eden-kramer-lab/replay_classification

State space models for categorization of replay content from multiunit spiking activity. Deng et al. 2016

Eden-Kramer-Lab

Website

replay_classification's Issues

State probability is not converging fast enough?

Use trapz instead of sum for the integrals

Replace the Riemann sums with trapezoidal numerical integration should be easy and make the integration more accurate.

State transition model is not working

For the current release, 0.6.0, the state transition model estimation is not working.

Better explanation for advantages of software package

Implement the observation model plots in the same manner as `replay_identification`

Change how training data is entered into the decoder because of the state transition matrix needs the lagged position

In order to get the correct empirical transition matrix, you need the position from the previous time step versus the current time step. This can be a problem when entering training data because if you naively index the array before computing the previous position, the previous position won't be correct.

Options for fixing this are:

• have the user give a training indicator function
• have the user give the lagged position

Use EM to estimate replay speed

Better documentation of class function usage

Make API more like scikit-learn

Make installable from conda

Need repr for classes

Alllow users to see what's going on in class instances more easily

Better normalization than place occupancy?

Add inbound-reverse/outbound-reverse to examples of simulated data in notebook

Implement saving/loading model using joblib

http://scikit-learn.org/stable/modules/model_persistence.html

Change order of dimensions for data in predict_state

I think it's awkward to have the number of signals in the first dimension. Maybe switch time to the first dimension?
data : array_like, shape=(n_signals, n_time, ...)

Fix tests

Implement filter and smoother using `log_sum_exp` trick to prevent under and overflow

See:
https://andrewgelman.com/2016/06/11/log-sum-of-exponentials/
https://www.xarg.org/2016/06/the-log-sum-exp-trick-in-machine-learning/

Add interactive visualizations via holoviews?

Might be better for diagnostics

http://holoviews.org/

Rename trajectory direction?

Property name is confusing. Options:

• experimental task
• task
• context

Better model for reverse time trajectory for the ripple decoding

Xinyi's model assumes that outbound reverse movements are the same as inbound forward movements in the state transition matrix. We could improve this by computing the reverse time trajectory.

Scale maximum likelihood so that peak is always 1

The likelihood gets small if there is a lot of data observed at a time point (in our case, if there were multiple spikes). This leads to numerical stability issues if this exceeds the floating point accuracy of the computer. We can scale the maximum likelihood as follows to avoid the issue:

exp(log(L) - max(log(L))

Differences from xinyi's code

• Trajectory encoding (Outbound vs. Inbound Labeling)
• Fixed error where inbound for one tetrode was miscategorized (at least in the code I saw)
• xinyi normalizes her likelihood
• time bin size - I’m doing everything at 1500 Hz sampling rate
• position bins in likelihood - I evaluate at the center of the bins