Hebbian Learning in Perforant Pathway

[abdel]

• Moved is_hebbian_perforant to the inside of AHA to fix the issue with self call
• Created local optimizer LocalOptim to facilitate manual weight updates, in a PyTorch-y way
• Created learning_rules.py to hold all the relevant 'local learning rules', e.g. Oja, Hebb
  • Initially, I found an implementation of Oja rule (see OjaRule) and utilised that as basis for the second learning rule
  • The Oja rule was adapted into PureHebb as the 'pure Hebbian learning' rule is described in the Ketz2013
  • The initial implementation was very inefficient especially dealing with large weights matrix (EC:CA3) and was re-written in a vectorized way which significantly increased the performance
• Extended nn.Linear as the LocalConnection class to allow us to default to untrainable weights and better control the initialisation
  • We can then instantiate both dg_ca3 and ec_ca3 using LocalConnection, and utilise the new LocalOptim to control weight updates for each set of connections
  • The outputs of each connection is then combined into pc_out
  • The weight update occurs using pc_out as the 'target' with respect to the corresponding connection input (e.g. dg or ec_in
• Added a new MSE metric to measure the difference between pre and post synaptic activity: loss/dg_ca3, loss/ec_ca3 and loss/pc
• Hid the existing PC buffer and PR implementations behind the negation of the is_hebbian_perforant flag
• Added additional checks in lake to ensure it is not 'assuming' that PR exists
• Created new aha_config_hebb.json with increased study steps and removal of PR metrics

[affogato]

But it is required

[affogato]

i see, it would have the actual value 'required'. don't fully understand how that works, but it's clear that it wasn't a bug

On Wed, 14 Jul 2021 at 17:41, Gideon Kowadlo @.***> wrote:

But it is required

-- Gideon Kowadlo http://gideon.kowadlo.net