It has been a long time since my last post – mainly because I’ve been flat-out busy with exciting stuff…
The start-up I got involved with during Dublin Start-Up Weekend back in July has gained momentum, and we’ve been busy developing our business plan and prototype system; and have been taking part in the NDRC’s pre-accelerator programme (http://www.ndrc.ie/news-events/news/12-teams-selected-for-ndrcs-autumn-pre-accelerator-programme)
Today I feel like I need to write something, as I’ve assembled my first ever Deep-Learning based application, and I’m a little bit freaked out (and pleased) with the result. One of the key components to the product I’m working on is an audio track comparison mechanism, to allow the system to find similar tracks to an initial ‘seed’ track.
I’ve used a Keras-based neural network to analyze tracks and produce matching metrics, and then wrapped this in some playlist generation logic, that takes a seed track and then returns the 5 closest matches.
The whole system is still very much an initial prototype, and there are lots of improvements we will need to do; however, the initial results are still eerily more accurate than I would have hoped. Sometimes the matches it is coming up with may seem strange, but then on further consideration I can see that there are notable similarities. What is strange to me is that I have no understanding of the criteria the matcher used, just that what it has come up with kinda makes sense.
As I get more familiar with machine learning I’m sure this sense of wonder will be replaced by understanding; but at the moment I am feeling a significant sense of wonder!