Neural Structured Learning | TensorFlow

Abstraction: TensorFlow framework for training with explicit graph or adversarial structure

Key points:

• Neural Structured Learning (NSL) trains neural networks using structured signals alongside feature inputs; structure can be an explicit graph or implicit adversarial perturbations
• Leveraging structured signals allows learning from both labeled and unlabeled data, improving accuracy when labeled data is scarce
• Models trained with adversarial perturbations gain robustness against malicious attacks
• NSL unifies Neural Graph Learning and Adversarial Learning under a single TensorFlow framework with Keras APIs, TF ops, and graph-building tools
• Structured signals are used only during training; inference/serving performance is unchanged

Connections: Tensorflow · Google · Graph Neural Networks · Adversarial Training · Semi Supervised Learning

Source: https://www.tensorflow.org/neural_structured_learning