Skip to content

Latest commit

 

History

History
45 lines (30 loc) · 6.73 KB

README.md

File metadata and controls

45 lines (30 loc) · 6.73 KB

release_tag

Automatic Differentiation

Algodiff is a small header-only library for automatic differentiation. Automatic (also known as algorithmic or computational) differentiation is a core technology underlying applications that span scientific and high performance computing, optimization, and sensitivity analysis. These areas in turn serve as the foundation for myriad computing applications across areas like machine learning, computational engineering and design, physical simulation, and graphics, to name a few.

Contents

  1. Getting Started
  2. The Derivative Function
  3. Exploiting Computational Abstraction
  4. Example

Getting Started

Algodiff can be compiled and linked using the open-source cross platform compiler CMake. For first-time CMake users refer to the documentation. Algodiff requires a CMake version of 3.0 or newer.

The Derivative Function

The notion of a derivative function (throughout the README I use the terms derivative and gradient interchangeably) is one of the most powerful constructs in applied mathematics. In the next section I describe how the structure of our software systems enable us to compile derivative functions from a programmatic point of view, but it's worth thinking first about what information the derivative function exposes about its integral.

Exploiting the Computational Abstraction

From the computational point of view, one can view an automatic differentiation tool as a kind of domain specific compiler which takes as its source language a certain class of procedures defined in some high-level programming language (e.g., C, C++) and transforms this program into one which computes the derivative function associated with the given procedure. In this sense, the source language of the compiler can be thought of as the set of programs associated with mathematical functions in compliance with certain smoothness properties (so the derivative function is well-defined). The target language can be though of as the set of programs associated with some more general class of mathematical functions (since the derivative function need not be continuous).

Even more interestingly, these derivative programs compute derivatives with the same numerical accuracy as the source program. Giving this a little thought, this seems to defy intuitions gained from finite differencing. The abstraction hierarchy of our computing model is the solution to the riddle of how this accuracy can be gained, and how these compilers work.

Even the most complex computer progams can be composed from a relatively simple set of primitives. This is evident when you consider that the microarchitecture of the computer is limited to a relatively simple finite set of operations which in turn are somehow accomplished via the relatively simple digital operation of transistors. In software too then, apparently complex programs like operating systems and compilers are the result of applying a sequence of simple architectural operations.

This compositional structure of a computing system corresponds directly with the compositional structure of programs which compute mathematical fuctions. Although the corresponding function may itself be complex, apparently the function is actually computed using a relatively simple set of operators. By viewing the aggregate function as the composition of relatively simpler mathemetical operations (ideally those that can be easily associated with programmatic objects), we can propagate derivatives through the program by strategically applying the chain rule. I concretize this notion and provide an example below.

Example

Consider the mathematical function equation. We can analytically identify that the function is continuous (it is the sum of a finite number of continuous functions), and determine that the derivative function is equation. In this example we'll consider the process of compiling a computer program which computes the function into a different program which computes the derivative function.

That is, given the source program:

float f(float x)
{
    return math::sin(x) + 5 * math::pow(x, 3) + 3; 
}

How can we compile it into a program which computes the derivative function? The core insight of implementing automatic differentiation tools is that all numerical programs are composed of a small finite set of mathematical primitives. This is analagous to the sense in which large classes of mathematical functions are compositions of small finite bases.

Restricting our attention to the operations induced on our independent variable x in computing the dependent variable (the output), in this case our procedure f is the composition of four primitives: math::sin(x), operator+(float x, float y), math::pow, and operator*(float x, float y). Mathematically if we apply the chain rule to f, we have d/dx sin(x) * d/dx x + d/dx 5 * x^3 + 5 * d/dx x^3 + d/dx 3. If we simply knew and hard coded the derivatives of these four operators, we could compute the aggregate derivative function.

In some sense, computational differentiation is simply the art and science of designing compilers endowed with the built in differentiation intrinsics. In the literature these intrinsics are called elementals. For the majority of applications in scientific computing and machine learning, the so-called polynomial core is a sufficient set to construct arbitrarily complex functions: this set includes the binary operations of addition and multiplication, the unary negation, and the initialization to a constant. These are smooth, continuously differentiable functions, and so are simple to reason about. Other smooth functions in this class include trigonometric operators like the sine and cosine functions, the exponential, and the reciprocal function. Trickier classes of functions include functions that are Lipschitz continuous, or more general functions like the heaviside function or the ternary conditional. Even more generic computational structures like Monads, closures, and control flow structures can also be computationally differentiated.