- Researching how differing sets of training data effect the outcome (ex. changing genre, specific artists, etc.)
- In general, finding a way for us to predict/direct some qualities of the outputted songs
- Looking into creating a larger database of images/songs to work with
- Better results through looking at album artwork vs. single artwork (might have stronger correlation between songs released as a single, than just songs pulled from a full-length album)
- Researching how using different sizes/resolution of images effect the songs
- Researching how different lengths of songs effect the outcome of the generated music
- Would analyzing images using object identification as well as point-of-interest identifiction increase the correlation between songs and images?
Personal Goals:
- Understanding on a deeper level the code actually involved in making the LSTM (how it works)
- Being able to apply the general framework of this project to other situations
We have done some really cool stuff this year and I am really proud of everything the team has accomplished. That being said, we now have a great foundation for exploring new things and improving existing aspects of the project. Here is a summary:
- Diving deeper into neural networks
We have a functioning neural network and a good baseline understanding of how it works. However, there is room for improvement in understanding what's actually going on in the individual layers of the neural net.
- Experiment with different models
It would be cool to try out a few different models and seeing which performs best. It would also be cool to try some models more closely related to the original that compare sequences of notes with following ones.
- Transfer Learning
It would be cool to experiment with transfer learning and see how we can use it in this context!