Stuff I did today:
1. Found the bug in beam search that I previously suspected. The issue was my heuristic wasn’t doing quite what I wanted it to do. I fixed that and re-ran L-BFGS; I’m currently waiting for it to finish.
2. Attended our group’s lab meeting, which was a NIPS retrospective + talk from Simon, who is still visiting Stanford.
3. Met with Simon to discuss my research project. He suggests using stochastic gradient descent instead of L-BFGS, which I will try later for comparison.
This doesn’t look like terribly much, and probably honestly isn’t, although my excuse is I now think my code is reasonably likely to find the correct alignments and I’m just waiting for it to run to find out.