AI Confidential

Green eyes, pointy ears and whiskers. To a person, it’s clearly a photo of a cat. Then researchers fed it a leading-edge artificial intelligence system published by Google and trained to recognize animals. The software responded that it was 99% certain the photo was … guacamole.

Confusing a furry mammal with guacamole is not the sort of mistake you might expect for AI, considering it’s the technology that we are told will soon drive our cars and diagnose our diseases. But to those of us who work in artificial intelligence – trying to make that more intelligent future real -this sort of “fail” is not altogether surprising.

All AI solutions are not created equal, but by educating yourself, asking the right questions, and carefully evaluating solutions against the problem your organization is trying to solve, you can find a system that can tell the difference. Here are some little-known AI secrets that will help anyone interested in the field understand its challenges, as well as how Hyperscience solves for these challenges with our proprietary ML solution:

Machine Learning (ML) systems can learn the wrong things. “Garbage in, Garbage out,” was an early aphorism of the mainframe era that warned users not to believe the seeming precision of a computer print-out based on incomplete or inaccurate information.The problem is greater, but harder to detect, with Machine Learning systems. If the training data used to construct ML models is skewed or limited, the system will offer skewed results.

Sometimes it’s easy to catch these problems, like the system that confused a cat for guacamole. Other cases are harder to detect. Last year, Amazon discovered that its AI system meant to pick out the best job applicants had learned the pervasive prejudice against women in engineering, and as a result, had to abandon the system.

Hyperscience takes an ML-driven approach to a specific problem of data extraction and document processing. Our solution uses proprietary ML models (primarily Deep Learning) that we have developed and trained ourselves to deliver out-of-the-box performance across diverse document types and text inputs. In addition, a series of built-in product features allow our solution to generate relevant training data and continuously train on an organization’s specific data – behind their firewall – to drive performance improvements with minimal human intervention.

AI requires humans to be successful. Getting efficient, reliable results from AI systems such as Machine Learning is based on a collaboration between people and machines. This is known as involving humans-in-the-loop. Humans not only train the computer model, but they also provide on-going feedback to refine the automated system and handle cases where the model is uncertain. Once you accept that there’s no silver bullet AI solution, the key is to investigate when and how an organization’s employees are involved. Make sure that the work it asks a person to do is minimal to maintain the highest level of automation possible.

The amount of prep work involved can be astounding. The power of today’s AI systems is they can find patterns in data from many varied sources. But this only works if the data is accurate and consistently presented. One study found that 80% of the time spent on Machine Learning projects is devoted to collecting, cleaning and labeling data. Only 20% of the time goes into actually building and running the AI models themselves.

Unlocking business value from AI and ML involves a clearly defined problem (such as extracting data from A/P invoices or automating the life insurance underwriting process) and constant iteration. Open source AI software, as an example, likely involves significant work and custom development, including gathering and cleaning data, testing models, checking results, and making improvements, to drive any meaningful automation. With Hyperscience, the work has been done for you. By training and developing our own ML models to read a variety of document types (everything from W-9 forms to mortgage applications to invoices and checks), we’re able to deliver over 85% automation and 95% accuracy on Day 1 with additional performance improvements over time.

Ready to learn more? Download The New Science of Process Automation E-Book to discover what the latest advances in AI and ML technologies can mean for your organization.

Credits

Copy: Saul Hansell

Back to Articles