# Log 04-26-2013

What I did today:

1. Write scorers for the particle filter and abstract SMC algorithms.

2. Make logging and execution fig-based (1 hour)

3. Code clean-up (30 minutes)

4. Properly seed random number generators (20 minutes)

5. Set everything up as a script and run it (??? hours; worked on it during the Gates game night, so was distracted by other things)

# Log 04-25-2013

What I did today:
1. Met with Percy (2 hours)
2. Gates remodeling lunch (1.5 hours)
3. Replaced O(T) updates with O(1) updates / cleaned up API (4.25 hours)

# Log 04-24-2013

What I did today:

2. Implement approximate model based on node potentials (no edge potentials yet) (30 minutes)

3. NLP reading group (1 hour 30 minutes)

4. CVX homework (2 hours 30 minutes)

5. Re-implement smoothers as filters (1 hour)

6. Fix add() method (1 hour)

7. Make particle filter robust (10 minutes)

8. Skype meeting with Andrew (1 hour 30 minutes)

# Log 04-23-2013

Productive day!

1. Batch mass computation for pi-system (30 minutes)

2. Numerically stable mass updates (45 minutes)

4. [SKIPPED] Particle filter smoothing (5 minutes before deciding to abandon as not worthwhile)

5. Meeting with Rok (105 minutes)

6. Think about how to link together different values of $\mu(a,b)$ / how to define a cost bound in terms of $\mu(a,b)$ (90 minutes)

7. Meet with Percy (105 minutes)

8. Housing search (45 minutes)

9. Write up notes from meeting with Rok (30 minutes)

10. Skim chapter 3 of Boyd’s book (15 minutes)

# Log 04-22-13

What I did today:

1. Implement rotation-based factorial HMM (40 minutes)

2. Work out details of different mass-adjustment schemes; it seems that the right thing to do is to have a function $\mu(a,b)$ which is the mass assigned to b under the model assumed by a, and that this function has to have a non-negative Mobius inverse. (1 hour 15 minutes)

3. Plan and implement one such scheme (35 minutes, although there appear to be at least some issues with this scheme)

4. Attend Arun’s talk (1 hour)

5. Understand why current HMM model has problems; I think I’ve confirmed that it’s due to some events dominating other events in probability. (20 minutes)

I’ve noticed in general that it’s difficult to work immediately after playing frisbee; I should find a way to counteract this, perhaps caffeine will work.

# Log 04-19-2013

What I did today:

1. ML reading group (2 hours)

2. Some debugging of Pi-system / particle filter code with visualizer (1 hour)

3. Go to Slav Petrov’s talk (1.5 hour)

4. Meet with Noah / Dan / Joe about bridge (1 hour)

5. Meet with Slav (0.5 hour)

6. Meet with Percy (1 hour)

# Log 04-18-2013

What I did today:

1. Figured out why some particles don’t get added to the Pi-system; it seems like it may be due to some trajectories dominating other trajectories in likelihood. (1.5 hours)

2. NLP lunch (1.5 hour)

3. Write visualizer (3 hour)

4. Vannevar meeting (2 hour)