Musical Instruments Sound Classification

Convolutional Neural Network: Convolutional Neural Network: As we all are aware of the fact, how deep learning and transfer learning is revolutionizing the world with its immense capability of handling any kind of data and learning so efficiently. So, similarly we have applied the same concept by picking a deep learning model i.e., Convolutional neural network which basically work son the principle of having filters. Each convolutional layer has some specific filters to identify and extract the features from the input image and learn it and transfer it to other layers for further processing. We can have as many filters as possible in the convolutional layer depending on the data we are dealing on. Filter are nothing but feature detectors in the input data. Along with the convolutional layer we also have other layers which does further pre-processing such as Maxpooling, Activation function, Batch Normalization and dropout layer. These all contribute to the CNN model creation and along with the flatten and output layer. The reason we do flattening is to feed the output of the CNN model to the dense layer which gives us the probability of the predicted value.

References:

https://www.ibm.com/cloud/learn/convolutional-neural-networks

Long Term Short Memory:  If you want to predict the next word for the sentence “I like numbers, my favourite subject is …” your answer would be “mathematics”. How do you come to that conclusion; it is because of the word “numbers”? Recurrent Neural Networks help us achieve this. They are neural networks with loops, allowing the information to stay for some time.

Figure 1: Recurrent Neural Network

Xt is the input, the Network analyses and throws out output ht. RNNs have performed decently but the problem comes when the sentence is too long and has immense data. E.g., predict the next word for the sentence “I like numbers, my favourite subject is mathematics. My brother likes space and technology, his favourite subject is ….” your answer would be “astronomy”. How do you come to that conclusion, it is because of the word “space and technology”? But how does the network know that word mathematics is not important now and “space and technology is important now”. Also, the problem with RNNs is it cannot remember the entire sequence for long. Long Short-Term Networks (LSTMs) are a special kind of RNN which are very much capable of handling long term dependencies. LSTMs has a cell state which holds the important word/information required for processing. The information in the cell state can be forgotten (removed) or added based on gates. It has 3 gates to help it decide:

1. Forget gate:To forget the information in the cell state. Done with the help of a sigmoid function which returns a value between 0 and 1. If 1 or closer to 1, remove the word else retain the word.  
2. Input gate:What input should be given to the cell state done with the help of tan-h function.
3. Update gate:Any information that can be added to the cell state. E.g. : When “space” was in cell state, update it with the word “technology” because word “technology” is important in decision making as well .

References:

https://colah.github.io/posts/2015-08-Understanding-LSTMs/