What does DNC stand for?
| Full form | Differentiable Neural Computer |
|---|---|
| Category | Technology & computing |
| Type | Initialism |
The initialism DNC stands for Differentiable Neural Computer in artificial intelligence (AI) and machine-learning research. DeepMind introduced the named architecture as a neural-network controller combined with an external memory that it can read and write. It is a model design, not a name for every computer with memory.
Separate the controller, the stored information and the operation that changes or retrieves it. The researchers’ account describes memory locations holding information and a controller using reads and writes while producing outputs. To make one small distinction visible, our teaching trace uses ordinary named slots rather than a trained neural model. Write red into slot A, write blue into slot B, then write green into A. Reading A now returns green; reading B returns blue. A’s earlier value was replaced in this toy state. That replacement does not mean the program learned a new rule, and an unwritten slot has no supplied value to retrieve. The worksheet below lets you vary this discrete trace and inspect which value survives. It is deliberately much simpler than a DNC: it has no neural controller, optimization, learned addressing or vector memory. Its purpose is to explain storage and overwrite behavior before reading a model diagram. A real research claim still needs the model, task and evaluation described by its authors. Do not turn a successful toy read into a claim that a neural network was trained or that a chatbot remembers a particular user across conversations.
Meaning in context
- Differentiable Neural Computer
- The memory-augmented architecture named by its researchers.
- Controller and external memory
- The controller and the locations it reads or writes have different roles.
Practical examples
Read the latest supplied value
- Before
Toy trace: A=red; B=blue; A=green. Draft: reading A must still return red.
- After
After the final write, reading A returns green; B retains blue.
The second write to A replaces its previous value in this discrete trace.
Keep the demonstration bounded
- Before
Toy result: a dictionary-style worksheet returns green for A. Draft: this proves a neural controller learned the task.
- After
The worksheet followed explicit writes and a read; it did not train or run a neural controller.
A correct deterministic result cannot establish an unperformed learning process.
Original illustrative examples, not quotations or reports of actual events.
Trace a tiny memory
What remains in a memory slot after another write, and what has this toy exercise actually demonstrated?
See all guidance
Read a model diagram
Separate the controller from memory locations.
The controller reads and writes the external memory.
Trace an overwrite
Track the most recent write to each named slot.
A=green replaces the earlier A=red in this toy trace.
Describe an experiment
Name what was actually executed.
The worksheet ran discrete slot operations, not neural training.
Distinguish a stored value, a read and an overwrite; a discrete memory trace is an explanation of those operations, not a trained DNC.
Documented origin and history
The researchers’ introduction of the named DNC architecture is dated October 12, 2016.
This dates the cited introduction, not the first-ever neural network with memory or the origin of all memory-augmented learning.
Researchers introduce the named architecture
Help improve this entry
Spot a mistake or have a clearer example? Suggest a change.
By AcronymIndex Editorial Team · Editorial policy