Neural Turing Machines

Neural Turing Machines

Neural Turing Machines (NTMs) are a type of artificial neural network model that extends the capabilities of standard neural networks by coupling them with external memory resources. They were introduced by Google’s DeepMind in 2014, aiming to enhance the ability of neural networks to store and access information over long periods, thereby improving their problem-solving capabilities.


NTMs are composed of two primary components: a neural network controller and a memory bank. The neural network controller is responsible for reading from and writing to the memory bank, while the memory bank stores information that the controller can access and manipulate. This architecture allows NTMs to learn complex, algorithmic tasks that traditional neural networks struggle with, such as sorting and copying sequences of data.

How it Works

The neural network controller interacts with the memory bank using read and write heads. These heads are trained to perform operations on the memory bank, such as reading data, writing data, and erasing data. The controller decides what to read or write based on its current state and the input it receives. The memory bank, on the other hand, is a large, addressable matrix where each cell can store a vector of information.

The controller uses a form of soft attention mechanism to read and write to the memory. This mechanism allows the controller to access a weighted combination of locations in the memory bank, rather than a single location. This soft addressing makes the operations of the NTM differentiable, which is crucial for training the model using gradient descent.


NTMs have shown promise in a variety of tasks that require long-term memory and complex data manipulation. These include tasks like algorithmic tasks, one-shot learning, reinforcement learning, and natural language processing. Their ability to learn algorithms from data makes them particularly useful in areas where explicit algorithmic solutions are hard to formulate.

Advantages and Limitations

One of the main advantages of NTMs is their ability to learn to store, retrieve, and manipulate information in their memory over long periods. This makes them more powerful than traditional recurrent neural networks (RNNs) and Long Short-Term Memory (LSTM) networks for certain tasks.

However, NTMs also have their limitations. They are more complex and computationally expensive than standard neural networks. Training NTMs can be challenging due to issues like vanishing and exploding gradients, and they require careful initialization and optimization.

Future Directions

Research on NTMs is ongoing, with many researchers exploring ways to improve their efficiency and ease of training. Some of the future directions include developing more efficient training algorithms, improving the architecture of the memory bank, and exploring new applications for NTMs.

In conclusion, Neural Turing Machines represent a significant step forward in the development of neural networks, offering a powerful tool for tasks that require complex data manipulation and long-term memory. Despite their challenges, they hold great promise for advancing the field of artificial intelligence.


  1. Graves, A., Wayne, G., & Danihelka, I. (2014). Neural Turing Machines. arXiv preprint arXiv:1410.5401.
  2. Graves, A., Wayne, G., Reynolds, M., Harley, T., Danihelka, I., Grabska-Barwińska, A., … & Hassabis, D. (2016). Hybrid computing using a neural network with dynamic external memory. Nature, 538(7626), 471-476.