Human in-the-loop in Machine Learning
1 like545 views
Human-in-the-loop (HITL) machine learning integrates human and machine intelligence to enhance algorithm training and testing, allowing for continuous improvement through human feedback. It involves various data labeling techniques to develop AI models for applications like natural language processing and computer vision. The HITL approach is evolving towards automation while ensuring human oversight, significantly impacting business workflows in the future.
Human in-the-loop in Machine Learning
- 2. Uncertain Confident Active Learning Output AI Classifier Human Annotation Human-in-the-loop Machine Learning 2
- 3. What is Human-in-the-loop? ▪ Human in the loop machine learning is a process that leverages both human and machine intelligence to create machine learning-based AImodels. ▪ When a system is unable to provide a solution or solve a problem, human intervention is needed in the training and testing phases of an algorithm development. ▪ This creates a recurring feedback loop where every iteration of algorithm provides better solutions.Humans can also modify the algorithm to improve accuracy. 3
- 4. What is HITL Machine Learning? Combination of human & machine intelligence. Incorporate human feedback into learning loop of machines. Humans are involved in training, tuning and testing of algorithms. 80% of the time, the algorithm is left alone with the human involvement limited to 19% and remaining 1% left to randomness. 4
- 5. Human-in-the-loop AI Increase efficiency of every human expert that work with the AIsystem Increase the level of quality by collecting and sharing data and data-driven insights Capture silent knowledge and expertise of human experts in a single system Human expert confirms, rejects or labels output AI model generates output Model learns from human and accuracy constantly improves 5
- 6. Combining Machine Learning with Human Contributions Crowdsourcing of subjective parameters. ▪ Contextualize data. ▪ Use user expertise in identifying and recognizing patterns. ▪ Process data that is difficult to process automatically. ▪ Ask users to improvealgorithms. ▪ Human input to improve machine inference. ▪ Use machine inference to monitor and improve user performance. 6
- 7. Types of Data Labelling The data labelling can be categorized into different types such as ▪ Semantic segmentation to train visual perception annotation-based model ▪ Landmark annotation to train facial recognition ▪ NLPannotation, text annotation and sentiment analysis to train language or voice recognition model. ▪ The data are annotated and labelled to develop AI devices that communicates with humans. ▪ HITLcan create different types of training data sets for different types of machine learning models built for different fields. 7
- 8. Object & activity detection Person tracking Face recognition Real - time live stream Content moderation Celebrity recognition Video Analysis 8
- 9. Image Annotation for Computer Vision 9
- 10. Typical Workflow for a Machine Learning Model Deploy the model Generate example data Train a model Clean Fetch Prepare Train modelEvaluate model Deploy to production Monitor/ Collect data/ Evaluate 10
- 11. Lifecycle of HITL Machine Learning ▪ Inception - A manual workflowgets created. ▪ Iteration - The workflow is improvediteratively. ▪ Transition - The workflow transitions from a complete manual to semi-automated. ▪ Monitoring - Complete automation shifts HITLinto a validation mechanism. 11
- 12. Inception ▪ Here, we follow a "Divide and Conquer" approach where we categorize the process into smaller components and train the machine learning models accordingly. ▪ The goal of the Machine learning model is defined and conceptualized. We fetch the raw unstructured data, clean, label the same and prepare for the next stage. 12
- 13. Iteration ▪ We increase the efficiency, time cost, quality and accuracy of the machine learning model to achieve better outcomes and obtain higher quality data. ▪ We train and evaluate the model to determine its efficiency in an iterative way. ▪ The iterative process on the workflow components benefits us to structure our machine learning modelefficiently. 13
- 14. Transition ▪ We deploy the machine learning model to production once desired human efficiency and quality levels are achieved. ▪ Combining these automated mechanisms with the human component creates the actual Human-in-the-Loop machine learning system. ▪ The transition adopts into the following. ▪ Fully manual - The human receives the automated system’s decision along with the input data, which can be used to evaluate the model’s performance before it can independently act. 14
- 15. Types of Transition ▪ Augmented manual - The human is augmented by the automation, acts as a gate-keeper for automated decisions, approving or declining — and correcting — those automated decisions. ▪ Semi-automated - A fraction of the automation’s decisions stop being monitored by a human. The rest remains under the Augmented manual model. ▪ Automated - The model reaches a desired accuracy level and humans stop being involved in the loop. 15
- 16. Monitoring ▪ HITL is never fully trust what is automated. Even at the Automated phase, there occurs the touch of humans to verify that the automated processes acts as intended. ▪ Once the monitoring process is operationalized, we must ensure that it is precise and accurate to avoid the potential risks ahead. 16
- 17. Who uses Human-in-the-loop Machine Learning? ▪ HITL can be used for manifold AI projects including NLP, computer vision, sentiment analysis, transcription, and a vast amount of other use cases. ▪ Any deep learning AI can benefit from human intelligence involved into the loop at some point. 17
- 18. Logistics ▪ Determine and predict demand. ▪ Patterns of the route. ▪ Characterize congestion areas. Retail management ▪ Why do users buy what they buy? ▪ Onset of fashions ▪ Products development ▪ Targeted advertisement and influence Application Areas 18
- 19. 19
- 20. Out ofstock Out ofstock Out ofstock No visible label present cannot box as an out - of - stock item. 20
- 21. Business Address Order Number 21
- 22. Conclusion ▪ HITL will significantly change the way business workflows are carried out in future by creating a pipeline that includes data collection, model training, testing, deployment andmaintenance. ▪ This is a very excitingtime to be involved in this field as industry and institutions are pushing the limits furtherevery day. 22
- 23. [email protected] www.mitosistech.com IND: +91-78240 35173 US: +1-(415) 251-2064