Chalk print identifies user

Neural networks predict professors based on chalk leftovers

An episode in CSI? No.

Sheer curiosity, perhaps. Researchers at the National Institute of Physics, headed by Dr. Christopher Monterola, would just like to know if chalk leftovers, its tip profile in particular, would give an idea as to who its user was.

In their paper accepted for presentation at the Samahang Pisika ng Pilipinas 2008 Congress, Dr. Monterola and his team composed of J. Tugaff, I. Crisologo, A. Longjas and R. Batac presented this as "a possibility", by using photometric stereo profiles as inputs to an artificial neural network [1].

Teachers conducting lectures often have signature ways of using chalk to write on a board. Some like to use just one side of the chalk, leaving the tip flat and broad for writing. Other unknowingly twist and rotate the chalk to maintain pointed tip and finer lines.

To quantify how differently teachers do their thing with that calcium carbonate stick, the team supplied fresh (unused) chalk samples for ten different teachers from the Physics and Mathematics institutes in the University of the Philippines. Whatever remains of these chalk samples were collected back for processing.

Images of the chalk tips were taken for at four different lighting conditions, and without lighting (for removing bias). Images were processed by using photometric stereo technique to get the chalk tip profiles.

The height profiles (basically just a two-dimensional array of height values) were reduced into the most significant values by principal component analysis, and the reduced vectors were used as input to a neural network (NN).

Dr. Monterola emphasized that NN's still astound scientists: patterned after neurons in the brain, artficial NN's can "learn" by "training"; i.e. forcing it to understand a certain pattern. Artificial NN research is still an active arena of research involving a wide array of applications.

In this work, the NN tried to classify whose chalk leftover a sample is. But how good really is a neural network prediction? For sure, NN can be less than 100% in accuracy, say 99%; but is 99% good enough?

To answer that question, NN accuracy is compared with that of chance. In this case, given a chalk and ten possibilities for a user, one has 1/10 chance of guessing it right (note that repeated guesses for quite a number of chalk samples further decrease this chance of correct prediction). The researchers pegged the chance proportion value at roughly 10%.

The NN right guess is at 50%, several times higher than chance. In this case, 50% is pretty much acceptable.





Chalk tip profiles can quantify only a single aspect of the complex classroom dynamics. In the end, what matters is not the amount of chalk that remained, but how much information - written using these chalk pieces - the students retained.

1. Tugaff, J., Crisologo, I., Longjas, A., Batac, R. and Monterola, C. (2008) "Chalk print? Feasibility of predicting the lecturer based on his/her chalk leftovers," accepted for presentation in the 26th Samahang Pisika ng Pilipinas Congress.