ML has made many strides over the past several years, but in most cases the overall training methodology has remained consistent. The current ML training process mimics Kolb’s Experiential Learning paradigm found in classrooms. This model drives students to learn from personal experimentation, often without any outside instruction. This technique can provide rapid understanding, but also has the drawback of not being able to take advantage of knowledge and experience provided by expert guidance.
Constructivism is a broader learning theory which incorporates the addition of knowledge gained from past experiences as well as social interaction and collaboration with an expert. This allows for students to learn from an instructor’s past experience and knowledge as a supplement to the experimentation process.
BALANCED’s HEWMEN platform utilizes Constructivist Augmented Machine Learning (CAML) methodology that allows for humans to interact with ML algorithms and techniques. CAML is a human-in-the-loop methodology scaled by using human computation video game techniques. This process allows algorithm guidance by augmenting inputs as well as directly modifying hidden layers, weight and connections throughout the training process. Humans are capable of identifying patterns and optimization opportunities during training and can subsequently modify the ML model to take advantage of the human’s intuition. In short, CAML allows for the direct transference of human knowledge into ML models.
Adding CAML to an existing ML pipeline can improve model accuracy, compress model size, or allow model improvement in absence of large data sets. This talk will show examples of CAML being used to guide ML model when analyzing medical and satellite imagery as well as knowledge transfer directly into the Leela Zero deep learning model (Open Source version AlphaGo Zero). The process is compatible with HEWMENs distributed ML techniques, such as federate learning, which allows for scaling of CAML on both human and machine components.
Key Takeaways Points:
1. See examples of how human-in-the-loop techniques can dramatically accelerate ML training as well as techniques to extract knowledge from trained ML models.
2. Demonstration of BALANCED’s HEWMEN platform, which combines video games, human intuition and distributed computing into a single cloud environment.