Paper
Neural Programmer-Interpreters
First arXiv submission: · ICLR 2016
Learns neural programs that share a recurrent interpreter and a program memory.
What it does
Learns from execution traces and composes subprograms for addition, sorting, and 3D-object canonicalization.
- Conditioning
- A program identifier, arguments, and observations of the environment.
- Program form
- Program embeddings stored in a key-value memory.
- Execution
- A recurrent interpreter calls subprograms and takes actions in the environment.
Related work
PAW cites this work on learned program representations and shared interpreters.
Limitations
Training requires execution traces and task-specific interfaces.
Related terms: neural programmer-interpreter, neural programs, program memory, compositional execution.
Cite this paper
Published version
Scott Reed, Nando de Freitas. Neural Programmer-Interpreters. International Conference on Learning Representations, 2016.
Publication BibTeX · Publication CSL-JSON
ArXiv preprint
arXiv:1511.06279, initial submission 2015-11-19.