Research Ideas:
- Neural Turing Machines
- Interpolating between two LSTM states
Research Tool Feedback:
-
Side by side, up to 5
-
Put all questions on the same page
-
Music Background solicitation before page
-
Amount of time spent on metric
-
Buy domain name (10 quid), AWS billing limit (20 quid)
Visitor agreement
- Does not apply to our work
- For when we use MSFT, read private poster, and decide to do something off of it
Feynman's repo will be repo we collaborate on.
Weekly 1-1s
- What did last week
- Plan for next week
- Bring up questions, get connected to Matt's network
Today's agenda:
- Evaluation: how do we know what we're doing is working
- First steps into the project
Baseline: log probability
Subjective evaluation: completion of a given composition
- Biases towards copying rather than creative generation
Train on untransposed, model transposed
Do we want to pursue this metric learning problem? YES
- AI(Mark): will send us a list of statistics used for evaluating chorales
- AI(Feynman+Marcin): Do we have metadata partitioning the piece up into certain key signature blocks?
- Is there a corpus of Bach pastiches (Bach-like data)? Can use this for training.
Bach transcription ==> Statistics defined by Mark ==> Generative model over statistics features (probability for Bach-like)
- Augmenting training data:
- Transpositions to the training set? Goal is for LSTM to not care about
Evaluation statistics:
- Mark: propose statistics
- Feynman+Marcin: implement in Python
- Construct factor graph for generative model on statistics
- Bring in to Microsoft, factor graph model evaluation in Infer.network
Single voice melody generation:
- Extract soprano lines
- Simple LSTM
Single voice melody generation
- Extract melody (soprano)) lines from Bach fugues
- Train 1-voice LSTM on
- Could extend to use bidirectional LSTM
- How to generate time/key signature? Sample before and preprocess metadata?
Chorale harmonization given the melody
- Generate single voice melody given first model
- Use the melody as an input, output the other voices
- Connectivity structure: current voice depends on history of current voice only and all voices at current time only