Upload (inputs to?) and outputs from the activity classifier to the server.

[kentaylor]

Currently only outputs from the aggregator are uploaded to the server. This minimises network traffic by only sending each activity once regardless of how long it has occurred. Currently classifier inputs are stored in a database in the device and classifier outputs can be added. MET could be added also.

There is a goal to improve activity classification by re-analysing aggregation of activity classifications on the server. This requires that outputs from the classifier are known to the server. Inputs may be useful also. Uploading this data will increase network traffic.

Some consideration of cost vs benefit is required.

[umranium]

Perhaps another idea is that we could upload probability matrixes. Currently our classifier outputs only the final classification, which is fed to the aggregator. Instead, perhaps a better way is for the classifier to output the approximated probability of the given data being a certain activity. The various activities and their respective probabilities could be taken to account in the aggragation algorithm (e.g. HMM). The activities can be simplified to a 32bit int ID in a look-up table, saving bandwidth.

We could have uploading raw-data (i.e. classifier inputs) as optional in the application. Some users (mostly alpha and beta testers) can have it on. This data can be used to optimize and tune our classifier and aggregator algorithms.

[kentaylor]

Yes.

[kentaylor]

Do we have the data to assign probabilities to activities? We have data for a number of activities so that we could calculate a probability for one activity vs another e.g. walking vs travelling. However, where an activity is outside one we have categorised we don't have data which means we can't state a probability of an activity not being that which we are classifying it to be. For example if someone is preparing a meal what is the probability it is classified as walking.