These results reflect object smoothing and learning all the CPTS: Layers ofModelTime-Slice AccuracyEdit Distance SmoothingPrimary ActivityAny Activity 0"single-trained" straw man (prior)84.8% (std 6.3).90.35% (std 5.0)14.7 std(6.9) 1"single-trained" straw man (prior)87.35% (std 3.8).92.95% (std 3.6) 6.7 std(3.5) 2"single-trained" straw man...

These results reflect one change: I smoothed the CPTs between EM iterations to eliminate zero entries. ModelTime-Slice AccuracyEdit Distance Primary ActivityAny Activity "single-trained" straw man (prior)84.4% (std 6.6).90.1% (std 5.1) 16.2 std(7.4) "full-trained" straw man (state-transitions)71.43% (std 5.3)77.71% (std...

These results reflect two changes. 1) I allowed EM to train the starting probabilities, pi_0 2) I changed the way "no observation" observations were handled The second change means that I tweaked the strawman so that it wasn't so...

Here is the strawman: I calculated the mean length of each activity from the two training runs that I did of each activity by itself. From that mean length I fit an exponential curve to the amount of time...

Today, I took some time out to make sure our straw man was going to do badly. The above graphs confirm this. The top graph is of forward particle filter inference with hand made models as priors and no...

Okay, backward inference seems to be done. Here is a graph comparing the two inferences. The top is forward inference. The bottom is backward inference. They both seem to perform better in different portions of the state space. That...

Forward approximate inference on the data is complete. The image above shows the results of filter the data forward. The x axis is time. The y axis is "activity". There is a heavy colored line when the activity is...

Data collection is complete 11 activities total 23 training runs of the activities by themselves (2 each plus an extra) 10 full breakfast runs with all 11 activities completed and interleaved 2 incomplete breakfast runs with a subset of...

Math review for learning with particle filters....

Here is a viz of the data run from this morning. x axis is time in seconds. y axis is objects. A circle is an observation. Some features to note are the vertical stripe on the upper left...that is...

"Current model" Okay I've got training data for the individual activities. That's 2 traces of 11 activities, mostly in the kitchen, all in the house. I started collecting a full run of all the activities when I ran out...

I have collected two training runs of all of the activities from the previous post. It took all day. Tomorrow I'm going to try and record the combo runs of all the activities together....

Alright, I'm pretty excited. Things are going well. I've got a full interrupt model running on toy data right now. I'm doing exact DBN inference and EM learning. A picture of the model is on the left. Starting from...

I've got the equipment, as I mentioned before, to record traces, but I'm spending a little more time making sure that I can do the inference on toy data. Right now I've got a small example demonstrating the ability...