Słownik AI
Kompletny słownik sztucznej inteligencji
Neural Turing Machine
Neural network architecture combining an RNN controller with an addressable external memory, enabling the learning of complex algorithmic tasks.
Addressable External Memory
External memory matrix external to the main neural network, accessible via attention mechanisms to store and retrieve information.
Controller Network
Neural network (typically RNN or LSTM) that generates read/write operations and addressing vectors for the external memory.
Read Head
Mechanism that allows the controller to read information from external memory by generating attention weights over memory locations.
Write Head
Component that performs write operations in external memory by combining erase and addition vectors with addressing weights.
Content-based Addressing
Memory addressing method based on the similarity between a key vector and the contents stored in each memory location.
Location-based Addressing
Addressing technique using shifts and interpolations to sequentially navigate through external memory.
Interpolation Gate
Parameter controlling the blending between previous addressing weights and new weights calculated via content-based addressing.
Shift Weighting
Probability distribution defining possible movements of the read/write head in external memory.
Memory Matrix
Two-dimensional data structure storing memory vectors, where each row represents an addressable memory location.
Erase Vector
Vector used by the write head to selectively erase information in memory locations targeted by addressing weights.
Add Vector
Vector added to selected memory locations to insert new information during the write operation.
Key Vector
Vector generated by the controller to search for similar content in memory during content-based addressing.
Key Strength
Scalar parameter controlling the concentration of attention during similarity search in external memory.
Read/Write Weightings
Probability distributions over memory locations determining the intensity of read and write operations.
Differentiable Memory
Memory designed so that all operations are differentiable, allowing training of the complete system through backpropagation.
Temporal Link Matrix
Advanced memory structure that tracks the chronological order of writes to facilitate sequential data navigation.