PATEGANOptions<T>

Options & Configuration · AiDotNet.Models.Options

Configuration options for PATE-GAN (Jordon, Yoon and van der Schaar, ICLR 2019), a GAN whose discriminator is a student trained on generated samples labelled by a noisy vote of teachers, each of which sees only one disjoint partition of the data.

Inheritance: ModelOptions → NeuralNetworkOptions → ModelHyperparameterOptions → FinancialNeuralNetworkOptions → RiskModelOptions<T> → SyntheticDataGeneratorOptions<T> → PATEGANOptions<T>

For Beginners

PATE-GAN never lets one network see all the real data. Each “teacher” sees a small slice; a “student” learns only from the teachers’ noisy votes about generated rows; the generator learns only from the student. Training stops once the privacy budget ε is spent.

How It Works

Defaults follow the paper’s appendix (“Hyper-parameter optimization”): depth-3 generator and student with d and d/2 hidden units (d the encoded row width), depth-1 teachers, ReLU hidden layers and sigmoid outputs, batch size 64, n_T = n_S = 5, Adam with learning rate 1e-4. The privacy budget (ε = 1, δ = 1e-5), the noise size λ = 1 and k = 10 teachers are the defaults of the authors’ released code; the paper selects k by cross-validation from {N/10, N/50, …, N/10000}.

Constructors

Constructor Summary
PATEGANOptions()

Properties

Property Summary
double Delta { get; set; } Gets or sets δ of the (ε, δ) guarantee.
double Epsilon { get; set; } Gets or sets the privacy budget ε; training stops once the moments accountant’s ε̂ reaches it (Algorithm 1).
double LaplaceScale { get; set; } Gets or sets the scale λ of the Laplace noise added to each vote count (Eq.
double LearningRate { get; set; } Gets or sets the Adam learning rate of every network.
int BatchSize { get; set; } Gets or sets the mini-batch size n for the teachers, the student and the generator.
int Epochs { get; set; } Gets or sets the maximum number of generator iterations; training also stops when ε is spent.
int MomentOrders { get; set; } Gets or sets the number L of moments the accountant tracks (Algorithm 1, line 2).
int NumTeachers { get; set; } Gets or sets the number k of teachers, each trained on one of k disjoint partitions of the data.
int StudentSteps { get; set; } Gets or sets the number n_S of student updates per generator iteration.
int TeacherSteps { get; set; } Gets or sets the number n_T of teacher updates per generator iteration.
int? EmbeddingDimension { get; set; } Gets or sets the width of the generator’s noise z ~ Unif([0, 1]^n); null uses the encoded row width d, as the paper does (section 4.2.2 relies on G([0, 1]^d) covering [0, 1]^d).
int[] GeneratorDimensions { get; set; } Gets or sets the generator’s hidden widths; null uses the paper’s [d, d/2].
int[] StudentDimensions { get; set; } Gets or sets the student’s hidden widths; null uses the paper’s [d, d/2].
int[] TeacherDimensions { get; set; } Gets or sets the teachers’ hidden widths; empty is the paper’s depth-1 teacher (a logistic regression).