search

tcstewar / 2015-Embodied_Benchmarks

Paper on Embodied Neuromorphic Benchmarks
GNU General Public License v2.0
0 stars 0 forks source link

L377 #12

Closed studywolf closed 8 years ago

studywolf commented 8 years ago

https://github.com/tcstewar/2015-Embodied_Benchmarks/blob/master/paper/paper.tex#L377

Adding in q_d to the input:

This then suggests an explicit neuromorphic benchmark.
The input to the neuromorphic hardware is $[q, q_d]$, the system state and desired state.
This input is fed to each neuron such that each neuron produces some output behaviour that is based on this input.
Since $[q, q_d]$ is multi-dimensional, we may give each neuron a random weighting of each $[q, q_d]$ value ($J_i=e_i \cdot [q, q_d]$, where $J_i$ is the input to neuron $i$, and $e_i$ is a randomly chosen vector\footnote{$e$ could also be chosen so as to regularly span the space of possibilities}). Given this input, the neurons will produce some output $A$. We now form a weighted sum of these outputs $Ad$, where $d$ is a matrix (number of neurons by number of elements in $[q, q_d]$) that is initially all zeros.

studywolf commented 8 years ago

wait nope, this is wrong. nevermind! just q for the input is right. Is the learning / u input considered post-hardware?

studywolf commented 8 years ago

answered in the next paragraph! haha

tcstewar commented 8 years ago

Ah, oops, my mistake again. q_d is an input to the controller, but I don't currently feed q_d into the adaptive part....

studywolf commented 8 years ago

i was just bein' silly issall!