Neural Turing Machines
E899070
Neural Turing Machines are a class of neural network architectures that augment standard networks with differentiable external memory, enabling them to learn algorithmic and sequence-based tasks in a manner analogous to Turing machines.
Statements (48)
| Predicate | Object |
|---|---|
| instanceOf |
differentiable computer
ⓘ
memory-augmented neural network ⓘ neural network architecture ⓘ |
| comparedTo |
Gated Recurrent Unit
NERFINISHED
ⓘ
Long Short-Term Memory NERFINISHED ⓘ |
| controllerType | recurrent neural network ⓘ |
| describedInPaper | Neural Turing Machines NERFINISHED ⓘ |
| developedAt | Google DeepMind NERFINISHED ⓘ |
| field |
deep learning
ⓘ
machine learning ⓘ neural computation ⓘ |
| hasAuthor |
Alex Graves
NERFINISHED
ⓘ
Greg Wayne NERFINISHED ⓘ Ivo Danihelka NERFINISHED ⓘ |
| hasComponent |
controller network
ⓘ
read head ⓘ write head ⓘ |
| hasExternalMemory | differentiable memory matrix ⓘ |
| hasLimitation |
difficulty scaling to large problems
ⓘ
training instability ⓘ |
| hasMemoryAccessMechanism | soft attention over memory locations ⓘ |
| hasProperty |
can generalize to longer sequences than seen during training
ⓘ
can learn algorithms from data ⓘ end-to-end differentiable ⓘ |
| hasPublicationYear | 2014 ⓘ |
| hasSuccessor | Differentiable Neural Computer NERFINISHED ⓘ |
| isInspiredBy |
Turing machine
NERFINISHED
ⓘ
neural networks ⓘ |
| memoryAddressingType |
content-based
ⓘ
hybrid addressing ⓘ location-based ⓘ |
| optimizationObjective |
sequence prediction loss
ⓘ
task-specific loss ⓘ |
| relatedConcept |
differentiable memory
ⓘ
memory-augmented neural networks ⓘ neural RAM NERFINISHED ⓘ |
| supportsOperation |
content-based addressing
ⓘ
differentiable read ⓘ differentiable write ⓘ location-based addressing ⓘ |
| trainedWith |
backpropagation through time
ⓘ
gradient descent ⓘ |
| usedFor |
algorithmic tasks
ⓘ
associative recall task ⓘ copy task ⓘ repeat-copy task ⓘ sequence-based tasks ⓘ sorting task ⓘ |
Referenced by (1)
Full triples — surface form annotated when it differs from this entity's canonical label.
subject surface form:
Neural Turing Machines