Cloud-based ‘Deep’ Learning Tool Explains Predictions
“Deep” machine learning technology said by its creator to have been used by large customers over the last five years to make high-stakes advertising decisions is being released to the masses as a cloud web application and an API.
St. Louis-based Rice Analytics said this week its automated RELR deep machine-learning tool differs from conventional artificial intelligence approaches by providing explanations for the reasons behind its predictions. A cloud-based version of the tool called SkyRELR is said allow machine learning that extends beyond other forms of AI and neural networks by drawing on “deeper patterns that are hidden to other forms of artificial intelligence.”
The other selling point, the startup said, is transparency: “Deep learning is not a black box,” the company stressed in a statement.
RELR (pronounced “reller”) is short for “reduced error logistics regression,” which is described as a neuromorphic algorithm designed to model the deep explicit and implicit learning mechanisms of neurons. Rice Analytics said its secret sauce is leveraging RELR’s ability to provide “explanatory models” in the form of explicit models that can be used to interpret the reasons for predictions.
The startup notes that RLER has been used in “production environments” for more than five years, primarily as a tool for targeting advertising for large consumer brands.
The SkyRELR implementation is touted as yielding models that are equivalent to a four-layer-deep learning artificial neural network. “RELR models the deep learning that is believed to occur within neurons, as 21st century neuroscience now believes that substantial deep learning likely occurs within neurons,” the startup stressed on its web site.
Another advantage is full automation, eliminating the need to “train” the deep learning machine. It also means enterprise customers don’t have to be data scientists to use the cloud-based version or the accompanying API. Hence, the startup claims RELR’s explanatory applications should extend beyond such business applications as targeted advertising to include scientific, medical and even public policy applications.
RELR technology is based on Rice Analytics’ patented error-reduction methods designed to ensure that predictions are stable and that learning is automated and free of human modeler bias. Those attributes result in predictions that can be easily generalized and replicated.
Hence, the company claims “RELR is much easier to support in production environments than its deep learning black box competitors because it is hard to pin down a reason for poor model performance when the reason for predictions is not known.”
Company founder Daniel Rice literally wrote the book on leveraging deep learning techniques that allow machine learning based on patterns often hidden from other forms of artificial intelligence.
Now that the deep learning technology has been wrung out in targeted advertising campaigns, the cloud-based version is intended to open up the technology to other enterprise customers who are not data scientists.