search

vanderbilt-data-science / ancient-artifacts

Dynamic image analysis to identify ancient artifacts in soil samples. We will work on microdebitage (the debris of ancient stone knapping first) and later expand to other materials (e.g., mortars).
MIT License
6 stars 2 forks source link

Examine data for calibration drift contributions to labeling and model/test in reversed orders #96

Open csbell-vu opened 3 years ago

csbell-vu commented 3 years ago

Consider the following situation:

  1. A particular sample of soil is run
  2. Due to sensor drift, adhesions, temperature, or other hardware-related factors, one of the measured parameters begins steadily increasing.
  3. A particular sample of microdebitage is run

In this case, the classification may perform better since the model may be able to depend on the drift of one of the measured features. In this case, the microdebitage may have larger (or smaller) values based on systematic error from the measurement device.

To combat against this, some approaches are:

  1. Determine if there is sequential drift in measurements. Do all of the particles seem to be getting larger? Would this make sense from the point of view of the sieves/operation?
  2. Try creating a model from samples created from running the soil first and then the microdebitage and measure the performance. Does it have the same performance if you were to run the microdebitage first and then the soil?
markus-eberl commented 3 years ago

We can definitively test these issues; I'll run samples for all scenarios. I'll also look into the ways how the particle analyzer ID's particles –– does the machine assign the ID sequentially? If so, we can easily order the particles in the sequence in which they were measured.