Hybrid computing using a neural network with dynamic external memory.

Area of Science:

• Artificial Intelligence
• Machine Learning
• Computer Science

Background:

• Artificial neural networks excel at sensory and sequence processing but struggle with complex data structures and long-term memory due to a lack of external memory.
• Existing neural network models are limited in tasks requiring variable representation and data manipulation over extended periods.

Purpose of the Study:

• To introduce a novel machine learning model, the differentiable neural computer (DNC), capable of interacting with external memory.
• To demonstrate the DNC's ability to learn and perform complex reasoning and structured tasks by leveraging its external memory capabilities.

Main Methods:

• Developed a differentiable neural computer (DNC) model combining a neural network with a read-write external memory matrix.
• Trained the DNC using supervised learning for reasoning and inference tasks, and reinforcement learning for goal-oriented sequence tasks.
• Evaluated the DNC on synthetic and real-world graph-based problems and a symbolic sequence-driven puzzle.

Main Results:

• The DNC successfully answered synthetic questions mimicking natural language reasoning and inference.
• Demonstrated learning of shortest path finding and graph link inference, generalizing to transport networks and family trees.
• Successfully completed a moving blocks puzzle with changing goals specified by symbol sequences.

Conclusions:

• Differentiable neural computers (DNCs) bridge the gap between neural networks and conventional computers by incorporating external memory.
• DNCs exhibit the capacity to solve complex, structured tasks previously inaccessible to standard neural networks.
• This advancement opens new possibilities for AI in areas requiring sophisticated reasoning, data manipulation, and long-term memory.