KI-Glossar
Das vollständige Wörterbuch der Künstlichen Intelligenz
Longformer
Transformer architecture using a combination of local sliding window attention and global attention to efficiently process very long sequences with linear complexity.
BigBird
Model implementing sparse attention through three patterns: local, global, and random attention, allowing processing of sequences up to 4096 tokens with theoretical preservation of universal properties.
Sliding Window Attention
Technique where each token only attends to a fixed number of neighbors in a sliding window, reducing complexity to O(n*w) where w is the window size.
Dilated Sliding Window
Variant of sliding window attention using jumps (dilation) to increase the receptive field without increasing computational complexity.
Global Attention
Mechanism where certain predefined tokens (like [CLS] tokens) can attract attention from all other tokens, allowing information propagation across the entire sequence.
Random Attention
Approach where each token randomly attends to a subset of distant tokens, preserving long-distance connections with low computational overhead.
Pattern-based Attention
Strategy applying predefined sparse attention patterns (like fixed or learned patterns) to determine which query-key pairs to compute.
Linear Complexity Attention
Class of attention methods reducing algorithmic complexity from O(n²) to O(n), enabling scaling for very long sequences.
Kernel-based Attention
Approach using kernels to approximate softmax attention, enabling linear complexity calculations through techniques like FAVOR+ (Fast Attention Via Positive Orthogonal Random Features).
Low-rank Approximation
Technique approximating the attention matrix through low-rank decomposition, significantly reducing memory and computational requirements.
Clustering-based Attention
Method that first groups tokens into similar clusters then applies attention at the cluster level, reducing the number of required computations.
Routing Attention
Mechanism that learns to route queries to the most relevant keys using content-based routing functions, avoiding unnecessary computations.
Reformer
Architecture using locality-sensitive hashing (LSH) to limit attention computations to the most similar pairs, with quasi-linear complexity in sequence length.
Performer
Model based on FAVOR+ attention that efficiently approximates softmax attention through positive orthogonal random features, enabling linear complexity.
Linformer
Architecture that projects the key-value matrix into a lower-dimensional space, transforming complexity from O(n²) to O(n*k) where k << n.
Routing Transformer
Model using k-means based routing to group tokens and apply attention selectively, optimizing computations for long-distance dependencies.
Sinkhorn Sorting
Algorithm using Sinkhorn iteration to transform attention into a differentiable permutation, applied in sparse attention architectures.
Efficient Attention
Paradigm encompassing all attention variants aimed at reducing computational complexity while preserving the modeling capabilities of Transformers.