Circular statistics Maximum likelihood Local likelihood Kenneth D. Harris

advertisement
Circular statistics
Maximum likelihood
Local likelihood
Kenneth D. Harris
4/3/2015
Relationship of cells to oscillations
Klausberger & Somogyi, Science 2008
Relationship of cells to oscillations
O’Keefe & Recce, Hippocampus 1993
How do we quantify:
• A cell’s average phase of firing?
• How phase depends on other variables?
Computing instantaneous phase
• Hilbert transform
• Peak fitting
Phase histogram
• Different cells prefer different phases
• How do we compute each cell’s mean phase?
Linear mean doesn’t work
Circular mean
• Treat angles as points on a circle
𝑧 = 𝑒𝑖 πœƒ
• The mean of these gives you
• Circular mean πœƒ
• Vector strength 𝑅
• If all angles are the same:
• πœƒ is this angle
• 𝑅 is 1
• If angles are completely uniform
• 𝑅 is 0
• πœƒ is meaningless.
𝑧 = 𝑅𝑒 π‘–πœƒ
R
πœƒ
von Mises distribution
𝑒 πœ… cos πœƒ−πœƒ0
𝑝 πœƒ =
2πœ‹πΌ0 πœ…
• πœƒ0 is central angle
• πœ… is concentration parameter
• Larger -> more peaked
• Zero -> uniform distribution
• 𝐼0 πœ… is a Modified Bessel function
• Needed to make probabilities integrate to 1.
Maximum likelihood estimation
𝑒 πœ… cos πœƒ−πœƒ0
𝑝 πœƒ; πœƒ0 , πœ… =
2πœ‹πΌ0 πœ…
• Given data set πœƒ1 … πœƒπ‘ , how do we estimate parameters πœ… and πœƒ ?
• Choose them to maximize
𝑝 πœƒπ‘– ; πœƒ0 , πœ…
𝑖
Maximum likelihood estimation
• Given data set πœƒ1 … πœƒπ‘ , how do we estimate parameters πœ… and πœƒ ?
• Log likelihood
𝐿=
log 𝑝 πœƒπ‘– ; πœƒ0 , πœ…
𝑖
=πœ…
cos πœƒπ‘– − πœƒ0 − 𝑁 log 𝐼0 πœ…
𝑖
• Solve
πœ•πΏ
πœ•πœƒ0
= 0 and
πœ•πΏ
πœ•πœ…
=0
− 𝑁 log 2πœ‹
Maximum likelihood von Mises
• πœƒ0 = πœƒ, (πœƒ = circular mean).
• πœ… is the solution of
𝐼1 πœ…
𝐼0 πœ…
= 𝑅, (𝑅 = vector strength).
• 𝐼1 πœ… is another modified Bessel function.
How does phase depend on another variable?
• Don’t use linear regression!
Locally-weighted likelihood
• Log likelihood
𝐿 πœƒπ‘– ; πœƒ, πœ… =
log 𝑝 πœƒπ‘– πœƒ, πœ…
𝑖
• Locally-weighted likelihood
• Spike phases πœƒπ‘– occur at positions 𝐱 𝑖 :
𝐿𝐱 πœƒπ‘– ; πœƒπ± , πœ…π± =
E.g. 𝑀 Δ𝐱 = 𝑒
𝑀 𝐱 − 𝐱𝑖 log 𝑝 πœƒπ‘– πœƒπ± , πœ…π±
𝑖
Δ𝐱 2
2𝜎2
Kernel smoother
∑𝑦𝑖 𝑀 𝐱 − 𝐱 𝐒
𝑦(π‘₯) =
∑𝑀 𝐱 − 𝐱 𝐒
Is local likelihood for a Gaussian
Place field estimation
π‘†π‘π‘–π‘˜π‘’πΆπ‘œπ‘’π‘›π‘‘π‘€π‘Žπ‘ ∗ 𝐾
π‘‚π‘π‘π‘€π‘Žπ‘ ∗ 𝐾
Local likelihood estimation
for a Poisson distribution
Phase field
• Local likelihood estimation for von Mises
Abstract space
Physical space
Harris et al Nature 2002
Confirmatory statistics:
• How to test that phase is independent of place?
• Shuffling method?
• Test statistic?
Download