The Living Thing / Notebooks :

Automatic differentiation

Getting your computer to tell you the gradient of a function, without resorting to finite difference approximation. I am mostly interested here in the sense of automatic forward or reverse mode differentiation, which is not, as such, a symbolic technique, but symbolic differentiation gets an incidental look-in, and these ideas do of course relate.

Infinitesimal/Taylor series formulations, the related dual number formulations, and even fancier hyperdual formulations. Reverse-mode, a.k.a. Backpropagation, versus forward-mode etc. Computational complexity of all the above.

There is a beautiful explanation of the basics by Sanjeev Arora and Tengyu Ma.

You might want to do this for quadrature, or for optimisation, either batch or SGD, especially in neural networks, matrix factorisations, variational approximation etc. This is not news these days, but it took a stunningly long time to become common since its inception in the… 1970s? See, e.g. Justin Domke, who claimed automatic differentiation to be the most criminally underused tool in the machine learning toolbox. (That escalated quickly.) See also a timely update by Tim Viera.

Related: symbolic mathematical calculators.


There are many ways you can do automatic differentiation, and I won’t attempt to explain the various approaches here since there is much of good material out there already with fancy diagrams and the like. Symbolic, numeric, dual/forward, backwards mode… Notably, you don’t have to choose between them - e.g. you can use forward differentiation to calculate an expedient step in the middle of backward differentiation, for example.

See, e.g. Mike Innes’ hand-on introduction, or his opinionated but clear introductory paper, [#Inne18].



Julia has an embarrassment of different methods of autodiff (Homoiconicity and introspection makes this comparatively easy.) and it’s not always clear the comparative selling points of each.

The juliadiff project produces ForwardDiff.jl and ReverseDiff.jl which do what I would expect, namely autodiff in forward and reverse mode respectively. ForwardDiff claims to be very advanced. ReverseDiff works but is abandoned.

ForwardDiff implements methods to take derivatives, gradients, Jacobians, Hessians, and higher-order derivatives of native Julia functions

In my casual tests it seems to be a slow for my purposes, due to constantly needing to create a new closure with a single argument it and differentiate it all the time. Or maybe I’m doing it wrong, and the compiler will deal with this if I set it up right? Or maybe most people are not solving my kind of problems, e.g. finding many different optima in similar sub problems. I suspect this difficulty would vanish if you were solving one big expensive optimisation with many steps, as with neural networks. update: I has doing it wrong. This gets faster if you avoid type ambiguity by, e.g setting up your problem in a function to avoid type ambiguities. I’m not sure if there is any remaining overhead in this closure-based system, but it’s not so bad.

In forward mode (desirable when, e.g. I have few parameters with respect to which I must differentiate), when do I use DualNumbers.jl? Probably never; it seems to be deprecated in favour of a similar system in ForwardDiff.jl. But ForwardDiff is well supported. It seems to be fast for functions with low-dimensional arguments. It is not clearly documented how one would provide custom derivatives, but apparently you can still use method extensions for Dual types, of which there is an example in the issue tracker. The recommended way is extending DiffRules.jl which is a little circuitous if you are building custom functions to interpolate. It does not seem to support Wirtinger derivatives yet.

Related to this forward differential formalism is Luis Benet and David P. Sanders’ TaylorSeries.jl, which is satisfyingly explicit, and seems to generalise in several unusual directions.

TaylorSeries.jl is an implementation of high-order automatic differentiation, as presented in the book by W. Tucker #Tuck11. The general idea is the following.

The Taylor series expansion of an analytical function \(f(t)\) with one independent variable \(t\) around \(t_0\) can be written as \[ f(t) = f_0 + f_1 (t-t_0) + f_2 (t-t_0)^2 + \cdots + f_k (t-t_0)^k + \cdots, \end{equation} \] where \(f_0=f(t_0)\), and the Taylor coefficients \(f_k = f_k(t_0)\) are the \(k\)-th normalized derivatives at \(t_0\):

\[ f_k = \frac{1}{k!} \frac{{\rm d}^k f} {{\rm d} t^k}(t_0). \]

Thus, computing the high-order derivatives of \(f(t)\) is equivalent to computing its Taylor expansion.… Arithmetic operations involving Taylor series can be expressed as operations on the coefficients.

It also has a number of functional-approximation like analysis tricks. TBD.

HyperDualNumbers, promises cheap 2nd order derivatives by generalizing Dual Numbers to HyperDuals. (ForwardDiff claims to support Hessians by Dual Duals, which are supposed to be the same as HyperDuals.) I am curious which is the faster way of generating Hessians out of ForwardDiff’s Dual-of-Dual and HyperDualNumbers. HyperDualNumbers has some very nice tricks. Look at the HyperDualNumbers homepage example, where we are evaluating derivatives of f at x by evaluating it at hyper(x, 1.0, 1.0, 0.0).

> f(x) = ℯ^x / (sqrt(sin(x)^3 + cos(x)^3))
> t0 = Hyper(1.5, 1.0, 1.0, 0.0)
> y = f(t0)
4.497780053946162 + 4.053427893898621ϵ1 + 4.053427893898621ϵ2 + 9.463073681596601ϵ1ϵ2

The first term is the function value, the coefficients of both ϵ1 and ϵ2 (which correspond to the second and third arguments of hyper) are equal to the first derivative, and the coefficient of ϵ1ϵ2 is the second derivative.

Really nice. However, AFAICT this method does not actually get you a Hessian, except in a trivial sense, because it only seems to return the right answer for scalar functions of scalar arguments. This is amazing, if you can reduce your function to scalar parameters, in the sense of having a diagonal Hessian. But that skips lots of interesting cases. One useful case it does not skip, if that is so, is diagonal preconditioning of tricky optimisations.

Pro tip: the actual manual is the walk-through which is not linked from the purported manual.

How about Zygote.jl then? That’s an alternative AD library from the creators of the aforementioned Flux. It usually operates in reverse mode and does some zany compilation tricks to get extra fast. It also has forward mode. Has many fancy features including compiling to Google Cloud TPUs. Hessian support is “somewhat”. Flux itself does not yet default to Zygote, using its own specialised reverse-mode autodiff Tracker, but promises to switch transparently to Zygote in the future. In the interim Zygote is still attractive has many luxurious options, such as defining optimised custom derivatives easily, as well as weird quirks such as occasionally bizarre error messages and failures to notice source code updates.

One could roll one’s own autodiff system using the basic diff definitions in DiffRules. There is also the very fancy planned Capstan, which aims to use a tape system to inject forward and reverse mode differentiation into even very hostile code, and do much more besides. However it also doesn’t work yet, and depends upon Julia features that also don’t work yet, so don’t hold your breath.

See also XGrad which does symbolic differentiation. It prefers to have access to the source code as text rather than as an AST. So I think that makes it similar to Zygote, but with worse PR?