Video Diffusion Models (VDM) are generative models that combine the advantages of variational autoencoders (VAEs) and diffusion models. The core idea of VDM is to perform the diffusion process in the latent space, rather than...
Hyperparameter tuning is a crucial step in machine learning model training, involving selecting the optimal combination of hyperparameters to optimize model performance. Hyperparameters are configurations set before the learning process begins...
Real-Time Engagement (RTE) is a collection of technologies, product architectures, and scenario solutions designed to enable two parties communicating and interacting remotely to realistically and seamlessly recreate the offline interaction scenario in real time...
Robotic Process Automation (RPA) is a software technology that automates business processes by simulating human behavior and interacting with digital systems and software. RPA software robots can...
Mixture of Experts (MoE) is a technique in machine learning used to build large models by decomposing them into multiple sub-networks or "experts" to improve model performance and efficiency. Each expert...
Genetic algorithms (GA) are heuristic search algorithms that simulate the process of biological evolution, belonging to the category of evolutionary algorithms. They are based on the ideas of natural selection and genetics, simulating the process of natural selection...
Generative AI is an AI technology that creatively generates new content based on existing data. It has wide applications in various fields, including text generation, image generation, audio and video generation, etc.
Contrastive Language-Image Pretraining (CLIP) is a multimodal pre-trained neural network model developed by OpenAI. It uses contrastive learning to train images and text...
Hallucinations of large models refer to the phenomenon where the content generated by a model is inconsistent with real-world facts or user input.
Embodied Intelligence (EI) is a multidisciplinary research field that involves cognitive science, artificial intelligence, robotics, neuroscience, and other areas. The core idea of embodied intelligence is that...
Training data is the dataset used to build predictive models in machine learning. It contains a series of input features and corresponding target outputs, which are used to teach the model how to perform operations based on the features...
Reward models are a core concept in reinforcement learning, used to evaluate an agent's behavior in a specific state. In large language models (LLMs), reward models guide the model by scoring the input question and answer...
Prompt engineering is an emerging discipline that focuses on developing and optimizing prompts to help users effectively utilize large language models (LLMs) for various purposes...
Model generalization is a core concept in machine learning, referring to a model's ability to perform on new data beyond its training data. In other words, generalization ability measures a model's performance on unseen samples...
In Natural Language Processing (NLP) tasks, the context window refers to the range of contextual information a model considers when processing a specific input. Specifically, it determines the scope of contextual information a model can consider simultaneously when generating or understanding text...
Dataset distillation, also known as dataset compression, is a technique designed to extract key information from large datasets and build a smaller dataset. This...
The Scaling Law describes how model performance changes with model size (such as the number of parameters), training dataset size, and computational resources used for training. It reflects the relationship between model performance and these factors...
Connectionism, also known as link theory, is a theoretical framework in cognitive science and artificial intelligence that advocates explaining cognitive processes by simulating the interconnections between neurons in the brain. This theory posits that...
The objective function is a core concept in mathematical optimization problems; it represents the functional relationship between the objective and the influencing factors. Simply put, the objective function is the function you are trying to calculate or optimize to achieve...
Slow Perception is a novel visual perception concept proposed by the Step Multimodal team. Through a progressive decomposition and flow approach, it allows models to perceive complex geometric shapes with greater finer detail, much like humans. In experiments,...
Judge models are auxiliary models used to evaluate the quality of the output of other models. They act as "referees," evaluating and scoring the output of Large Language Models (LLMs). The models are evaluated based on the input question and the...
Agentic RAG is a method that combines AI agents and retrieval augmentation generation (RAG) systems. It revolutionizes question-answering by introducing an agent framework. This differs from methods that rely solely on large models...
Forward chaining is a method in the field of artificial intelligence that uses inference engines for automated reasoning. It's a data-driven reasoning strategy that starts with known facts and derives new information by applying a series of rules...
Discriminative models are an important class of models in machine learning, primarily used for classification and regression tasks. Their core objective is to learn the mapping relationship between input variable x and output variable y, i.e., the conditional probability distribution P(y|x). (This is in contrast to generative models...)