Free Information Center
Neural radiance flow (NeRF) is a computer vision technique that models dynamic 3D scenes by combining neural radiance fields with scene flow estimation, enabling realistic rendering of complex motions and deformations from sparse input views.
Expectation–maximization for deep learning is a computational approach that integrates the EM algorithm into neural network training to handle incomplete data and latent variable models. It iteratively estimates missing data and optimizes model parameters, improving learning in unsupervised and semi-supervised contexts.
The F1 score is a statistical measure used to evaluate the accuracy of a binary classification test by considering both precision and recall. It represents the harmonic mean of precision and recall, providing a balanced metric especially useful when the class distribution is uneven.
AlphaFold is an artificial intelligence system developed to predict protein structures from amino acid sequences with high accuracy. It represents a significant advancement in computational biology, enabling faster and more reliable protein modeling than previous methods.
KPConv (kernel point convolution) is a convolutional operation designed specifically for deep learning on 3D point clouds. It enables the extraction of geometric features directly from irregular point data, improving tasks such as 3D shape understanding and segmentation.
Constrained decoding is a technique in natural language processing and machine learning that restricts the output of a generative model to satisfy specific constraints. This method is used to guide the generation process to produce valid, relevant, or contextually appropriate outputs according to predefined rules or conditions.
Glow is a type of generative model based on normalizing flows, designed for efficient and exact likelihood estimation and high-quality image synthesis. Developed by researchers at OpenAI, Glow offers invertible transformations enabling both data generation and density estimation.
Unsupervised learning is a machine learning paradigm that analyzes data without pre-labeled responses to identify patterns and structures.
The projected gradient descent (PGD) attack is a widely used adversarial method in machine learning that iteratively perturbs input data to fool neural networks while ensuring perturbations remain within a specified constraint. It is a cornerstone technique for testing model robustness against adversarial examples.
Deep operator network (DeepONet) is a machine learning framework designed to learn nonlinear operators mapping between infinite-dimensional function spaces. It extends deep learning approaches to functional inputs and outputs, enabling efficient approximation of complex operators arising in scientific computing and engineering.