Are AI Models Working Harder Than They Need to?
IEEE Spectrum AI Jackie Snow
A UT Austin professor says AI models don't need all that multiplication—lookup tables can do the job instead. Her weightless networks are already 1,000x smaller on sensors, and she's eyeing chatbots next.
Lizy K. John has a pretty simple complaint about modern AI: it's doing way more math than it needs to. Neural networks, from chatbots to song recommenders, run on endless multiplication between inputs and learned weights. John, an electrical and computer engineering professor at the University of Texas at Austin, has spent five years building an alternative called weightless neural networks, which swap out multiplication for lookup tables. Instead of crunching numbers, the network just checks binary inputs against stored answers, more like flipping through an index than solving equations over and over.
The idea isn't new. A UK company built a commercial version in the 1980s for pattern recognition, and it vanished. A small group at the Federal University of Rio de Janeiro kept the concept alive quietly for decades. John got pulled in a few years ago through a friend, paired it with a student's FPGA experience, and within six months had lookup-based networks running 1,000 times smaller than conventional models on tiny chips, no GPU required.
The results on small problems are striking. For medical monitoring tasks like ECG or EEG tracking, her networks hit comparable accuracy at a fraction of the size: 14 kilobytes versus 17 megabytes for the best conventional model on one task, a difference of more than 1,000x. That means sensors can process data locally instead of streaming raw readings to a phone or server, which saves battery and keeps personal data from ever leaving the device. On keyword spotting, the kind of always-listening detection that catches a wake word like
My take
,
Read more about this at: IEEE Spectrum AI
