One-shot learning and information amplification via Bayesian uncertainty approximation for human-AI interaction

Abstract

In this study, we consider an alternative methodology for estimating the Bayesian uncertainty of humans and artificial intelligence (AI) systems using meta-models and a new paradigm of human-AI interaction that achieves an intrinsic quality advancement of convergent knowledge by harnessing the mutually independent learning abilities of humans and AI. Since deep learning algorithms advance through highly nonlinear approximation characteristics that are difficult to realize in biological intelligence, AI (Machine learning model) has begun to outperform human intelligence in several areas. Nevertheless, human intelligence involves unique characteristics that have yet to be fully exploited by machine learning (ML) algorithms. One-shot learning is a unique human learning ability, in which people rapidly develop a general knowledge model using one or a few data samples. According to recently developed theories in computational neuroscience, uncertainty is a key variable guiding human one-shot learning and inference. If human uncertainty can be directly estimated via ML, computational control of human one-shot learning will become possible. Therefore, in this study, based on the nonlinear approximation characteristics of ensemble ML, we developed an ML-guided human learning system designed to estimate the uncertainty of a human learner and generate a data sequence that can be used to modulate human uncertainty to induce one-shot learning using the estimated information (AI-guided human learning). Furthermore, we assumed that the human one-shot learning induced by the ML model could approximate the knowledge distribution that could not be achieved by the ML model in other situations. Therefore, feeding back one-shot-learned human knowledge into the ML model could ultimately create a virtuous learning cycle in which the performance of the model improves further under limited learning conditions (human-guided AI learning). This phenomenon is defined as information amplification in this study. When the human one-shot learning through ML and the ML model that learns recursively through the feedback of one-shot learned human information form a virtuous cycle, the efficiency of the cycle can be further enhanced using the uncertainty information of the ML model. The investigation of these topics comprised two steps. In Step 1, we proposed a new meta-modeling framework to estimate the uncertainty of human and ML models, aiming to validate the key hypothesis that human one-shot learning can induce information amplification. In Step 2, we validated the one-shot learning and information amplification framework via an uncertainty estimator-based human learning control system implemented in a human-ML co-learning environment. In Step 1, we first modeled human decision making as an inference process of a Bayesian neural network and showed that human uncertainty can be expressed as Bayesian uncertainty. To do so, we proposed an ensemble neural network algorithm designed to infer Bayesian uncertainty in a wide data area through a single behavioral sampling on a small set of sample data. This ensemble model for estimating human uncertainty was validated via an interpretation experiment involving expert subjects (physicians) and using specialized data (medical images). Second, we developed a meta-model algorithm to approximate a deterministic neural network converging under the same output distribution for the training and validation datasets as an interpretable Gaussian process, called the proxy Gaussian process. From the equality of output predictions when the same kernel function defines the Gaussian process and kernel ridge regression models, we showed that the Bayesian uncertainty of deterministic neural networks can be approximated through the proxy Gaussian process. In Step 2, we considered an environment in which human and ML models learn cooperatively based on the Bayesian uncertainty estimation of human decision making and deterministic neural networks obtained in Step 1. A reinforcement learning-based control system designed to estimate human uncertainty and induce human one-shot learning was embedded in the co-learning environment in which the inflow of external information was blocked. Regarding the information flow between human and ML models as a Markov chain, mutual information with the ground truth variable was lost as information exchange progressed, which may have reduced performance. However, our approach exhibited a recursive performance improvement that can be interpreted as an increase in the amount of mutual information in the one-shot learning environment. That is, a human learner who performs one-shot learning in a closed co-learning loop can be considered an information amplifier. This hypothesis was verified through a prospective co-learning experiment in which an expert (physician) subject learned a small number of medical images and interacted with an ML model real time. Overall, this work presents an effective framework for estimating Bayesian uncertainty in human decision making and deep neural networks, and we verified the performance of a practical ML model developed to induce human one-shot learning through uncertainty estimation. Furthermore, we proposed the concept of information amplification and demonstrated it through large-scale behavioral experiments involving expert subjects. Our results pave way for a paradigm based on a convergent co-evolution of human and machine intelligence that aims to acquire integrated knowledge not otherwise available to either human or ML models alone.