1-Bit AI Models

In conventional AI inference, values are represented as floating-point numbers — a weight might be 0.347, an activation might be −1.892. The model's prediction emerges from thousands of such multiplications and additions applied in sequence.

1-bit processing represents values as single bits: 0 or 1. The operations that combine them — AND, OR, NOT, XNOR — are the elementary operations of digital logic itself. Any processor — from a server CPU to a sub-1MHz microcontroller — handles these operations natively and at maximum speed.

The critical distinction is that 1-bit processing in the context of Logic-Based Networks is not quantisation. Quantisation takes a neural network trained on floating-point values and rounds its weights to lower precision. Logic-Based Networks are trained from the ground up in a binary representation. Data binarisation converts raw sensor inputs into binary features before training begins. The model learns to classify on binary inputs using binary logic. There is no floating-point precision being discarded because there was never floating-point precision to begin with.

Consider what happens when a processor executes a 32-bit floating-point multiply. It requires a dedicated floating-point unit (FPU), multiple clock cycles, and substantially more energy than an integer operation of equivalent precision. Many embedded processors — particularly lower-cost Arm Cortex-M0, M0+, and M1 cores — have no FPU at all. Floating-point operations on these devices are emulated in software, making neural network inference on them not merely slow, but impractical.

A bitwise AND or OR operation, by contrast, executes in a single clock cycle on every processor in common use. There is no dedicated hardware required, no emulation fallback, no precision-dependent penalty.

For 1-bit AI inference, an entire clause — a logical rule combining multiple conditions — can be evaluated in a handful of clock cycles. An inference that would require thousands of floating-point operations on a neural network can be completed with a small number of bitwise operations on an LBN.

This is why the benchmark numbers are what they are: 54× faster inference, 52× less energy per inference compared to equivalent neural networks, as measured by MLPerf Tiny. The gap is not surprising when you trace it back to the hardware arithmetic. It is a predictable consequence of computation type.

The reason 1-bit processing is viable as an AI technique — rather than merely a compression hack — is that real-world classification problems can be expressed in propositional logic without meaningful loss of accuracy.

Propositional logic allows a model to express rules of the form: "if condition A is true AND condition B is true AND condition C is false, then output is class X." These are not approximations of some underlying floating-point truth. They are the actual learned patterns in the data, expressed in a form that maps directly to binary computation.

Logic-Based Networks learn these clauses through a reinforcement-based training process. Clauses that predict correctly are reinforced; clauses that predict incorrectly are suppressed. Over training iterations, the network converges on a set of logical rules that characterise the classes in the training data.

The output is a model composed entirely of propositional clauses. Inference is the evaluation of those clauses against a binarised input. The computation is bitwise throughout — from input preprocessing to output classification.

Inference on microcontrollers

An Arm Cortex-M4 running at 80MHz can execute an LBN inference in microseconds. The same task — anomaly detection on a vibration sensor, for example — would take hundreds of milliseconds on a comparable neural network. LBNs run on all of them, with no hardware modifications and no software runtime to deploy.

Battery-powered AI

Energy per inference determines how long a battery-powered device can operate. At 52× lower energy per inference, an LBN-equipped sensor running on a coin cell can sustain continuous inference for months where a neural network equivalent might sustain it for days.

No GPU, no NPU, no specialist hardware

1-bit processing needs none of this. Any standard microcontroller or low-power processor becomes a capable inference engine. The hardware bill of materials stays lean, and the AI deployment does not depend on specialist silicon.

Binary neural networks (BNNs) are neural networks in which weights and/or activations are constrained to ±1 values — a form of extreme quantisation applied to a neural network architecture. BNNs still use a neural network training procedure (backpropagation), still produce a weight-based model, and still require a neural network inference runtime. The 1-bit constraint improves efficiency relative to full-precision networks, but BNNs retain the fundamental architectural properties of neural networks: opacity, sensitivity to weight precision, and dependence on the network's layer structure.

LBN 1-bit processing is different in kind, not degree. The model is not a neural network with binary weights; it is a logical structure that was never a neural network. The training procedure is different, the model representation is different, and the inference procedure is different. The term "1-bit processing" in the LBN context refers to the full-stack binary nature of the computation: binary inputs, binary clauses, binary outputs, bitwise evaluation.

The distinction matters practically. BNNs still require a neural network deployment stack and still carry the interpretability limitations of neural architectures. LBN-based 1-bit processing produces an inherently explainable model with no neural network infrastructure required.