One of the most common ways we learn something new is by watching an expert and imitating their actions. This basic concept is at the heart of imitation learning, a form of meta-learning in which a neural network is trained to learn new activities. Thus, the behavioral policies learned are much more flexible than traditional rules-based systems and allow us to adapt to and filter out the noise that is inherent in human-based activity.
Another form of meta-learning, few-shot detection, allows us to recognize the content of an image without seeing the multitude of training images generally associated with standard convolutional neural networks. This allows us to rapidly adapt to new, custom, or updated software that we see, a common occurrence when working with a global enterprise.
We've built our computer vision infrastructure to closely resemble how humans understand software screens, with convolutional neural networks that recognize each individual control, along with its corresponding label. This intelligence allows us to define a unique interaction model per control type and understand the difference between clicking on a button or clicking in a checkbox. By truly interpreting the screen at the field level and not just as a screen of X/Y coordinates, our AI can track noisy human interactions and transcribe them cleanly.
Natural language models provide a deeper understanding of a process that transcends traditional sequence matching. Many processes involve free-form text, which often shows up as random noise inside of a larger sequence. By leveraging word-embeddings, parts-of-speech tagging, and named-entity recognition, we add structure to what looks random, filling in the blanks that often come from working with unstructured text.
Our Virtual Process Analyst combines visual data acquisition, imitation learning, natural languge understanding, few-shot learning and data mining into a new discipline we call replication learning. Humans are multi-sensory, which allows for the rapid learning and adaptation we see in everyday life. By leveraging joint models containing visual, time-series and language information, our replication learning models can quickly learn and adapt, uncovering patterns that are hidden under the surface.
Integrations with legacy systems are one of the largest areas of friction in enterprise projects. Our technology uses only computer vision for data acquisition, negating the need for costly, complex, and expensive data integration and mapping. In addition, we've automated the data science—so after your IT group deploys our agents you can start receiving insights in weeks.
Most of today's processes don't exist within a single monolithic application, but span across multiple systems: off the shelf, proprietary, desktop, and cloud. Any discovery solution dependent on transaction logs or database access will have to be specifically mapped for each one of these applications. Cloud-based systems and desktop applications present unique challenges to log-based solutions, which may not even produce the types of logs necessary for transaction mining. Our computer vision-based approach offers 100% compatibility across legacy systems, custom applications, thin or thick clients, and green screens. If a human can see the software screen, our software can interpret it.
Process discovery via interview-based methodologies or recording software that produces screen-level granularity suffers from an inability to ensure complete process coverage. Process permutations can be missed and nuance can be lost, leading to poor quality documentation and inadequate planning. Our computer vision-based solution perfectly captures every activity at the field level, saving hours of transcription, tracing, and validation.