Blog

Summary

Supervised learning using a neural network with resilient backpropagation. Simple, it was trained with only 33 rats, and it seems to present good results. Pointed looking to non-exploratory actions to explain significant errors.

Notes

  • Had the goal of modeling the behavior of a normal rat. Injected a saline solution in rats to create stressful condition present in other similar experiments.
  • Sample size of 33 rats.
  • Used a software called Prostcom to analyze videos recorded in the VCR. Later, produced a dataset to do Supervised Learning.
  • Written in Matlab.
  • Elevated plus-maze is divided into 13 sections, 3 per arm and an extra for the center.
  • Used a neural network with resilient backpropagation, converging “in less than 1,000 epochs.” Later, generated 33 different networks, or virtual rats. The output of the neural network: forward, backward, right, left.
  • The measured variables were: time in open arms, time in closed arms, number of entries in open arms, number of entries in closed arms, percentage of time in open arms, percentage of time in closed arms, number of entries in any arms, and statistical significance.
  • The only variable with a significant difference is the time in closed arms. Authors attribute the error to the large difference in sample times (from 2 to 100s), training data errors, and other behaviors such as stopping to groom itself, turning randomly inside the same square…

Thoughts

  • Briefly looking online (Conde, Costa, & Tomaz, 2000), this Prostcom doesn't seem to have many references. It's a software from 1986, created to help a person to transcribe a video recording of a maze into a dataset.
  • If one can reproduce these results, maybe an argument in favor of not having to look for a large sample size. They seemed to be capable of producing a reasonably accurate model with small data and quick training.
  • Would be interesting to add more output variables to the model. Maybe it's not about predicting the square in which the rat will move, but what is the next action. For example, the prediction could be that it will stay in the same square and will groom itself.
  • I am not sure why measuring all these variables. They seem to be more useful to demonstrate the accuracy rather than helping to train the model.

References

Conde, C., Costa, V., & Tomaz, C. (2000). PROSTCOM: Un conjunto de programas para registro y procesamiento de datos comportamentales en investigaciones de fisiologia y farmacologia. Biotemas, 13(1), 145-159.