How a Single Electronic Component Can Learn and Process AI Data

This patent describes a tiny electronic component called a resistive processing unit (RPU) that acts like a brain cell in an artificial intelligence network, storing and processing information directly within its changing electrical resistance.

What it covers

This patent details a two-terminal resistive processing unit (RPU) that functions as a neuron in a neural network. Its electrical resistance represents a 'weight' for that neuron, which is a crucial value in AI learning. The RPU is designed to change its resistance in a random, or 'stochastic', and 'non-linear' way based on electrical signals applied to its two terminals (Claim 1). This change in resistance allows the RPU to 'locally perform a data storage operation' (Claim 3) for the neural network's training. Crucially, it also 'locally performs a data processing operation' (Claim 1) using this changed resistance. For example, in a system learning to recognize images, each RPU could adjust its resistance based on input signals, effectively 'remembering' a part of the image, and then use that 'memory' to help process new images, all within the same tiny component.

What it doesn't cover

• —Does not cover RPUs that use more than two terminals for their operation.
• —Does not cover RPUs where the resistance changes in a predictable, non-stochastic manner.
• —Does not cover RPUs where the change in resistance is a simple linear relationship, rather than a non-linear one (Claim 1).
• —Does not cover systems where data storage and processing are handled by separate, distinct components, rather than locally within the RPU (Claim 3).
• —Does not cover RPUs that use encoding methods other than stochastic pulse sequences or magnitude modulation for input signals (Claims 6 and 7).

The clever bit

The novelty lies in designing a single, two-terminal resistive unit that can both store data (by changing its resistance) and process data, performing these functions locally and stochastically. This dual capability, especially the non-linear and stochastic resistance change, is key to efficiently implementing neural network 'weights' and 'training' directly in hardware.

Why it matters

This patent addresses a core challenge in artificial intelligence: how to make AI systems more energy-efficient and faster by integrating memory and processing. By allowing a single component to both store and process data, it paves the way for 'neuromorphic' hardware that mimics the brain's structure. This approach can significantly reduce the need to constantly move data between separate processing and memory units, which is a major bottleneck in traditional computer architectures.

Real-world examples

• 1.Neuromorphic computing chips
• 2.AI accelerators for edge devices
• 3.Resistive random-access memory (RRAM) arrays used for in-memory computing
• 4.IBM's AI hardware research