November, what has been done, and what to do next

The first half of the November, I spent time on testing the circuit, filtering, mounting techniques for embroidered circuit… and then built the 12 neuron forward propagation embroidered ANN >> https://www.kobakant.at/false-lies/attiny-neural-network-version2/

in the mean time, I had an exchange with Yuansong from CONFIRM group, who is an expert on ANN and machine learning. he confirmed my attempt of ANN simulation with python is in fact correct. (yey!) so this made me go ahead with the embroidered ANN I mentioned above, as I used the python simulation value to give all the weight and bias data. He also consulted some of my burning questions about how to recognize series of actions, movements and suggested me to look into RNN (Recurrent Neural Network). He suggested LSTM (Long Short-Term Memory) and GRU (Gated Recurrent Units).

on 15th of Nov, Milan from CONFIRM visited the studio and we looked at the freshly made embroidered ANN together. Yes, it can recognize 3 hand gestures (rock, paper, scissors) smoothly. The outcome light comes out in slight delay (which I do not know why yet) but even this delay feels good as if the network is “thinking”. We also started to look into RNN together. Yuansong gave me a link to a tutorial, but it was a bit too advance for me… so at the moment I am looking into this one >>http://karpathy.github.io/2015/05/21/rnn-effectiveness/

If I could understand this, and instead of alphabets to use sensor input value and the network could generate the sequence of movements after learning my movements, just like this example of Shakespeare text generator, will be great.

Now, I have my new to do lists:

  • check if one can use one pin on ATTINY as both TX and RX for software serial, or if one can use D5/A0 that is connected to reset pin for analog read or software serial.
  • completing the first garment design for pattern development. Ben (pattern maker) is scheduled to work with me in the beginning of December to work on pattern/ toile.