A directed graph whose nodes denote elementary computations (operations) or data placeholders (variables/tensors) and whose edges denote the flow of data values between those computations; used to represent, analyze, and execute mathematical computations.

Principle

Decompose a complex computation into nodes connected by data dependencies so that evaluation, transformation (e.g., algebraic simplification), and differentiation can be performed by traversing or transforming the graph—often as an acyclic structure for forward evaluation and reverse-mode differentiation.

Demonstration

A neural network forward pass can be represented as a graph where nodes are matrix multiplications, bias additions, and nonlinearities; tensors flow along edges. Backpropagation is implemented by traversing the same graph in reverse to accumulate gradients.

Misapplication

Treating a computational graph as a control-flow model (using its edges to reason about execution ordering of statements with side effects) or assuming every computational graph is acyclic and ignoring allowed cyclic motifs for recurrent definitions, which leads to incorrect assumptions about evaluation or convergence.

Consequence

When used correctly, it enables automatic differentiation, symbolic manipulation (common-subexpression elimination), parallel execution scheduling, and memory/recomputation trade-offs informed by data dependencies.

Reversal

View the program as text or an imperative sequence of instructions where operations are ordered by control flow rather than by explicit data edges; inverting the concept emphasizes execution steps instead of value dependencies.

Boundary

Applies to representations that expose computations and value flow; it does not, by itself, encode runtime scheduling, low-level resource allocation, or non-data control effects unless extended. Distinguish static 'define-then-run' graphs from runtime-constructed graphs.

Semantic Tension

Competes with the term 'dataflow graph'—both emphasize value movement but dataflow focuses on token passing and firing semantics, whereas computational graph emphasizes algebraic composition and differentiation for numerical computation.

Synthesis

A computational graph is the structured, directed representation of a computation in which nodes are operations or data holders and edges are value dependencies; this structure makes analytic tasks (derivatives, optimizations, parallelization) tractable by graph algorithms while remaining distinct from control-flow or runtime scheduling concerns.