ML engineer with rich experience

We wish to make a sensorimotor autoencoder: That is a neural network that condenses the information found within a sensorimotor feedback loop. We lack the technical expertise to fully evaluate and certainly to execute the idea but can attempt to communicate the idea thusly: Start with a fully connected RNN. Arrange the neurons in layers and remove neurons from inner layers to produce the autoencoder shape. Remove all connections to other cells in the same layer, including the self-connection. This will force connection into the horizontal plane - every node in a layer is connected to all other nodes in all other layers left and right, but are not connected to any notes in it’s own layer - up and down, including itself. Initiate these connections, these weights with a decreasing bias the further away a neuron is. Neurons in the 3rd layer, for example, should naturally have strong connections to neurons in the 2nd and 4th layers, but weaker connections to the 1st and 5th layers, etc. Each timestep certain cells in each layer will be active. Force a constant, small percentage of the cells to be on. In other words, the activation function must take the entire layer into account so that the top 4% of activated neurons actually fire. Be sure to treat the backwards motion of data (that coming from the Motor neurons) as a feedforward network. In other words, The sensory network trains the motor network and the motor network trains the sensory network in an online fashion. Backpropagation should not be used. Other optimizations are possible, but this is what we think of as the simplest sensorimotor autoencoder. We need help evaluating and building this kind of network. Of course development and testing should be done in an interactive environment so that the actor controlled by this network is in a feedback loop with the environment. Many common actor-testing environments are not suitable for this design as they typically have a few actions the actor can take, whereas with many motor neurons the actor should ideally have many actuators or muscle fibers to control.