This is a readme file regarding the simulation codes for the paper:

"Rapid Bayesian learning in the mammlian olfactory system"
Naoki Hiratani, and Peter E. Latham
Nature Communications, 11, 3845 (2020)
doi: https://doi.org/10.1038/s41467-020-17490-0

Questions on the manuscript and the codes should be addressed to Naoki Hiratani (N.Hiratani@gmail.com).

"bayesian_learning.py" is the main simulation code corresponding to the results depicted in Figure 3, as well as the results for "the proposed model" in Figure 4-6. "invariant_learning.py" is an extention of "bayesian_learning.py" in which a circuit for acquisition of concentration-invariant representation is added.

All other simulation codes are also available upon reasonable request. Please see the maniscript for the derivations and further details.

In "bayesian_learning.py", inputs are

• coM: the average number of odors simultaneously presented. Throughout the manuscript, coM = 3 unless otherwise stated.
• M: the total nuber of the odor. We used M = 100, except for Fig. 3 and Fig. 4C.
• N: the total number of glomeruli. N = 400 throughout the manuscript.
• sigmax: the standard deviation of input Gaussian noise (sigmax = 1.0).
• Zrho_init: a constant that detemines the initial value of the weight precision parameter (Zrho_init = 0.5, except for Fig. 6BD where Zrho_init = 0.3).
• ik: id of simulation. Curves in the figures are mean over 10 independent simulation unless otherwise stated.

And the output file "bayesian_learning_readout..." contains the odor estimation performance and the weight error after each trial.

Similarly, in "invariant_learning.py, inputs are

• coM: the average number of odors simultaneously presented (coM = 3).
• M : the total number of odor (M = 50).
• N : the total number of glomeruli (N = 200).
• sigmax : the standard deviation of input Gaussian noise (sigmax = 1.0).
• Zrho_init : a constant that determines the initial value of the weight precision (Zrho_init = 0.5)
• circuit_type: "circuit_type = 0" corresponds to the model without piriform to granule connection depicted in Fig. 7A(iii), while "circuit_type = 1" corresponds to the model depicted in Fig. 7A(iv).
• ik : id of simulation.

The output file "invariant_learning_readout..." contains the odor estimation performance and the weight error of both granule cells and piriform neurons after each trial.

UPDATE

2020-09-11: Replaced the link to the published version of the paper.