Summary: weekly mtg 20170524 (Byron, Steve, Matt, Aaron, Emily, me) - mobeets/nullSpaceControl GitHub Wiki

In this meeting I only have like 30 mins so we don't really discuss much.

• reschedule meeting?
• discussion outline
• neural tuning changes

Questions we could ask:

Train a network to control the Perturbation decoder using the Intuitive repertoire

• How good can cursor control be while still generating from the repertoire? (e.g., for OMPs)
• Can we predict the behavioral asymptote?
• Can we predict cursor trajectories?
• If not, what additional constraints are necessary? (e.g., dynamics, sensory delay, etc.)

Train two networks to control the Intuitive and Perturbation decoders using the Intuitive repertoire

• Can we predict tuning changes?

To do:

• remind Steve and others of outlining meeting tomorrow at 4:30pm
• send everyone email about best-mean problem
• send Emily and Matt manuscript

GAN idea:

• train the gen/disc networks to match intuitive activity
• then, with the gen network frozen, train an association network where the output_dim is the latent_dim, which selects when to select which neural activity.
  • i.e., let g = gen, f = association, h = decoder
  • then given goal velocity x, you want to optimize h(g(f(x))) with the loss being something about the cursor's error through the decoder
  • or maybe this can even help us optimize over the sequence of x's?
  • and then you can just swap out the outputs and retrain f for different decoders