As exercises get revised, the derivative files (solution images and scripts) will become orphaned. We need a little clean-up script to remove those files from the repo to keep everything tidy.

What case policy should we adopt for this repo?

Just a detail that was a bit confusing... In Micro-tutorial 7, in the step-by-step, it says after filtering to take the average, even though it is to take the final value.

Is there any modifications are taking place? As I was going through the tutorials of optimal control, but those are removed and showing as private video.

In the helper function
'plot_trajectory(system, params, initial_condition, dt=0.1, T=6, figtitle=None)':

solution = solve_ivp(system, t_span=(0, T), y0=x0, t_eval=t, args=(params), dense_output=True)

v0 is set to x0, but the argument passed to the helper function is 'initial_condition'

When I load the tutorial I get this

Is there an image missing?

In the notebook the log-likelihood is given like this for Poisson GLM:

inducing to the wrong answer in the exercise.

But in the video it is:

which is the correct one.

I noticed that there is a typo in the links for the "hint" figures (at least so far in W1D1 and W1D2): in the notebooks as opened in colab the figures are not showed (https://raw.githubusercontent.com/NeuromatchAcademy/course-content/master/tutorials/W1D1-ModelTypes/static/W1D1_Tutorial1_Solution_7183720f_0.png) because of a dash instead of an underline (the correct one is https://raw.githubusercontent.com/NeuromatchAcademy/course-content/master/tutorials/W1D1_ModelTypes/static/W1D1_Tutorial1_Solution_7183720f_0.png), probably because a commit has renamed the folders.

First embedded video in Colab Python Workshop1 has only ~1min

The first video is missing from W1D1-tut2. It shows that the video was removed by the uploader.
Also, tutorial videos 6 and 7 on the W1D1 playlist are not present in any of the tutorials.

Porting over the discussion from #121:

Context: the notebook was running fine on colab but failing the notebook check:

---------------------------------------------------------------------------
TypeError                                 Traceback (most recent call last)
<ipython-input-15-0edb6db6447f> in <module>
     23   X_mean = np.mean(X,0)
     24   X_reconstructed = reconstruct_data(score,evectors,X_mean,K)
---> 25   plot_MNIST_reconstruction(X ,X_reconstructed)

<ipython-input-4-148b545ffdc9> in plot_MNIST_reconstruction(X, X_reconstructed)
     51     for k2 in range(3):
     52       k = k+1
---> 53       plt.imshow(np.reshape(X_reconstructed[k,:],(28,28)),extent=[(k1+1)*28,k1*28,(k2+1)*28,k2*28],vmin=0,vmax=255)
     54   plt.xlim((3*28,0))
     55   plt.ylim((3*28,0))

~/miniconda3/envs/nma/lib/python3.7/site-packages/matplotlib/pyplot.py in imshow(X, cmap, norm, aspect, interpolation, alpha, vmin, vmax, origin, extent, shape, filternorm, filterrad, imlim, resample, url, data, **kwargs)
   2649         filternorm=filternorm, filterrad=filterrad, imlim=imlim,
   2650         resample=resample, url=url, **({"data": data} if data is not
-> 2651         None else {}), **kwargs)
   2652     sci(__ret)
   2653     return __ret

~/miniconda3/envs/nma/lib/python3.7/site-packages/matplotlib/__init__.py in inner(ax, data, *args, **kwargs)
   1563     def inner(ax, *args, data=None, **kwargs):
   1564         if data is None:
-> 1565             return func(ax, *map(sanitize_sequence, args), **kwargs)
   1566
   1567         bound = new_sig.bind(ax, *args, **kwargs)

~/miniconda3/envs/nma/lib/python3.7/site-packages/matplotlib/cbook/deprecation.py in wrapper(*args, **kwargs)
    356                 f"%(removal)s.  If any parameter follows {name!r}, they "
    357                 f"should be pass as keyword, not positionally.")
--> 358         return func(*args, **kwargs)
    359
    360     return wrapper

