Autonomous and continuously self-improving learning system
A system and methods are provided in which an artificial intelligence inference module identifies targeted information in large-scale unlabeled data, wherein the artificial intelligence inference module autonomously learns hierarchical representations from large-scale unlabeled data and continually self-improves from self-labeled data points using a teacher model trained to detect known targets from combined inputs of a small hand labeled curated dataset prepared by a domain expert together with self-generated intermediate and global context features derived from the unlabeled dataset by unsupervised and self-supervised processes. The trained teacher model processes further unlabeled data to self-generate new weakly-supervised training samples that are self-refined and self-corrected, without human supervision, and then used as inputs to a noisy student model trained in a semi-supervised learning process on a combination of the teacher model training set and new weakly-supervised training samples. With each iteration, the noisy student model continually self-optimizes its learned parameters against a set of configurable validation criteria such that the learned parameters of the noisy student surpass and replace the learned parameter of the prior iteration teacher model, with these optimized learned parameters periodically used to update the artificial intelligence inference module.
1. Apparatus for self-improving the classification of features detected in an image dataset, the apparatus comprising:
memory for storing instructions; and
at least one processor to execute the instructions to:
partition the dataset into a first subset and a second subset;
receive from a domain expert user a curated dataset based on review of the first subset;
extract learned intermediate features from the second subset using self-supervised pretext tasks;
train a teacher model using the learned intermediate features and the curated dataset;
execute the teacher model to generate pseudo labels for the second subset;
apply an active learning process to select a subset of the pseudo labels;
apply a multiple instance learning process to the subset of pseudo labels to generate a reduced subset of pseudo labels;
employ a weakly supervised training process to train a noisy student model using the reduced subset of pseudo labels to generate a set of learned parameters for classification of detected features; and
validate the set of learned parameters.
2. The apparatus of claim 1 , wherein the instructions further comprise instructions for cleaning and grading data within the second subset after partitioning the dataset.
3. The apparatus of claim 1 , wherein the instructions further comprise, during validation of the set of learned parameters, comparing an accuracy of classification of detected features against at least a portion of the curated dataset.
4. The apparatus of claim 1 , wherein the instructions further comprise, during validation of the set of learned parameters, identifying a portion of the reduced subset of pseudo labels for error correction by a domain expert user.
5. A system for classifying features detected in an image dataset, the system comprising:
apparatus for generating a stream of video data;
a local computer having an AI inference module for analyzing the video data to detect and identify potential abnormalities;
a remote server coupled to the local computer via a wide area network to receive the video data, the remote server comprising
memory for storing instructions and the video data; and
at least one processor to execute the instructions to:
partition the video data into a first subset and a second subset;
receive from a domain expert user a curated dataset based on review of the first subset;
extract learned intermediate features from the second subset using self-supervised pretext tasks;
train a teacher model using the learned intermediate features and the curated dataset;
execute the teacher model to generate pseudo labels for the second subset;
apply an active learning process to select a subset of the pseudo labels;
apply a multiple instance learning process to the subset of pseudo labels to generate a reduced subset of pseudo labels;
employ a weakly supervised training process to train a noisy student model using the reduced subset of pseudo labels to generate a set of learned parameters for classification of detected features;
validate the set of learned parameters; and
periodically transmit the set of learned parameters to the local computer to update the AI inference module.
6. The system of claim 5 , wherein the instructions further comprise instructions for cleaning and grading data within the second subset after partitioning the video data.
7. The system of claim 5 , wherein the instructions further comprise, during validation of the set of learned parameters, comparing an accuracy of classification of detected features against at least a portion of the curated dataset.
8. The system of claim 5 , wherein the instructions further comprise, during validation of the set of learned parameters, identifying a portion of the reduced subset of pseudo labels for error correction by a domain expert user.
9. The system of claim 5 , wherein the apparatus for generating a stream of video data comprises an endoscope.
10. A system for classifying features in a video stream output by a colonoscope, the system comprising:
a local computer having an AI inference module coupled to colonoscope for analyzing a video stream to detect and identify potential abnormalities;
a remote server coupled to the local computer via a wide area network to receive and store the video stream, the remote server comprising
memory for storing instructions and the video stream; and
at least one processor to execute the instructions to:
partition the video stream into a first subset and a second subset;
receive from a domain expert user a curated dataset based on review of the first subset;
extract learned intermediate features from the second subset using self-supervised pretext tasks;
train a teacher model using the learned intermediate features and the curated dataset;
execute the teacher model to generate pseudo labels for the second subset;
apply an active learning process to select a subset of the pseudo labels;
apply a multiple instance learning process to the subset of pseudo labels to generate a reduced subset of pseudo labels;
employ a weakly supervised training process to train a noisy student model using the reduced subset of pseudo labels to generate a set of learned parameters for classification of detected features;
validate the set of learned parameters; and
periodically transmit the set of learned parameters to the local computer to update the AI inference module.
11. The system of claim 10 , wherein the instructions further comprise instructions for cleaning and grading data within the second subset after partitioning the video stream.
12. The system of claim 10 , wherein the instructions further comprise, during validation of the set of learned parameters, comparing an accuracy of classification of detected features against at least a portion of the curated dataset.
13. The system of claim 10 , wherein the instructions further comprise, during validation of the set of learned parameters, identifying a portion of the reduced subset of pseudo labels for error correction by a domain expert user.
14. The system of claim 10 , further comprising a colonoscope for generating the video stream coupled to the local computer.