TRaining AI Naturally (TRAIN)

The dirty secret of AI is that it’s currently powered by armies of humans. Machines cannot read natural language like humans do, so massive efforts are underway to manually label the data necessary for training machine learning models to understand human language. By augmenting a much smaller subset of labeled data than is typically needed with traditional and cutting-edge machine learning algorithms, Jaxon creates huge quantities of training data as well as fully-trained, highly-accurate models that can immediately be used for text classification. Jaxon reduces manual labeling requirements by 90% or more, accelerating the process by months and slashing the cost by over 75%. Jaxon continuously incorporates state-of-the-art technologies from the machine learning community, including deep learning architectures, classical algorithms, and the inclusion and ensembling of any custom logic already available (a means of capturing specific information from human subject matter experts). Successful machine learning is an iterative process of finding the right mixture of feature extraction, algorithmic selection, and parameterization to fit a specific combination of data and problem. With Jaxon, customers reap the benefits of:  Speed: substantially faster time-to-label Discoverability: exposing knowledge previously untapped Cost Effectiveness: dramatically reduce manpower needed to build and maintain Robustness to Domain Shift: keeping up with an ever-changing taxonomy (data evolves) Jaxon is on a mission to eliminate the need to have humans manually label data for AI training and make neural architecture, data prep, and pipeline design a snap. This “easy button” ensures that AI projects get put into production in weeks rather than months or years. Working with the defense end-user, the first steps will be to rigorously define a problem specification and determine exactly what level of accuracy is necessary for each task that requires Jaxon. This level will then be used as the benchmark in order to assess final performance. Jaxon will estimate classifier performance using metrics such as an F1 score and a per-class confusion matrix and demonstrate its effectiveness within the Air Force using Air Force data. The Air Force is implementing an advanced Artificial Intelligence strategy that will allow for faster and more accurate decision-making in many areas. Jaxon is able to accelerate the machine learning model creation process in a variety of areas, adapting to each user’s specific problem, data, and accuracy needs. Based on Jaxon’s specialization in unstructured text, a particular pain point within AI, it is a powerful tool for any department interested in intelligence or using text of any kind to inform decision making. Jaxon’s integration into ACE would allow airmen to create fully-trained, reusable models with minimal data labeling from anywhere with just a few clicks, expanding the Air Force’s machine learning initiative exponentially.