~/miniconda3/envs/nma/lib/python3.7/site-packages/matplotlib/cbook/deprecation.py in wrapper(*args, **kwargs)
    356                 f"%(removal)s.  If any parameter follows {name!r}, they "
    357                 f"should be pass as keyword, not positionally.")
--> 358         return func(*args, **kwargs)
    359
    360     return wrapper

~/miniconda3/envs/nma/lib/python3.7/site-packages/matplotlib/axes/_axes.py in imshow(self, X, cmap, norm, aspect, interpolation, alpha, vmin, vmax, origin, extent, shape, filternorm, filterrad, imlim, resample, url, **kwargs)
   5613                               resample=resample, **kwargs)
   5614
-> 5615         im.set_data(X)
   5616         im.set_alpha(alpha)
   5617         if im.get_clip_path() is None:

~/miniconda3/envs/nma/lib/python3.7/site-packages/matplotlib/image.py in set_data(self, A)
    692                 not np.can_cast(self._A.dtype, float, "same_kind")):
    693             raise TypeError("Image data of dtype {} cannot be converted to "
--> 694                             "float".format(self._A.dtype))
    695
    696         if not (self._A.ndim == 2

TypeError: Image data of dtype complex128 cannot be converted to float

The error is happening because np.linalg.eig is returning complex eigenvectors, and imshow fails if its passed a complex array.

But why is it not failing on colab? Not 100% sure but I was able to reproduce it locally only inconsistently as I was playing around with numpy versions. I don't think it has anything to do with numpy version per se, but that as I was changing the version I was getting numpys that were linked against different linear algebra libraries.

np.linalg.eig has a line internally that checks if the complex part of the eigenvalues are all 0, and returns data with float or complex type depending on the outcome.

The notebook is using np.linalg.eig on a covariance matrix, so the eigenvalues should be real. But I imagine what's happening is that there's a small amount of numerical error in the computation that the linear algebra libraries handle differently. That would account for the difference in return type.

Solution: The W1D5 notebooks should use np.linalg.eigh instead of np.linalg.eig. This function will always return real-valued data. Unlike np.linalg.eig, it will also order the outputs by sorting the eigenvalues. This is done by hand in the tutorials, but not necessary.

Some notebooks have retina plot styles (double the resolution), i.e.

%config InlineBackend.figure_format = 'retina'

For images to display correctly the size has to be written in the img tag (half the true size):

Fix this to prevent giant images, e.g. https://colab.research.google.com/github/NeuromatchAcademy/course-content/blob/master/tutorials/W1D3-ModelFitting/student/W1D3_Tutorial7.ipynb#scrollTo=W25y2KDnYRgN

It can mess up the formatting sometimes to open a colab in Jupyter, save, and bring it back to colab.

I guess colab embeds the filename as metadata and that's what it uses in the UI. So we end up with files like "Copy of Copy of Model Fitting Tutorial 3.ipynb", which is kind of messy.

Not sure if best solved in automated workflow or in making sure canonical versions are manually assigned standardized names during revision.

The nbviewer badge for for T2 lands on 400 : Bad Request and displays a plotly dictionary. Any ideas?

The video links in PythonWorkshop link to video Bayesian Statistics Day, will this part be replaced soon? Or there's no video for the workshop?

Exercise: Predict spike counts with Linear-Gaussian model
After running this chunk of command it's showing "stem() got an unexpected keyword argument 'use_line_collection'"

Hey there,
Maybe more a comment than anything major, but something I stumbled over:
The simulated part leaves off with the impression that exponential ISI distributions should yield most entropy (when the mean is fixed) but then in the interactive tool the "uniform" neurons show the higher entropies.
I think a sentence or two that point out the different conditions would be helpful.

At the linked line, the code calculates AIC using n_samples.

However, in the previous cell (show below), n_samples is last defined for the test dataset.

