd LDR… but I spotted it at Alibaba.
Here are some lengthy embroidery process images.
after the embroidery is completed, I needed to calibrate the readings of LDR. The ATTINY 85 chips does not have enough pins to use software serial for debugging, so I need to switch the program when debugging. It is a small detail but very annoying! I use FTDI to connect with the ATTINY.
To communicate between the neurons to pass the back propagation information, I had to hack the software serial as it does not leave enough pins on ATTINY. I used same pin for RX and TX, switching the pin to OUTPUT when writing out data to other neurons, then switch back to INPUT for reading mode. It is a bit buggy and had to slow down the serial communication speed to 2400. Surprisingly it works!
Here is a button button for telling the ANN the correct answer for the gesture. it is a mechanical push button mounted on a button. It works very well. When one press either of the button, it triggers the back propagation on the output layer, either yes (1) or no (0).