Applying Machine Learning to Psi Research | Mark Boccuzzi

With the advent of Big Data (data sets so large or complex that they render traditional information technology tools and techniques ineffective), data analysts have turned to Artificial Intelligence and machine learning in order to find meaning in and make predictions from data. Machine learning algorithms iteratively learn from data, allowing computers to find connections in data without the need for them to be explicitly programmed. The past few years have brought major advancements to the field of machine learning, including new developments in deep learning (the implementation of complex, multilayered neural networks) and advanced applications such as natural language processing, computer vision, medical diagnosis, user preferences, and image recognition and feature extraction. In addition, many of the companies and universities that are driving this development effort have started to release their machine learning tools as either low cost or open source software. Google, Inc.’s TensorFlow, IBM’s Watson, and Microsoft’s Computational Network Toolkit along with a wide range of powerful APIs (application program interfaces) now make machine learning highly accessible. With these new tools come new potential methods for examining psi-related data sets. In a recent exploratory study, machine vision recognition software was used to classify and extract details from a set of images that were created from random event generator (REG) data that were collected from 10 sessions in which meditators focused on feelings of “Love” (five sessions) and “Hate” (five sessions). The data, collected using FieldREG software (Psyleron, Inc.), were then processed though Windbridge Institute custom visualization software that converted them into complex 3D images.

Traditionally, FieldREG data are analyzed by looking for deviations from randomness; however, the current exploratory study employed image concept recognition software (Clarifai, Inc.) to classify the images based on their visual attributes. While there was considerable overlap in the resulting classifications, the software successfully grouped four of the five “Hate” images together with a unique descriptor not found in any of the other images, nor previously considered by the investigator. What is notable here is that the software made this test possible with just a few hours of investigator time in contrast to previous image classification approaches which would have required a specifically defined classification scale and a large number of research participants. The results of this exploratory study demonstrate the potential value in the application of machine learning to visual data sets both in terms of time/cost efficiency but also in potential hypothesis generation. As machine learning systems can quickly find connections in data that humans cannot, applying them to psi datasets may produce new, testable hypotheses which could profoundly move the field forward.

Recorded at the Society for Scientific Exploration Conference in Boulder, Colorado 2016.

Join the SSE to support to support the Society’s commitment to maintain an open professional forum for researchers at the edge of conventional science:

The SSE provides a forum for original research into cutting edge and unconventional areas. Views and opinions belong only to the speakers, and are not necessarily endorsed by the SSE.

Published on November 22, 2018