The Science Behind State Space Models

ABR pioneered a new class of neural network — state space models — that compress temporal dynamics into compact, real-time representations. The result: AI that is smaller, faster, and built for the edge.

Foundation

The Legendre Memory Unit

ABR's patented breakthrough that helped define an entirely new category of neural networks.

The Legendre Memory Unit (LMU) is ABR’s foundational invention — a mathematically principled architecture that uses orthogonal polynomial projections to compress continuous-time signals into fixed-dimensional state representations.

Unlike transformers, which require attention over entire sequences, the LMU maintains a compact, sliding state that captures long-range dependencies with constant memory and linear compute. This makes it inherently suited for streaming and real-time workloads.

The LMU became the first in a new class of neural network called State Space Models that are orders of magnitude more computationally efficient than popular attention-based architectures — unlocking AI capabilities on devices that were previously too constrained for meaningful intelligence.

Timeline

2019 - LMU Published
ABR introduces the Legendre Memory Unit, establishing a mathematical foundation for state space neural networks.

2020 - Patent Granted
Core LMU architecture patented, securing ABR's IP as the originator of the state space model class.

2022 - Edge-Optimized SSMs
ABR develops production-grade state space models specifically optimized for voice and time-series at the edge.

2024 - Full-Stack Deployment
Complete model-to-silicon pipeline demonstrated: training, quantization, compilation, and hardware execution.

Advantage

Why State Space Models Matter

SSMs fundamentally change the efficiency equation for temporal AI — making real-time intelligence practical on the smallest devices.

Constant Memory, Linear Compute

Unlike transformers that scale quadratically with sequence length, SSMs process inputs in constant memory with linear time complexity — critical for streaming applications.

True Real-Time Processing

State space models are inherently causal and recurrent, processing each time step as it arrives. There is no buffering, no look-ahead — just instant, streaming inference.

Dramatically Smaller Models

By compressing temporal dynamics mathematically rather than through brute-force attention, SSMs achieve comparable accuracy with a fraction of the parameters and memory footprint.

On-Device, Offline Operation

The efficiency of SSMs means complex voice and DSP workloads run entirely on-device — no cloud required. Data stays private, latency stays near zero.

Long-Range Dependencies

The Legendre basis provides a mathematically optimal compression of history, allowing SSMs to capture dependencies over thousands of time steps without degradation.

Ultra-Low Power Consumption

Fewer operations per inference step translates directly to lower energy usage — enabling always-on intelligence on battery-powered devices measured in milliwatts.

Hardware Enablement

World Record Low-Power Acceleration

ABR designed and built an optimal state space model accelerator — purpose-built silicon that executes SSM inference to demonstrate maximum efficiency. By co-designing the models and the hardware together, ABR achieved world record low power for streaming voice AI.

This accelerator demonstrates that when state space models meet dedicated hardware, the results redefine what’s possible at the edge — delivering real-time intelligence inside power envelopes that were previously considered impossible for meaningful AI workloads.

<30mW
Streaming voice AI power
Real-time
Always-on inference
Full Stack
Model-to-silicon co-design
Edge-first
No cloud dependency

Build the Future of Edge AI

ABR is a leader in AI innovation, with patented technologies and groundbreaking research. Let’s talk about what state space models can do for your product.