Inspired by the way the human brain understands language
by Max Smolaks 14 November 2019
Austrian startup Cortical.io has announced its first hardware product, an enterprise appliance that can filter and route messages by understanding the semantic content – the meaning and intent.
The platform uses unsupervised machine learning to establish what kind of information users need, and can then filter out not just traditional spam, but also marketing communications, notifications, group messages and any other email content that’s not useful to the recipient.
“The goal is to reduce the wasted efforts of handling irrelevant or misdirected emails by first line business operations – including support, sales, purchasing,” explained Steve Levine, CMO at Cortical.io
The technology developed by the company differs from traditional natural language understanding (NLU) methods – it builds up “fingerprints” of individual words and sentences, based on their meaning, which is defined by their connections to other words.
Cortical.io calls this approach “semantic supercomputing” – and the co-founder and CEO Francisco Webber told AI Business the technology has countless applications, from analyzing contracts and other business documents to fighting hate speech online.
Words are not numbers
Cortical.io was founded in 2011 to commercialize an approach to NLU called semantic folding, inspired by neuroscience. Unlike traditional statistical analysis NLU methods, it encodes the semantics of natural language text in a sparse distributed representation called a semantic fingerprint.
“Ever-increasing unstructured data is overwhelming the world and the available processing power and current statistical approaches to deal with it,” Webber said. “We are taking the concept of supercomputing to the next level with the introduction of semantic supercomputing and the ability to deliver real-time processing of semantic content.”
The company’s Messaging Classification Appliance came into existence thanks to a partnership with server vendor Supermicro. The appliance is based on Xilinx Alveo field programmable gate arrays (FPGAs) – accelerator cards that can be reconfigured to speed up very specific workloads.
According to Cortical.io, the machine only needs a small number of emails to train and can easily master the specialized vocabulary of any business domain – from finance to medicine.
Once trained, the appliance can work across multiple languages, including Cantonese and Mandarin Chinese. It is expected to go on sale in Q1 2020.
“Our goal is to make possible the broad implementation and deployment of AI solutions for automating business processes and solving the most challenging use cases that depend on human understanding, decision making and execution,” Webber said.