This AI Skips the Maths. It Could Run on a Band-Aid.

Key points

• Lizy K. John's team at the University of Texas at Austin has built neural networks that replace multiplication with lookup tables, cutting energy use by up to 1,000 times on medical monitoring tasks.
• The team's arrhythmia detector, a device that watches for dangerous heart rhythms, fits on a bendable plastic chip with only about 10,000 logic gates, compared to billions on a standard chip.
• For one keyword-spotting task, their model uses 42 to 79 nanojoules per decision, against more than 5,000 nanojoules for the current leading industry model.
• The approach has already replaced roughly half of a transformer model, the architecture behind chatbots like ChatGPT, but full language-model support is still in progress.
• The technique is not new, but almost no research teams are working on it.

Every time a chatbot answers your question, a neural network, the layered mathematical system that gives AI its pattern-matching power, performs millions or billions of multiplications. Multiply an input number by a learned weight. Add it up. Repeat. For a single sentence, that arithmetic runs at a scale most people struggle to picture.

Lizy K. John thinks that is a lot of unnecessary work.

John is a professor of electrical and computer engineering at the University of Texas at Austin. For the past five years she has focused on what she calls weightless neural networks. Instead of multiplying numbers together, these networks pass simple yes-or-no signals through a series of lookup tables, small stored lists that return a pre-calculated answer instantly, the same way a printed times-table book returns 7x8=56 without making you work it out from scratch. The result is a network that can, depending on the task, run using 1,000 times less energy or fit into a chip 1,000 times smaller than a conventional model, while getting comparably accurate answers.

The work was covered by IEEE Spectrum.

Why does cutting out multiplication matter so much?

Multiplication is expensive in hardware terms: it takes transistors, time, and power. A lookup table just reads from memory, which is far cheaper. The human brain, John points out, uses about 20 watts to hold a conversation. No multiplication involved.

Her team's most striking result so far is a model that detects arrhythmia, an irregular heartbeat that can signal serious heart problems, on a chip built on a bendable plastic sheet. That plastic can only fit around 10,000 logic gates. A modern smartphone chip holds billions. A conventional AI model cannot run on something that small. Hers can.

For medical monitoring tasks including ECG readings (electrical heart signals), EEG readings (electrical brain signals), and blood pressure, her team's models match standard accuracy at a fraction of the energy. The best competing small model for one problem they studied is 17 megabytes. Their version is 14 kilobytes, roughly 1,200 times smaller. Because the model fits on the sensor itself, health data never needs to leave the device for processing. That is both an energy saving and a privacy protection.

What could patients actually see change?

In the near term, John's first target is continuous medical monitoring. She describes a sensor patch you wear like a sticking plaster for a week, one that processes readings locally and flags a problem without needing a phone or a hospital connection. Right now, most wearable sensors collect raw data and beam it wirelessly to a server because the chip on the sensor is too small and too power-hungry to run analysis itself. A weightless network flips that.

For keyword spotting, the process where a device like an Alexa speaker listens constantly for its wake word, her team's model uses 42 to 79 nanojoules per decision. The current industry-leading model uses more than 5,000 nanojoules for the same task. That is the difference between a device that lasts weeks on a battery and one that lasts days.

Can this ever reach chatbots?

Not yet, but the team is working on it. A transformer model, the architecture behind ChatGPT and most modern AI assistants, has two main sections that repeat in sequence: an attention layer, which weighs how words relate to each other, and a multilayer perceptron, which does the bulk of the number-crunching. John's team has already replaced the multilayer-perceptron half with a lookup-based equivalent. The attention layer is still unsolved.

The technology is also not new: a company in the United Kingdom built a commercial product using it in the 1980s for pattern recognition, then the approach largely vanished. Researchers at the Federal University of Rio de Janeiro kept the idea alive quietly for years. John's contribution is making it run efficiently on real hardware, first on an FPGA, a flexible chip that can be reconfigured for different tasks, and now on purpose-built substrates.

Training still happens on conventional GPUs, the specialised number-crunching chips that power most AI research. But because FPGAs already contain small lookup tables by design, John believes they could eventually handle training too, closing the loop.

The honest caveat: every result so far comes from small, specific tasks. No peer-reviewed demonstration exists yet for large-scale language models. John is candid about that gap. The confidence that it will scale simply is not there yet, outside her own lab.