I believe building a backprop-learned MLP would be much better than using the perceptron learning algo. In my experience, perceptron learning works quite slow; happy to hear other thoughts.
On Wed, Feb 22, 2017 at 11:40 AM, Kazmi,Auon H <[email protected]> wrote: > Hi Xixuan, > > THanks! Is starting with single layer perceptron (and thus using > perceptron learning algo rather than SGD) a better idea? > > > > Auon > > ________________________________ > From: Feng, Xixuan (Aaron) <[email protected]> > Sent: Tuesday, February 21, 2017 9:33:02 PM > To: [email protected] > Subject: Re: Regarding new Perceptron module > > Hi Auon, > > I have made some effort to try implementing multi-layer perceptron a few > years back but never got the time to finish it. Latest commits on these 2 > branches has the history > https://github.com/haying/madlib/commits/nn_design > https://github.com/haying/madlib/commits/mlp > > I were using the convex programming framework in MADlib to implement it > with stochastic gradient descent. You may or may not do it in the same way. > > Thanks, > Feng, Xixuan (Aaron) > > On Wed, Feb 22, 2017 at 8:59 AM, Kazmi,Auon H <[email protected]> wrote: > > > Hi all, > > > > I am currently working on implementing Perceptron Learning Algorithm in > > Madlib. This module would serve the purpose of a single neuron when > > implementing any multi-layer neural network. > > > > Since, it involve state transitions, optimizations,etc., could you guys > > suggest a module that could help a little in this context? > > > > > > > > > > > > > > Regards, > > > > Auon > > > > > > -- > - Aaron >
