Many predictions have foretold that there will be a reduction in the interaction between humans and machines in proportion to the advances made by artificial intelligence (AI) and machine learning (ML) systems. What is, however, being witnessed today is that human input is still a necessity for most common forms of AI/ML training. Does this mean that as AI/ML technology continues to progress, human-in-the-loop (HTL) scenarios will disappear?
As AI/ML evolves and baseline accuracy of the models improves, human interaction will change from the creation of generalized ground truth from the scratch to human review of the worst-performing ML predictions. This evolution will serve to improve and fine-tune models iteratively and cost-effectively. The iterative nature of the process will still rely on human input, the work will require more subject matter expertise and consensus on what could be the most appropriate answer.
Deep learning algorithms thrive on labeled data and it can be improved progressively as more training data is added. A human review of every element of ML-generated output would be a painstaking task. Hence, the best approach would be to analyze ML predictions programmatically, prioritize the areas self-reported as low confidence for human review and editing, and subsequently reintroduce the results as new training data.
The industry too has shifted from simple bounding boxes and speech transcription to pixel-perfect image segmentation and millisecond-level time slices in audio analysis. As a result, labeling accuracy will increasingly become a primary concern. Human input in situations require judgment when facing a series of potentially disastrous results such as car moving into the same lane as an autonomous vehicle, is no longer theoretical, it has become the next frontier in data annotation.
Therefore, establishing ground truth has become more subjective, requiring increasingly higher levels of subject matter expertise and labeling precision. In the future, smaller teams of annotation specialists with deep subject matter expertise may well displace the commodity data labeling services currently offered by crowdsourcing and global business process optimization organizations. This shift will require expensive labor, strict quality controls, specialized toolsets, and workflow automation to optimize the process versus huge teams of low-cost labor.
It is evident that AI/ML is still a rapidly evolving field and human-machine interaction to support model training will continue to be a critical input for the near future. However, what will change is the nature of HTL workflows and the human expertise involved as the annotation problems to be solved become increasingly more complex and demanding.
Click here to read the original article published in mc.ai
Please give your feedback on this article or share a similar story for publishing by clicking here.
