Is an NTM the same as a Turing machine?

No. It is a differentiable neural architecture inspired by the computation-and-memory idea, with finite practical resources.

Can it learn from input-output examples?

Yes. The original research trained it on examples of tasks such as copying and associative recall.

Is a differentiable neural computer another name for it?

No. The DNC is a related later architecture with additional memory-management mechanisms.

AI Glossary / Deep learning / Advanced

Neural Turing Machine Network

Definition

A Neural Turing Machine, or NTM, is a neural architecture that combines a trainable controller with an external memory matrix. Differentiable read and write mechanisms allow the system to learn how to use that memory while solving a task.

What it means

A conventional neural model often stores temporary information within its internal state. An NTM gives its controller a separate numerical workspace that it can address and modify during a sequence.

Graves, Wayne and Danihelka introduced the architecture in 2014 and explored simple algorithmic tasks. The name refers to an analogy with a Turing machine's combination of computation and memory. It does not mean that a trained NTM automatically discovers arbitrary programs or performs every task a conventional computer can execute.

External means separate from the controller's internal state, not necessarily a remote database or permanent storage service. A particular implementation must define memory capacity, initialization and whether state survives between tasks.

How it works

The controller receives inputs and read information, produces task outputs and determines memory-access settings. It can use a recurrent or feedforward neural network.

A read head assigns nonnegative weights across memory locations and returns their weighted combination. A write head modifies locations using weighted erase and add operations. Smooth access makes the computation amenable to gradient-based training.

Addressing can use similarity to stored content and mechanisms for moving attention between locations. Training examples provide the desired task behavior; the system learns parameters that support useful sequences of reads and writes.

Learned parameters and working memory serve different roles. The former encode how the system operates, while the latter can change during one execution. Later memory-augmented architectures, including differentiable neural computers, developed additional mechanisms; their features should not automatically be attributed to the original NTM.

Three practical examples

A research benchmark teaches a model to copy a sequence and tests whether it succeeds on lengths outside its training examples.
An educational demonstration visualizes which memory locations a controller reads while recalling an earlier association.
A modeling study compares internal recurrent state with an explicit memory workspace for a structured sequence task.

Think of it this way

Think of a worker with a reusable scratchpad. The worker's learned skill determines how to use the pad, while the notes on it change during each job. A new job may begin with a blank pad even though the skill remains.

The analogy separates parameters from temporary memory. An NTM's scratchpad is numerical, and a soft read may blend several locations rather than retrieve one exact written note. It also lacks a person's automatic understanding of what should be remembered or discarded.

Worked example

Reading from a small memory

Suppose a teaching memory has two rows: M1 = (2, 0) and M2 = (0, 4). A read weighting of (0.75, 0.25) returns 0.75M1 + 0.25M2 = (1.5, 1).

A sharper weighting of (1, 0) returns (2, 0), exactly the first row. With (0.5, 0.5), the read vector is (1, 2), which is neither stored row.

This shows why attention precision matters when a task requires exact recall. A downstream controller might still use a blended representation effectively, but the memory interface alone does not guarantee exact symbol retrieval.

The calculation assumes the memory and read weights are already given. It does not show how the controller learned those weights or how a full copy task would be trained. Exact agreement should be measured over complete outputs when the application depends on exact sequence reproduction.

Illustrative case study

A fictional research team, Birch Sequences, trains an NTM on short lists of symbols. Many predicted symbols are correct, so the team initially reports only average symbol accuracy.

A reviewer asks how often the complete list is reproduced without any error. Longer lists reveal that small per-symbol mistakes can make whole-sequence success much worse. The team adds length-based evaluation, repeated trials and memory-capacity tests.

It also documents that working memory resets for each example. The result is described as learned sequence behavior within tested conditions, rather than a permanent memory system. The case illustrates how a model's impressive local behavior can differ from the exactness required by a complete task.

Common misconceptions and mistakes

External memory is automatically persistent storage. Persistence depends on the implementation and surrounding system.
A read head always selects one location. Soft attention can combine several locations.
Learning to copy proves general-purpose programming ability. It establishes performance on the tested task and conditions.
Writing to memory retrains the controller. Working-state changes are distinct from updates to learned parameters.

Business use

NTMs are most useful to business readers as a research concept explaining how learned computation can interact with an explicit workspace. They also provide context for later memory-augmented models.

A proposed practical application should justify why learned memory access is appropriate compared with a conventional data structure or algorithm. Where exact copying, sorting or record retrieval is required, established deterministic procedures offer a clear comparison. Assess complete-task accuracy, memory limits and behavior on longer or unfamiliar inputs. A model's ability to use a scratchpad does not establish the reliability, durability or access controls of a production storage system.

Check your understanding

Answer: An NTM's memory changes while processing a sequence, but its trained controller parameters remain fixed.

Is that necessarily online training?

Answer: No. It can be ordinary execution with changing working state. Online training would involve a specified parameter-learning process.

Related terms

Controller
Attention
Recurrent neural network
Differentiable neural computer
Memory-augmented neural network
content-based addressing
differentiable memory
algorithmic learning

Learn next

Calculate the read vector for three different memory weightings and compare exact selection with blending. Then study a copy-task evaluation that reports whole-sequence success and performance across sequence lengths. Keep working-memory updates separate from parameter training.

Sources

Graves, Wayne and Danihelka, Neural Turing Machines: https://arxiv.org/abs/1410.5401 Original paper, memory and addressing equations: https://arxiv.org/pdf/1410.5401 Graves and colleagues, Hybrid computing using a neural network with dynamic external memory: https://www.nature.com/articles/nature20101 Google DeepMind, Differentiable neural computers: https://deepmind.google/blog/differentiable-neural-computers/

Previous: Neural tangent kernel Next: Neuralese