From 82a33fd654d7b53cdfdfbe277a7ccb5847839d9c Mon Sep 17 00:00:00 2001 From: Kevin Shan Date: Wed, 7 Dec 2016 18:21:34 -0800 Subject: [PATCH] Elaborate on help text --- @MoDT/MoDT.m | 16 +++++++++++----- 1 file changed, 11 insertions(+), 5 deletions(-) diff --git a/@MoDT/MoDT.m b/@MoDT/MoDT.m index e358acb..a508406 100644 --- a/@MoDT/MoDT.m +++ b/@MoDT/MoDT.m @@ -13,7 +13,7 @@ % User-defined model parameters % mu_t - [T+1 x 1] time frame boundaries (ms) % nu - t-distribution degrees of freedom -% Q - [D x D] covariance of cluster drift, or scalar for diagonal +% Q - [D x D] drift regularization matrix, or scalar for diagonal % C_reg - Diagonal ridge added to C to ensure well-conditioning % Training data (set using attachData) % spk_Y - [D x N] spike data @@ -122,12 +122,18 @@ % frame t is is defined as half-closed interval [ mu_t(t), mu_t(t+1) ) mu_t = zeros(0,1); - % nu: Degrees-of-freedom parameter for the t distribution. Default = 7 + % nu: Degrees-of-freedom parameter for the t distribution. This may range + % from 1 to Inf, with smaller values corresponding to a heavier tails. + % nu=1 corresponds to a Cauchy distribution, and nu=Inf corresponds to a + % Gaussian distribution. Default = 7 nu = 7; - % Q: [D x D] cluster drift covariance. If given as a scalar, it is treated - % as a diagonal matrix with Q along the diagonal. Default = 1 - Q = 1; + % Q: [D x D] cluster drift regularization matrix. If given as a scalar, it + % is interpreted as a diagonal matrix with the given scalar along the + % diagonal. Smaller values correspond to more regularization (producing + % an estimated mu that is smoother over time). The units of this quantity + % are (feature space units)^2/(time frame). Default = 0.033 + Q = 0.033; % C_reg: Scalar that will be added to the diagonal of the scale matrix C % during the M-step. Setting C_reg > 0 can help ensure that C is