Not trying to "Schmidhuber" this or anything, but I think my 1999 NIPS paper gives a cleaner derivation and explanation for working on the Jacobian. In it, I derive a Jacobian operator that allows you to compute arbitrary products between the Jacobian and any vector, with complexity that is comparable to standard backprop.<p>[*] G.W. Flake & B.A. Pearlmutter, "Differentiating Functions of the Jacobian with Respect to the Weights," <a href="https://proceedings.neurips.cc/paper_files/paper/1999/file/b9f94c77652c9a76fc8a442748cd54bd-Paper.pdf" rel="nofollow">https://proceedings.neurips.cc/paper_files/paper/1999/file/b...</a>
A really cool post and a great set of visualizations!<p>Computing sparse Jacobians can save a lot of compute if there's a real lack of dependency between part of the input and the output. Discovering this automatically through coloring is very appealing.<p>Another alternative is to implement sparse rules for each operation yourself, but that often requires custom autodiff implementations which aren't easy to get right, I wrote a small toy version of a sparse rules-based autodiff here: <a href="https://github.com/rdyro/SpAutoDiff.jl">https://github.com/rdyro/SpAutoDiff.jl</a><p>Another example (a much more serious one) is <a href="https://github.com/microsoft/folx">https://github.com/microsoft/folx</a>
This paper is written by three Europeans who clearly understand these mathematical ideas.<p>Is this type of analysis a part of a particular mathematical heritage ?<p>What would it be called ?<p>Is this article relevant ?
<a href="https://medium.com/@lobosi/calculus-for-machine-learning-jacobians-and-hessians-816ef9d55a39" rel="nofollow">https://medium.com/@lobosi/calculus-for-machine-learning-jac...</a>
The classic reference on the subject is "Numerical Linear Algebra" by Lloyd Trefethen. Skip to the last chapter on the iterative methods for computational aspects. You'll learn a lot more and faster with Matlab.<p><a href="https://davidtabora.wordpress.com/wp-content/uploads/2015/01/lloyd_n-_trefethen_david_bau_iii_numerical_line.pdf" rel="nofollow">https://davidtabora.wordpress.com/wp-content/uploads/2015/01...</a><p>A short overview is chapter 11 in Gilbert Strangs's Intro to linear Algebra
<a href="https://math.mit.edu/~gs/linearalgebra/ila5/linearalgebra5_11-1.pdf" rel="nofollow">https://math.mit.edu/~gs/linearalgebra/ila5/linearalgebra5_1...</a><p>AD comes from a different tradition - dating back to FORTRAN 77 programers attempt to differentiate non-elementary functions (For Loops, procedural functions, Subroutines, etc). Note the hardware specs for some nostalgia
<a href="https://www.mcs.anl.gov/research/projects/adifor/" rel="nofollow">https://www.mcs.anl.gov/research/projects/adifor/</a>
Picking my way through this slowly... I'm familiar with autodiff but some of these ideas are very new to me. This seems really, really exciting though.
The optimal control framework Casadi has had the ability to compute sparse jacobians and hessians for a long time (maybe a decade?), which come up all the time in trajectory optimization.This not only provides massive speed ups in both the differentiation and linear solver time, but also greatly reduces the memory requirements. If this catches on in machine learning, it will be interesting to see if we can finally move past first order optimization methods.