Core Concepts
Fully connected neural networks training is ∃R-complete, posing computational challenges beyond NP.
Abstract
The article discusses the complexity of training fully connected neural networks, proving that it is ∃R-complete. The authors show that even for simple architectures with two inputs and outputs, the problem is equivalent to determining real roots of multivariate polynomials. They strengthen previous results by demonstrating that algebraic numbers are required for optimal training instances. The content covers the geometric nature of functions computed by neural networks and introduces gadgets to represent variables and constraints. The reduction from ETR-Inv to Train-F2NN establishes the computational hardness of training neural networks beyond NP.
Abstract
Training a two-layer fully connected neural network involves finding weights and biases to fit data points optimally.
Decision problem associated with this task is ∃R-complete, equivalent to determining real roots of polynomials.
Algebraic numbers are needed for optimal training instances even with rational data points.
Introduction
Neural networks are widely used in computer science but face computational challenges in training.
Abrahamsen, Kleist, and Miltzow showed that training two-layer networks with linear activation functions is ∃R-complete.
Preliminaries
Definition of fully connected two-layer neural network architecture.
Introduction of ReLU activation function as commonly used in practice.
Results
Main result: Train-F2NN is ∃R-complete even for simple cases like two input/output neurons.
Algebraic universality demonstrated where solutions require irrational weights or biases.
Discussion
Implications of ∃R-completeness on algorithmic challenges and heuristics for solving problems efficiently.
Proof Ideas
Reduction from ETR-Inv to Train-F2NN using gadgets representing variables and constraints.
Further Work
Investigate complexity implications on learning theory due to the computational hardness of training neural networks.
Stats
Our main result is that the associated decision problem is ∃R-complete.
We prove that algebraic numbers are required as weights for optimal training instances.