Any more syllabus?

[colinleach]

In amongst the chatter about data analysis and statistics in R, it may be worth noting that we left the LinAlg stuff at 2 docs-only concept and a vague thought about adding factorization sometime. I guess there is more we could do, if sufficiently motivated?

[depial]

I think there are some concepts that could still be added (metaprogramming and multi-threading jump to mind). I would like to get to the magic number of 40 (as per the track build status), or 42 if we're going to nerd out as per usual 😆

We'd be at 36 concepts with the LinAlg concepts, so I've been meaning to get to writing exercises for them. For the first one, I had considered doing a feed forward network since this has a simple linear algebra aspect with a ML flair, but I'm thinking it might be too light on the linear algebra (it only really needs matrix multiplication and elementwise operations). I'll keep thinking though.

[depial]

I've been mulling over the LinAlg basics concept exercise, and I've got something that resembles a feed forward network, but rather simplified. I've been trying to get size(), permutedims() and matrix multiplication into the exercise, while trying not to over complicate it. A possible solution is the following:

Possible FFN solution

function issemidefinite(M)
    M .== permutedims(M)
end

function initializeweights(x, dimout)
    rand(dimout, size(x, 1))
end

function forwardpass(x, Ws)
    current = x
    for W in Ws
        preactivation = W * current
        current = maximum.(0, preactivation)
    end
    current
end

The function issemidefinite has the bonus that once we introduce eigenvalues, we could revisit this to see if the matrix is positive or negative semidefinite.

I assume the story would be fairly straightforward, that somebody needs some functionality for a feed forward network, and the student has been assigned these tasks. I'm not sure if that's enough or if I should look into doing something else. Any thoughts?

[colinleach]

This seems like a promising way forward, and much better than any suggestion I could make.

I'll probably spend a few more days on practice exercises (currently in the middle of julia/rectangles).

At some point, I also need to finish some part-written R syllabus stuff. Even classes: after being so dismissive of this topic, I find I now need it for some exercises. Something about Karma?