KI-Glossar
Das vollständige Wörterbuch der Künstlichen Intelligenz
Multi-modal Model
AI architecture capable of processing, understanding, and simultaneously generating multiple types of unstructured data such as text, images, audio, or video in a common representation space.
DALL-E
Text-to-image generation system developed by OpenAI, combining a VQ-VAE transformer with a diffusion model to create photorealistic images from text descriptions.
Vision-Language Transformer (ViLT)
Transformer architecture that jointly processes image patches and text tokens without feature pre-extraction, enabling end-to-end learning for vision-language tasks.
Multi-modal Foundation Model
Large-scale pre-trained model on diverse multi-modal data, capable of being fine-tuned for numerous downstream tasks without requiring training from scratch.
Multi-modal Perceiver
Unified architecture that processes different modalities (text, image, audio) as token sequences in a single latent space, using cross-attentions to model their interactions.
GPT-4V(ision)
Multi-modal version of GPT-4 capable of accepting images and text as input, using a hybrid architecture to reason about visual content and generate contextualized text responses.
Audio-Visual Speech Recognition (AVSR)
Multi-modal system that combines audio signals and video (lip movements) to improve speech recognition robustness, particularly in noisy environments.
Visual Language Model (VLM)
Class of multi-modal models that extend language transformer architectures to understand and reason about visual content, often through cross-attention mechanisms.
Contrastive Pre-training Multi-modal
Self-supervised training paradigm that learns aligned representations between modalities by maximizing the similarity of positive pairs and minimizing that of negative pairs in the latent space.