Human gradient artificial intelligence for self-aware learning systems

Abstract

Artificial intelligence (AI) generally defines intelligence as an optimal association between input features and predictive behavior. In essence, it is defined as knowledge updating under adaptation, describing the static relationship between input features and decision accuracy in a human-interpretable form. This definition is not suitable for the cutting-edge colony of AI frameworks where intelligence evolves from fast, detailed, and organized learner AI agents. This paper introduces an innovative Human Gradient Artificial Intelligence (HGAI) approach to model intelligence as a measurable continuous gradient of knowledge updating under evolution with human-inspired behaviors such as metacognition-driven self-assessment, self-regulation, and self-reflection. In HGAI, pretrained machine learning models define AI agents to understand their evolutionary behavior by modeling gradient of knowledge updating with softmax probability, cross-entropy, stochastic gradient descent and regularization, while considering predictive performance over two phases of discrete time-steps. AI agents are also pretrained in the principal component basis of the original feature space. Additionally, confidence evolution over gradient is defined as an aggregated effect of high probabilities of predictions and the prediction losses to reflect metacognition-driven self-assessment. Trustability of knowledge update is also defined with the rolling statistics of losses and prior updated knowledge to integrate self-regulation and self-reflection. Experimental results of kernel-SVM, ridge-regression, Lasso, and Perceptron models, integrating fast, detailed, and organized learner AI agents with Digits and Fashion-MNIST datasets, show prediction accuracy in the second phases are generally higher (92-99%) for the fast and detailed learners. However, it is lower for the organized learners, satisfying higher accuracy in the second phase than in the first. The acceptance rate of gradient knowledge update is also high in all cases (85-100%). Additionally, confidence evolves steadily for all three types of learners of all the models, as models update their gradient knowledge. Hence, HGAI reflects gradient-based, self-aware evolutionary intelligence.