# You've seen this code before!

### Generate training data
np.random.seed(0)
n_samples = 50
x_train = np.random.uniform(-2, 2.5, n_samples) # sample from a uniform distribution over [-2, 2.5)
noise = np.random.randn(n_samples) # sample from a standard normal distribution
y_train =  x_train**2 - x_train - 2 + noise

### Generate testing data
n_samples = 20
x_test = np.random.uniform(-3, 3, n_samples) # sample from a uniform distribution over [-2, 2.5)
noise = np.random.randn(n_samples) # sample from a standard normal distribution
y_test =  x_test**2 - x_test - 2 + noise

### Fit polynomial regression models
max_order = 5
theta_hat = solve_poly_reg(x_train, y_train, max_order)

It's the training dataset that's being used to calculate AIC and as the training set is larger, it results in an inaccurate AIC (though it doesn't affect model selection).

Solution:
either

1. replace n_samples in the line with `len(y_hat) or
2. use the test dataset for AIC calculation.

This will also require redoing the solution figure for this cell in the tutorial 7 notebook.

We should add a pull request template that auto generates a checklist for late-stage tutorial notebook reviewing. It should have all the things we want to double-check that aren't automated (naming scheme, etc).

What is the standard way for content creators to add non-standard python modules/libraries?

One possibility is to add a file requirements.txt alongside the notebooks, and insert a cell with a single after the title of the notebook:
$pip install -r ./requirements.txt

At least for this new demo notebook: https://github.com/NeuromatchAcademy/course-content/blob/master/tutorials/demo/student/Neuromatch_Tutorial_Format.ipynb

The equation for the optimal theta_hat

\begin{align}
\hat\theta = \sum_{i=1}^N \frac{x_i y_i}{x_i^2}
\end{align}

seems to be incorrect? Shouldn't it be:

\begin{align}
\hat\theta = \frac{\sum_{i=1}^N x_i y_i}{\sum_{i=1}^N x_i^2}
\end{align}

Best,

Jan

In part C of W2D2_Tutorial3 the equilibrium variance of the OU process is calculated as:
Var = sig^2 / (2*(1-lambda))

However, I think that this is not the correct analytical solution of the equilibrium variance. A simple mathematical prove (for equabilirum mean=0) is shown below (the ~ sign means "follows the distribution"):

x0 = x0
x1 = lambda*x0 + N(0, sig^2) ~ N(lambda*x0, sig^2)
x2 = lambda*x1 + N(0, sig^2) ~ lambda*N(lambda*x0, sig^2) + N(0, sig^2) ~ N(lambda^2*x0, (lambda^2+1)*sig^2)
x3 = lambda*x2 + N(0, sig^2) ~ N(lambda^3*x0, (lambda^4+lambda^2+1)*sig^2)
.
.
.
xn ~ N(lambda^n*x0, (lambda^(2*n-2) + ... + lambda^4+lambda^2+1)*sig^2)

Note that:

lambda^(2*n-2) + ... + lambda^4+lambda^2+1 = (1-lambda^(2*n)) / (1-lambda^2)

And since lambda < 1, n -> inf, so lambda^(2*n) -> 0, therefore:

lambda^(2*n-2) + ... + lambda^4+lambda^2+1 = 1 / (1-lambda^2)

Therefore,

xn ~ N(lambda^n*x0, sig^2/(1-lambda^2))

So the equilibrium variance should be sig^2/(1-lambda^2) instead of sig^2/(2*(1-lambda))
And if you use this new analytical solution in exercise 3C, you can find that the fit to diagonal line is better.

The syllabus that's writing on the front page does not reflect the content of the lectures. Someone needs to go in and clean that up.

Re #6. If the QC workflow fails, revise the notebook on Colab then repeat Step 3.

What indicates that the QC has failed?

Re #7. Once the final pieces (e.g. video links) are in place, repeat Step 2 and Step 3, then remove the "draft" status from the PR on github and @ the reviewers for the merge.

How is the merge allowed without reviewers?

