Autonomous and continuously self-improving learning system

#This invention relates generally to the field of digital image processing, machine learning and computer predictive models, and more particularly to a system and methods for autonomously learning hierarchical representations from analysis of large-scale unlabeled raw data and continually self-improving detection capability using self-labeled data points.

#Machine learning is an application of artificial intelligence (AI) wherein systems have the ability to automatically learn and improve from experience without being explicitly programmed. Machine learning systems are used for a variety of applications including image analysis, speech recognition and detection of credit card fraud. As implied by the word “learning,” the machine learns by training with real world examples to form deductions about newly presented information. For example, a machine learning system could be trained with several thousand images of content classified and labelled by a human observer, so that the content of a future previously unseen image can be deduced. This approach is referred to as “supervised learning” in that the machine learning is supervised by a human.

#Previously known systems generally consist of two fundamentally different types of tasks in machine learning—unsupervised learning and supervised learning. The goal of supervised learning is to create a mapping from inputs to outputs (x to y), given a training set made up of input-output pairs (xi, yi). The core learning mechanism and fundamental building block of any supervised learning algorithm is “backpropagation,” a method first popularized by Rumelhart et al, “Learning representations by back-propagating errors,” Nature, 323(9):533-536 (1986). Backpropagation is the process of iteratively adjusting the parameters of a model (e.g., weights and biases) to minimize the prediction error. In the context of computer vision and image classification, one of the more influential successes of supervised learning was the publication of AlexNet, as described by Krizhevsky et al. in “ImageNet Classification with Deep Convolutional Neural Networks,” Advances in Neural Information Processing Systems, 25, pp. 1097-1105 (2012). Since then, supervised learning has become a de facto standard for state of the art learning tasks.

#FIG. 1 is a schematic diagram of a system implementing the autonomous and continuously self-improving system of the present invention.

#FIG. 2 is a schematic illustration of data flows within inventive system.

#FIG. 3 illustrates the flow of data through the unsupervised feature extraction module and the supervised feature extraction module of the system of FIG. 1.

#In view of the foregoing drawbacks of the previously known systems, the present invention is directed to a configurable system for self-analyzing, self-cleaning, and self-extracting features from unlabeled data to provide a continually self-improving learning system for predictive modeling of predefined targets. In an exemplary embodiment of the invention, an artificial intelligence (“AI”) inference module for extracting desired features from an input stream resides on a local computer. Apparatus, such as an endoscopy or colonoscopy system associated with the local computer provides a data stream, e.g., video output, that serves as input feed to the AI inference module running on the local computer. The AI inference module analyzes the input feed and generates annotations for the user of the local computer, e.g., which may take the form of annotations in a display of the video output presented to the apparatus user, to signal detection in the video output of potential abnormalities, such as polyps.

#In accordance with the foregoing exemplary embodiment of the present invention, the video output from the procedure is uploaded to a remote server that is programmed with the continually self-improving learning system of the present invention, such that the video output serves as an additional dataset for training of the continually self-improving system. Analysis of the new video output data may generate additional learned features that improve the unsupervised learning capability of the AI inference module. The AI inference module then is updated at the local computer for use in subsequent procedures or re-evaluation of the recorded video output to improve accuracy of the results generated by the preceding instantiation of the AI inference module at the local computer.

#In accordance with one aspect of the invention, a method is provided for self-analyzing and self-extracting features from large-scale unlabeled data for use in self-learning, without human supervision. The method involves a series of unsupervised feature extraction processes, including self-supervised pretext task learning and cleaning/grading, which extract a set of self-learned micro-level and macro-level features/parameters. Those micro- and macro-level features provide intermediate and global context features representative of the unlabeled dataset. The learned features from the unsupervised processes are applied to self-cleaned, self-graded unlabeled data before being sent as inputs to a semi-supervised self-learning process.