Overall plan

[billbrod]

Overall plan for contents we'll include

Concepts to cover

• Pytrees
• GLM math
• Basis: eval vs. conv
• How to pick basis, number of basis funcs, etc
• Initial analysis with pynapple (hide?)
• inference on relevant features
• Comparison with pynapple
• cross-validation
• regularizers

Introduction

We should have a page in our docs with the introduction: what's the format for the workshop, who are we, what do we want to get out of it, etc (and remember to present it too)

Tutorials

Conceptual intro

• describe model as linear model of log firing rates (which we use to handle non-negativity constraint)
• call it a LNP model, rather than GLM
  • "common to call this a GLM in neuroscience but this is what's typically done". also called Poisson regression
  • note that "GLM" also refers to many other things: in stats, there are many other generalized linear models (link to wikipedia), and in fMRI, this refers to a general linear model, which is just linear regression
• start with a brief user's guide that everyone should read, 2 paragraphs, that describes the things everyone needs to know, including: it's a convex problem (and what that means).
• mention that the loss function is not least squares? (which is probably what they're used to)
• and that it's probabilistic
• then go through Edoardo's theory and exercises with more links to Pillow's tutorials and other resources

Sketch:

• What GLM?
  • LNP
  • probabilistic, log-likelihood
  • convex
  • box: what it's not? terminology confusions
• Why GLM?
  • LNP, justify each component: start with linear model of firing rate, point out firing rate can't be negative, and we'd like a way to transform firing rate to spikes
  • bases for non-linearity: capture interesting phenomena while preserving above properties

Current injection

Single cell current injection data set from Allen Institute

1. No basis, just self-excitation
2. Basis for self-excitation (introduce basis convolution)
3. Current injection input (with basis)

Head direction

Head direction-tuned cells in thalamus, all with arrays

1. Get initial tuning from pynapple
2. Just coupling, can recover ring connectivity between cells, which can show they're tuned for direction
3. Ridge regression and discuss interpretation
4. extra: fit tuning explicitly with GLM. adding direction shouldn't change tuning much.

Spatial position

grid cells in MEC

1. Initial tuning with pynapple
2. Only coupling, show no position tuning recovered
3. Introduce pytrees
4. Add position input, (introduce basis evaluate), get tuning

Head direction + spatial position

Same dataset as above?

1. Use all of head direction, spatial position, and coupling as inputs to GLM (using pytrees)
2. Discuss importance of cross-validation and regularizers here

V1

Macaque V1 in response to white noise

1. Use spike-triggered average to get receptive field
2. Use receptive field to build more complex model matrix (1d filtered signal as input), 2 part procedure
3. Show which cells are simple cells (based on STA output?)
4. Coupling helps tuning curve sharpen? (This is what Pillow 2008 showed)

Auditory

gerbil primary auditory cortex, passive listening to pure tones, noise, and animal vocalizations

1. also 2 part procedure, first preprocess input (split into frequency channels / cochleagram), then use GLM
2. No basis on frequency input
3. standard analysis: try to get tuning properties from pure tones. spectral-temporal receptive field (strf): response in Hertz by delay (msec).
4. could then predict spiking in response to more complex stimuli. see if coupling improves performance? lateral connectivity?
5. inter-trial interval is several seconds

Todo

• [x] merge pytree PR (@billbrod )
• [ ] V1 spike-triggered analysis with pynapple (@gviejo)
• [ ] GLM math: build up as necessary in tutorials. most of basics in tutorial zero, interleaved with code (@BalzaniEdoardo)
• [x] create repo (@billbrod )
• [ ] NWB files for each dataset (@gviejo)
• [ ] Helper functions to convert params to tuning (@BalzaniEdoardo), with array and pytree
• [x] binder setup (@billbrod )
• [ ] Write tutorials, starting from current injection (all)

[billbrod]

For binder:

• We want gpu access. How many CPUs? Memory?
• Can get data already in the image (download on my cluster account, then symlink)
• In setup, have binder download jupyter notebook versions of all the tutorials?

[gviejo]

Data for head-direction : https://pynapple-org.github.io/pynapple/generated/gallery/tutorial_HD_dataset/

import numpy as np
import pynapple as nap
import requests, math, os
import tqdm

path = "Mouse32-140822.nwb"
if path not in os.listdir("."):
    r = requests.get(f"https://osf.io/jb2gd/download", stream=True)
    block_size = 1024*1024
    with open(path, 'wb') as f:
        for data in tqdm.tqdm(r.iter_content(block_size), unit='MB', unit_scale=True,
            total=math.ceil(int(r.headers.get('content-length', 0))//block_size)):
            f.write(data)

data = nap.load_file(path)  # Load the NWB file for this dataset

[billbrod]

also, should talk about GPUs