Feature visualization
Feature visualization is a technique in machine learning that helps interpret complex models by visualizing the features learned by the model.
Technology & Innovation
ColBERT (contextualized late interaction over BERT)
ColBERT (Contextualized Late Interaction over BERT) is a neural information retrieval model that leverages BERT embeddings combined with a novel late interaction mechanism to efficiently rank documents based on query relevance. It aims to balance retrieval effectiveness with computational efficiency.
Top-k sampling
Top-k sampling is a probabilistic method used in natural language processing to generate text by selecting the next word from the k most likely options. It balances creativity and coherence in language models by limiting the choice to a subset of probable candidates.
EMNIST
EMNIST (Extended MNIST) is a dataset for handwritten character recognition, extending the original MNIST dataset with additional characters and complexity.
Offline-to-online RL fine-tuning
Offline-to-online RL fine-tuning refers to the process of enhancing reinforcement learning models trained on offline data by further training them online.
Domain adaptation (machine learning)
Domain adaptation is a subfield of machine learning focused on adapting models trained on one domain to perform well on a different but related domain.
t-SNE
t-SNE (t-distributed Stochastic Neighbor Embedding) is a machine learning algorithm used for dimensionality reduction and data visualization, particularly effective for high-dimensional datasets. It maps complex data into a lower-dimensional space while preserving local similarities.
HumanNeRF (free-viewpoint rendering of humans)
HumanNeRF refers to a class of neural radiance field techniques specialized for free-viewpoint rendering of human subjects. It enables photorealistic 3D reconstruction and novel viewpoint synthesis of dynamic humans from multi-view images or video sequences.
YOLO (object detection)
YOLO (You Only Look Once) is a real-time object detection system that identifies and classifies multiple objects within an image or video frame using a single neural network. It is known for its speed and efficiency compared to traditional detection methods.
Equivariant neural network
Equivariant neural networks are a class of neural network architectures designed to respect symmetry transformations of input data. These networks maintain equivariance under group actions, meaning that transformations applied to inputs correspond predictably to transformations in the output, enabling more efficient learning and better generalization for certain tasks.