KI-Glossar
Das vollständige Wörterbuch der Künstlichen Intelligenz
Neural Turing Machines
Architecture combining a neural controller with a readable and writable external memory, allowing the model to explicitly store and manipulate complex information over time.
Memory Networks
Neural systems integrating a long-term memory composed of key-value pairs, allowing the model to store, retrieve, and use contextual information to improve performance in continuous learning.
Sliding Window Buffer
Memory strategy maintaining the N most recent samples, automatically eliminating the oldest ones to preserve a current temporal representation of data with constant memory complexity.
Core Set Selection
Algorithmic method selecting a minimal but representative subset of past data to keep in memory, optimizing the trade-off between performance and storage capacity in continuous learning.
Memory-Augmented Neural Networks
Neural architectures integrating addressable external memory components, allowing explicit storage and retrieval of information to solve tasks requiring long-term memory in online learning.
Episodic Memory
Memory component storing specific episodes of past experiences with their temporal context, allowing the model to recall and reuse similar situations in new online decisions.
Memory Synthesis
Technique artificially generating memory samples from past data distributions, virtually extending memory capacity without increasing physical storage.
Gradient Projection Memory
Method projecting gradients of the current task onto the orthogonal space of important directions from previous tasks, implicitly preserving past knowledge without explicit memory.
Gated Memory Units
Mechanisms controlling writing, reading, and selective forgetting in neural memories, optimizing memory space usage in continuous learning.