Can I specify unit-to-unit connections in multi-layer network?

Hi there. I would like to build a network with structured connections among multiple hidden layers. Is there a way to specify unit-to-unit connections among hidden layers in a network? All I can seem to find is functions that support layer-level connection specifications.

As a hack I've created multiple single-unit hidden layers and connected them in a way that maps onto the structure I'm looking for. But, this runs really slow, and it doesn't seem to learn as well as a network with a single hidden layer (as a test I connected all single-unit hidden layers to the output, which performed worse than a single hidden layer with an equal number of units).

Any advice is greatly appreciated.

-- Jeff

Answer 1

[Greg Heath]

Yes, you can do it.

See the NNTBX documentation for custom networks.

But why ??

The MLP with a single hidden layer and continuous bounded activation functions is a universal approximator for piecewise continuous I/O relations.

Hope this helps.

Thank you for formally accepting my answer.

Greg

[jeffrey cockburn]

Thanks Greg.

I'm simulating neural structures to investigate information processing under various conditions. So yes, a single hidden layer can solve the problem, but I'm more interested in how information processing breaks down when various structures are knocked out.

In then end I created a single hidden layer, but manually manipulate the input and layer weights as needed. This seems to be working pretty well.

-- Jeff

[Ken Kueh]

Hi Jeff,

May I know how you manually manipulate weights and biases?

Thanks