Legendre Memory Units

If the order of your data is important for understanding it, then the LMU is for you.

The LMU provides the best possible compression and representation of information over time, allowing you to get state-of-the-art results with smaller networks and less training.

Purchase license

Two white buses parked along the roadside.

Achieve state-of-the-art time series processing

The LMU is ABR’s proprietary advantage, making large-scale time series AI at the edge possible. Take a look at our published and easily reproducible results below.

The LMU outperforms current technology. LSTMs are more trainable than vanilla RNNs, but don’t scale beyond 500 to 1000 samples. LMUs can handle over a million samples. Transformers are more scalable than LSTMs on GPUs, but still limited by bad scaling at O(N²). LMUs boast O(N) compute and O(1) memory.

About Legendre Memory Units

The LMU has been recognized by DeepMind and others as the first state space neural network. It is proven to be not just better, but optimal, for solving time series problems. This allows us to bring down power demands and energy costs while remaining highly accurate, even with less data and smaller models. In fact, the LMU has proven so successful that we developed a TSP Chip specially built to run it for even greater efficiency gains.

With the LMU, we are bringing a wide range of features, from full speech recognition to advanced signal processing, to the edge. See our case studies on how this has already been done for keyword spotting in hospitals and smart garments monitoring heart rates.

Click here for a non-technical introduction to LMUs

Cutting edge research on LMUs

Legendre Memory Units

The first published paper on LMUs — our state-of-the-art algorithm, and a new kind of neural network

Hardware Aware Training for Efficient Keyword Spotting

LMU-based keyword spotting achieves state-of-the-art accuracy and low parameter counts

Language Modeling using LMUs

LMUs outperform transformers on language modeling, with better data efficiency and improved scaling

Parallelizing Legendre Memory Unit Training

Up to 200 times faster training with a simplified LMU variant that can be run in parallel

Introducing the LMU to the San Francisco Bay ACM

Discussing the LMU in-depth with DataSig

A deep dive on the LMU for speech and language processing

Harness LMUs yourself

The LMU is a product of the cutting edge research from ABR, and intellectual property of the company. Commercial licenses for our patented technology can be purchased below. Non-commercial and academic licenses are free.

Purchase license

Talk to our sales team

Learn more about how low power edge computing can support your AI integration.