Non-spiking Neural Network with LMU in Nengo

Hello Good People,

Greetings! I am new in Nengo. I just wanted to know is it possible to implement a time-series forecasting of a musical tone using Non-Spiking Neural Networks with LMU implementation? Thanks.

Hi @MNHaque and welcome to the Nengo forums! :smiley:

As to your question, the LMU has been demonstrated to work particularly well with time-series data, and has been shown to match or outperform LSTMs in equivalent tasks, so I would think that it would be applicable to the task you are inquiring about. :slight_smile:

Hello Sir,

Thank you so much for your reply. Could you please provide me with any referral source code to test a Time series music tone with LMU? That would be highly appreciated.

The application of the LMU (i.e., exactly how to tune the LMU) is very much application specific, and unfortunately, we don’t have many examples. The best I can do is point you to the generic image processing example, which you might be able to adapt to use in your audio task.

In that example, the image is modified to be a stream of pixels (which is sort of similar to a stream of values that a music tone would be), so the major thing you’d need to change is the parameters (units, order, and theta) of the LMU itself. Documentation for these parameters can be found here, and I recommend you read the LMU paper (which I linked in my previous post) to get a better insight as to how those parameters would relate to your audio task.