Sparse modeling AI is edging out traditional deep learning to become the technology of choice for product manufacturers and medical researchers because it ticks off all the boxes for modern quality control: explainability, energy efficiency, and speed. Just ask some of the customers from Hacarus, a Japan-based startup that’s developed a standout AI-fueled visual inspection solution.
“Sparse modeling AI is a surprising revelation for our clients that need to innovate faster while meeting high-quality standards, whether it’s electric vehicles, luxury watches, or drug discovery,” said Kenshin Fujiwara, CEO and founder of Hacarus. “They’re amazed at how we’re tackling traditional AI’s dirty secret, reducing the high energy costs of data collection and training, which saves time and our planet’s resources.”
A green, explainable AI alternative
According to some studies, training a single AI model using traditional machine learning (ML) can equal the carbon emissions of five cars across their entire lifecycles. That’s because ML algorithms attempt to “understand” every detail gleaned from huge amounts of data from scratch. In contrast, sparse modeling doesn’t require training from tens of thousands of images to yield a strong model for prediction. Because it starts with built-in assumptions, restrictions, and hypotheses, sparse modeling saves time by ignoring what’s already known. This reduces computational time and energy consumption.
On the factory floor, manufacturers need far fewer samples of both good and bad product parts to train the AI model, speeding up visual product inspections to detect defects and anomalies without sacrificing sustainability. Inside research labs, sparse modeling yields more explainable AI. For example, scientists exploring new drug treatments can more easily distinguish chemical compound reactions. In one pilot project with a pharmaceutical company in Japan, Hacarus’ solution performed 56 times faster than a deep learning algorithm.
“Sparse modeling is ideal for any precision engineering equipment or research company developing advanced products with less data,” Fujiwara. “Electric vehicle parts are a great example because it’s a brand new sector. Automakers and suppliers can create a reliable, AI-based model with as little as 20 images. It delivers the equivalent results of deep learning in a fraction of the time and energy.”