Cognition has bound everyone since long, our journey of cognition deals with a fun team. We are a small team (until we combine forces with Dr. Eric Yttri, at CMU for decision and behavioural experiments on Brain computer based interface for neuromorphic circuits on mice). I've been thinking about where we initialize the network with a connectivity that we know solves part of the problem and then see what/how connections are added to solve the remaining part.Allowing two different 4 neuron sets of memory pieces interact with each other that is rather than an elman rnn we watch interaction between neurons passing sensory input through each other With professor Christopher Cueva at MIT brain and cognitive science, we are trying tolook into the mechanism of how brain stores sensory information and how does it lead to decision making on multiple continuous variables.
200 test trials, 100 times steps, 1500 paramter updates, error 0.006%
Input chasing target output
Firing rate of RNN unit 4 vs time
Mean squared error during training vs Number of parameter updates
Firing of 50 hidden units vs time
Mean squared error during training vs parameter updates
Firing of 50 hidden units with time when trials increased to 200