Add GLMPPSeq: PPSeq with GLM Background Model

[smejak]

Summary

This PR introduces GLMPPSeq, a generalized version of PPSeq that incorporates a Generalized Linear Model (GLM) for modeling time-varying background firing rates. This extends the original PPSeq framework to handle non-stationary neural data where background activity varies over time.

Key Features

GLMPPSeq Model:

• Time-varying background: Models background firing rates using a GLM with RBF (Radial Basis Function) covariates, replacing the constant background rate in vanilla PPSeq
• Flexible covariate system: Uses smooth temporal basis functions to capture slow changes in population activity
• Sparsity regularization: Supports L1 penalties on both sequence amplitudes (l1_amp) and template neuron participation (l1)
• Configurable RBF basis: rbf_width parameter controls temporal smoothness of background (scale factor relative to basis spacing, default=0.2)
• Efficient Newton-Raphson updates: Vectorized GLM parameter optimization with numerical stability safeguards

Implementation Details

• EM algorithm with three update steps:

  1. E-step: Update sequence amplitudes given current parameters
  2. M-step (background): Update GLM coefficients via Newton-Raphson
  3. M-step (templates): Update Gaussian template parameters via moment matching

• Proper GLM initialization in initialize_random() and initialize_default()

• Consistent use of phi for covariates throughout the class

• Device-aware tensor operations for GPU acceleration

Backward Compatibility

• PPSeq and CAVI classes remain unchanged
• New model is in separate GLMPPSeq class to avoid breaking existing code

Testing Recommendations

After merge, users should verify:

• Model initializes without errors across different parameter settings
• fit() runs successfully on both CPU and GPU
• GLM background rates properly capture temporal dynamics
• L1 regularization produces expected sparsity

[xanderladd]

this looks good.

a few things to consider going forward:

• are we clipping eta (line 129) and delta (line 270) to agressively? to leinent? we can set these thresholds empirically
• same as above for scales = torch.relu(scales - self.l1)
• if we add inference mode T_for_RBF will be annoying - RBF basis wont have correct time axis
• do we ever want to add RBF basis with different lengthscale / width

[xanderladd]

we should allow this 0.1 to configured - sets the learning rate for newton updates

[xanderladd]

If there is a better way to do T for RBF, it would be nice - if you want to do inference, it needs to be set again and can cause some issues... though model isn't currently set for inference. but we can maybe implement this in a cleaner way next time