US 12,737,636 B2Granted

Autonomous AI process through reinforcement learning with expert feedback

AssigneeTecholution Consulting LLC

InventorLuv Tulsidas

Life of this patent

Estimating the term…

Loading patent…

Abstract

The present application relates to a process and system for automating Reinforcement Learning with Expert Feedback (RLEF) in AI training processes. The process provides a user-friendly and automated platform (system) that facilitates the training of AI models and enables continuous learning through Reinforcement Learning with Expert Feedback. The process reduces the time and iterations required to build an autonomous grade AI model that performs with high accuracy and consistently when deployed and operating in real world situations. The autonomous AI system integrates a live feedback loop mechanism that combines an AI model, this is referred to as the “apprentice”, with instructors who are domain experts at the particular job or tasks that the AI is attempting to automate. This setup allows for the swift training or retraining of the apprentice-level AI system in the field of interest, leading to a rapid continuous improvement of the model towards achieving expert level accuracy where the AI is able to perform the job at the level of the domain expert autonomously. The instructors provide valuable feedback and guidance to the AI model, enabling it to learn and adapt effectively. By leveraging this feedback loop, the AI system can enhance its performance and make more accurate predictions or decisions.

Claims14

Dependent claims

A process for automating AI model training processes in an AI system, comprising the steps of:

a) capturing and labeling data, pairing an apprentice level AI model with instructors, and training/retraining the apprentice level AI model by correcting errors in the apprentice AI model in real time;
b) validating real-time information and dataset approval through a verification module to ensure utilization of verified and approved data exclusively during the training process;
c) assembling the verified and approved data in an AI collection module to ensure the apprentice level AI model is trained exclusively on reliable data;
d) configuring an API to trigger training and retraining script of the apprentice level AI model using assembled expert-level data;
e) automating training and retraining script of the apprentice level AI model by domain expert(s) to further rapid on-demand training, retraining, and validation of the apprentice level AI model as it evolves autonomously;
f) selecting the expert-level data collected in the AI collection module, utilizing train/retrain controls for further training, retraining, and data analysis;
g) conducting automated backtesting on the apprentice level AI model using a dataset not utilized for training, in order to evaluate the validity and performance of the newly trained apprentice level AI model;
h) utilizing a co-pilot live module for real-time correction of AI's decisions, therewith enabling domain experts to supervise and rectify incorrect decisions made by the apprentice level AI model;
i) utilizing a co-pilot historical module to identify and rectify AI model decisions/inferences and allowing domain experts to visually review, investigate and analyze the decisions and incorrect inferences made by the AI system in the past data set on specific parameters and the subsequent corrections applied;
j) accumulating real-time data from the deployed model, eliminating irrelevant data, and generating an AI model for further training;
and
k) generating a performance and accuracy report of the AI model based on the accumulated real-time data and feedback provided by instructors acting as experts.

9 more elements

A system for automating reinforcement learning with expert feedback in AI training processes;

comprising:
a) a live feedback loop mechanism that pairs AI model with instructors to rapidly train/retrain an apprentice level AI system;
b) a user interface facilitating interaction between human expert trainers and the AI model;
c) a real-time data accumulation and processing module that generates an AI model by filtering out irrelevant data and combining unseen and confused data;
d) an AI-assisted approval process module that accurately labels captured data and classifies it;
e) a verification module that validates the approval of real-time information and datasets;
f) an AI collection module that assembles approved data;
g) an API module that triggers the training and retraining of the model using the assembled data;
h) a backtesting module that automatically assesses the validity and performance of the newly trained AI model by backtesting it on a dataset not used for training, using training and retraining controls, wherein the key metrics of this version of the model is compared to prior versions of the model trained to assess if the accuracy of the model is improving/qualified for production deployment;
i) a click button or algorithmically API call for deploying the trained AI model to production once training has been completed;
j) a plurality of models deployed automatically or on demand to production where a first model runs as the primary decision maker while a second is silently making decisions in production as a background model for comparing its performance with previously deployed and proven model;
k) a plurality of background models deployed to production other than system and hardware limitation where the system is deployed;
l) A co-pilot module that identifies and rectifies errors in real-time;
and
m) a feedback module that feeds the corrected data back into the approval process, enhancing the model's performance through an ongoing cyclical process of model training.

12 more elements

Dependencies are inferred from the claim text.