Favorites.
This is probably one of the most intriguing projects I’ve ever gotten to participate in and this is because it deals directly with Machine Learning. We are making a Neural Network which can hopefully correctly determine the genre of music that we pass into it via audio file. This is of course a branch of AI and we have done many, many spikes into Neural Networks and how they are structured, how they work, etc. It hasn’t been without its difficulties, for example this past week: one teammate had their power out because of a natural disaster for 2 days, one had gotten sick and was out of commission, and another (me) ended up going to the emergency room on his doctor’s orders. Sounds like fun right?
This week was met with a lot of challenges clearly, but we were still able to get our work done and get it done in a timely manner. My teammates are fantastic and really do exceptional work. We all are learning something new and exciting, which drives us on more. We all want to learn about machine learning, and that is why each of us picked this project, so we are excited with every step even though the learning curve may feel less like a curve and more like a mountain. For example, and for the purposes of this post, my favorite technology used here has got to be the neural network. We are using a Convolutional Neural Network (CNN) because it is better at processing images. What do images have to do with audio you ask? Well we convert the audio into spectrogram images that have different features of the audio waveform presented. The more we process the audio, the more spectrograms we can get, and hopefully the more accurate of a model for determination will be.
The neural network is my favorite because it feels like you are peering into the mind of something, someone. It is made to model our own neural network we like to call, the brain. This is a little self evident in the name, but I don’t think we appreciate the mathematical complexity governing how just our simple neural networks operate instantaneously, versus what goes on in our brains. Backpropagation, Gradient Descent, Cost Functions… these are all essential mathematical concepts when it comes to deploying just the simplest of neural networks, and I find it fascinating. I’ve discovered through this project that the math behind what we do sometimes may be counterintuitive, but that is just because we are looking at the same problem from another perspective. For example, take y = mx + b, now we’ve all dealt with this equation before, its the equation for a line; but it doesn’t scream “line” when we just look at the equation right? In math we would call this the analytical approach, versus the graphical approach, but they mean the same thing. If I draw a (straight) line on a cartesian plane, it can be represented by that same equation, no matter where I draw it. In the same way, all of the mathematical components represent common sense things, for example, a cost function operates on your “output layer”. If you have 10 choices your machine can make, then your output layer will have 10 nodes, each representing one of those choices. Each node is given a number between 0 and 1, to represent the probability that the given choice is the correct one that the node represents. A cost function takes the squares of all of these numbers, and adds them. What you get is a sum, that we want to be as small as possible, ideally 1. Why is that?
Well, if you think about it, we want each output to be as correct as possible, so if there is only one right choice, then we want all the wrong choices to have a probability of 0 and the right one to have a probability of 1 (out of a range of 0 to 1). Just looking at what the cost function does mathematically seems obscure but when we understand its intention, it makes a lot of sense.
I love learning about how machines learn and I look forward to continuing our study in this via a project, I hope somewhere your curiosity is peaked, and you learn how to appease our coming robot overlords. I mean, I’m joking but, am I?
Leave a Reply