Hi I am currently trying to set up an simple neural net test function to compare the efficiency of alternative high-dimesional Bayesian optimization algorithms. (I know such approaches don't generally scale to high-dimensions, but that is why is it a current research topic of mine!)
Essentially, I would like to instantiate a nerual net, however, rather than training the net using the native matlab training methods, I would like to assign weights to every bias/layer connection and then use a matlab performance function to assess the performace of that set of weights. Based upon that weight/performance input/out relationship I would then use my own coded Bayesian optimization algorithms to re-generate and set a new set of weights and then use the matlab performance function to assess the performance of that new set of weights. Please note that I must use these Bayesian optimizers that I have coded as this is the entire purpose of the experiment at hand; I would really like to avoid using the Bayesian optimizers to generate inital weights from which the net is then trainined as this will only increase the computational needs and ultimately deteriorate the scalability of this test function (as I hope to use this test schematic for much larger and extensive testing in the future).
For now, the goal is to create a net with 500 weights (biases and layer connections) that must be assigned, which I think could be done with a fairly shallow fully connected 4 or 5 layer feed forward net with around 10 nodes per layer? Such that I have a 500 dimensional optimization problem to be solved by my Bayesian optimizers. Currently I am looking into using the breast_cancer example data from Matlab which has a 9x699 input and 2x699 target.
In my mind this should be pretty straight forward, but I know that the large array of capabilities/funcitonality of the matlab nets could make such a task more difficult than it seems on the surface. Hopefully someone familiar with the functionality can help whip up a quick answer to hopefully a simple question?