In the Parameter Exploration section the code fails if one increases inh_rate variable too much. As the result there are no spikes and this results in "ValueError: zero-size array to reduction operation minimum which has no identity" when computing distribution statistics (_amin function). If this was not left on purpose I'm afraid this may confuse some students.
Details are on a screen-shot below.

.

Below an issue which may be understood politically wrong... maybe retyping the parameter as ISIs would change the understanding?

Machine Learning: Classifiers and regulizers
Should be regularizers

The argument interval in the function restrict_spike_times in W1D1_Tutorial doesn't appear inside. In the function, t_interval should be removed and the argument interval modified to t_interval

hope that helps!

Maybe something like the following is easier to read? @patrickmineault @GunnarBlohm

Mon, July 13: Introduction to Computational Neuroscience and NMA

Description Introduction of datasets (spikes, EEG, fMRI + behavior), and questions about them. These questions will foreshadow the whole summer school.

Day 1 Schedule

In the micro-tutorial 1, Gunnar says he introduced himself, but he hasn't =)

Presumably because Colab copies the notebook to a google server, but does not copy the static data.

But also I think Colab markdown has stricter limitations on HTML constructs.

When you pull a colab from somewhere else, the link that colab auto-adds at the top of the ipython notebook links to the specific branch that the colab was saved to, not to master. Example:

https://github.com/NeuromatchAcademy/course-content/pull/80/files

One points to the day06 branch, the other to Madineh's GH.

Should we adopt the usual spacing in Jupyter notebook (4 spaces) or Colab's 2 spaces?

#121

this seems fishy:

Can you compare two lists like that?

The underlying problem is fixed by #155, but these will need to be regenerated.

This applies only to the colab badge embedded within the notebook itself. The README badges are correct.

"The cell below defines the function to simulate the LIF neuron when receiving external current inputs. You can use v, sp = run_LIF(pars, I) to get the membrane potential and spike train with giveN [?] pars and input current I."

The playlist for W2D3 here
https://github.com/NeuromatchAcademy/course-content/tree/master/tutorials#w2d3---decision-making
links to the W3D3 playlist

The syllabus lists 4 tutorials, though there are only 2, and the themes are different. Here's an updated entry:

The naming scheme we have right now in tutorials won't work for a high number of tutorials. Instead of placing the Bayes day tutorials in the folder called Bayes, how about:

WxDy-SnakeCaseSubject, .e.g. W2D1-BayesianStatistics

Inside, we can have the student version at the top level, the solution inside of TA_solutions, and static content specific for that day inside of static.

When would it be a good time to implement case consistency across the repo?

Some direct links have been shared on social media, like for example:
https://github.com/NeuromatchAcademy/course-content/tree/master/tutorials/reviewing-tutorials.md

This markdown file should be named ReviewingTutorials.md following the new guidelines.

Merged despite tests failing, because tutorials needed be available by tomorrow morning - will take a look. It didn't generate the student version for this reason.

I think the cells with import .... (and some function definitions) are intended to be hidden by default, but when cloning the repo, installing the dependencies using conda and then running jupyter notebook the cells are not hidden.

In Exercise 1: Try to approximate causation with correlation

with plt.xkcd():
    fig, axs = plt.subplots(1,2, figsize=(10,5))
    plot_connectivity_matrix(X[:,[1]], ax=axs[0])

The activity matrix X is plotted as the ground truth instead of the connectivity matrix A

Write a tutorial writing guide. Focus on the biggest pain points (what took the most time). We have a short review guide that could also be improved: https://github.com/NeuromatchAcademy/course-content/blob/master/tutorials/reviewing-tutorials.md

The post-pod review process is going to cut at least one full tutorial, and probably more. We will need to stay on top of removing vestigial tutorials or else they will get picked up by the automated README script, etc.

I don't know whether we want some mechanism for preserving entire "bonus" tutorial documents, but that is a topic for discussion.

When you analytically minimize MSE in the notebook, the result should be sum(xy)/sum(xx) as in the video, and not sum(x*y/x**2)

Content-creators user group