Here I am outlining the ideas for a potential workshop paper for NIPS 2016.
Paper ideas:
1st paper
- compare FFN + index encoding vs. FFN(seq[:4] + seq[-4:]) vs. LSTM
- the effect of learning rates on all these architectures
2nd paper
- pimp LSTM so that it comes close to or outperforms FFNN.
3rd paper
- apply LSTM to MHC class II prediction
Things to include in the paper(s)
- a page of exposition on the data + problem
- the challenge of reducing short sequences into fixed input size models
- talk about how LSTM would be ideal, and why it is a candidate
- the importance of the learning rate
Some random ideas:
- view LSTM as encoding technique, get LSTM to mimick
kmer_index_encoding()to get comparabale result, then tweak it to beat it. - add attentional gate to LSTM
- consider deep LSTMs
- check what is going on with forget, and input gate.