Part of AGI experiments
Attractor Learning
Training a recurrent network to settle into its attractor state faster - LISSOM, but differentiable.
Imagine a network that settles its activity over time through lateral feedback connections, relaxing into a stable attractor pattern. Can it be trained to reach that settled state faster? This line explores exactly that - in the spirit of LISSOM's self-organizing lateral dynamics, but made fully differentiable so the standard deep-learning toolset (backprop, autograd) applies. The settling-time objective is an unusual target that ties recurrent dynamics to trainable structure.