Farah Baracat

Day-to-day

Friday 20.10.2023

• Looked into OpenSim (version 4.4) (following Arnault tutorial): attempt to load the ARM model and a motion file created from Neuromotion.
  • GUI is not working, I don’t see anything. Tried version 4.3 –> still same issue
• Moved to coding: managed to understand a bit how to simulate MU spike trains and a pool of motor neuron pool
• Switched to SNN and iEMG:
  • fixed the plotting in case of splitting the rep into different bins (plot all bins together)
  • fixed/revised issue with snn data preparation (wasn’t relying on the correct MU id probably)
  • attempt to train on rep 1 and test on rep 2 for a single finger. Looked too promising to be true…will have to look and debug more.

Monday 23.10.2023

[IEMG project]

• Cleaned code/ merged changes. Tested trainning on rep 1 and testing on rep2 for thumb alone, thumb-index (kind of works), thumb-index-middle (doesn’t work), all (not working).
• Attempted to train (and correctly plot) with a batch greater than 1. It doesn’t work in this case of time series. I am not sure why exactly.
• Agnese suggested we use percentage mvc to have a bigger loss –> good idea!
• A new branch removing hidden layer…didn’t test much yet.

Tuesday 24.10.2023

• More testing with a single layer network: managed to make it fit on a thumb
• Refactor notebooks into a python script to start optimization with wandb
  • setup sweep for tau syn, learn the tau_mem
    • Sweep for thumb
    • Sweep for index
    • Sweep for middle
    • Sweep for ring
    • Sweep for little
  • added a class for filtering the output vmem using a single leaky neuron: vmem is injected as current.

Wednesday 25.10.2023

• Documentation of the current stat of the network
  • Summarize resuls of the individual sweeps
• Run sweep for all fingers together (2 different weight initialization ranges)

Thursday 26.10.2023

• What is MU angle within a muscle?
• Got through part of the paper on Fuglevand’s model to understansd the neuron parameters
• More sweeps with sclaing the loss function

Monday 30.10.2023

• Sweep rep2 on rep1
  • changing the weight initialization:
    • Uniform 0-0.1
    • Uniform 0-0.05.
• Checkout a new brnach for adding varaibility on the time constants

Friday 3.11.2023

• Problem: can we really train the network on 2 sec windows? is this practical? how do we make the predictions then? We want to see the ouput of the network every 200 -250 ms. Does it matter if we are training on longer times?

• Ran sweeps on SNNMonster: it was a nightmare to setup tmux…had problems with the imports then with tmux running with conda…but finally found a fix.

• “Relevance of the spiking model to motor system” might be a good section to add