Why humans are indispensable for many automation projects
Artificial Intelligence (AI) is changing the world of work. But automation is not making human intelligence obsolete. For the vast majority of automation projects, Artificial Intelligence relies on human intelligence to play to its strengths. In this article we explain what you need to know about the human-in-the-loop models.
What does Human-in-the-Loop mean?
Artificial Intelligence aims to automate intelligent actions. Tasks that are normally performed by humans are completed by an AI. This opens up great potential for the automation of data-intensive tasks in all industries and all areas of a company. In many cases, however, removing humans from the process entirely does not prove efficient. Instead, human-in-the-loop combines human and AI in a joint process.
Humans learn through experience, an AI learns through data. Supervised learning uses qualified data sets that have been specially prepared to train the algorithm. By “labeling” inputs with their respective outputs, the AI can derive patterns and apply them to the new data sets.
Now the model is fed with the data. Supervised learning is used in two areas: In Classification, an algorithm sorts data (e.g., in texts or images) into given classes based on patterns and similarities. In a Regression, an algorithm estimates the quantitative magnitude of a target variable based on the relationships of different variables in the data. Regression is used where numbers need to be predicted – for example, in sales forecasting.
Improvement and Finetuning
In the next step, humans adapt the Machine Learning model to avoid typical problems, such as overfitting, and to integrate important special cases into the model. To do this, cases with a lower level of certainty, in particular, are checked manually and the result of the check is fed back to the model.
Human-in-the-Loop: Advantages & Use Cases
Compared to full automation, Human-in-the-Loop models have numerous advantages for various use cases:
- Higher accuracy: Artificial Intelligence can never achieve 100% accuracy. The technology is based on statistical methods that only know probabilities. Humans can often make statements with higher certainty. In many cases, this combination can systematically produce better results than a machine-only approach.
- Harnessing expertise: By integrating specific experiential knowledge, blind spots in the algorithm can be illuminated. For example, NASDAQ, with its tool for monitoring trading activity on the stock exchange, relies on a Human-in-the-Loop approach in which experts pass on their knowledge to the algorithm via feedback loops.
- Borderline cases: Big data is not publicly available for every problem. For example, if a chatbot needs to understand user input in a rare language or local dialect, the algorithm relies on human feedback. Facebook uses a Human-in-the-Loop model to analyze communication on the social media platform because not every aspect of human communication can be extrapolated from the data.
- Detecting bias: The problem of hidden biases is most familiar from behavioral research. Even data sets are not neutral in every respect, but may contain unrecognized but systematic dependencies on influencing factors that have not been taken into account. Just as the way a question is asked can influence the answers, the way data is obtained can influence how AI interprets the data. If they come from public sources, the path of data extraction may not be fully traceable. Without human intervention, an algorithm would pick up the bias from the data. Human-in-the-Loop can often identify any bias before it becomes entrenched in the algorithm.
Human-in-the-Loop in practice: Classification of customer data
How Human-in-the-Loop can lead to better automated processes is demonstrated by the automation of verification processes in the context of legally mandated know-your-customer policies, (partial) automation is indispensable. For example, an intelligent algorithm can compare and validate uploaded selfies with the photo of the ID document. Suppose an algorithm is to compare a selfie photo with the photo of an ID card. This application requires a high level of accuracy that an AI alone cannot achieve:
- Employees create the data sets used to train the algorithm in the first step. In the example, employees would create sets from photos of selfie and ID document and assign a corresponding value (correct / incorrect) to each. The algorithm is fed with this training data and thus learns to recognize features based on which it can correctly classify the image pairs. The algorithm outputs one of these values for each set. Now a human checks the output of the algorithm and feeds it again with this result. This creates a feedback loop – and the algorithm gets better and better at classifying the data.
- There is always data for which the algorithm cannot deliver a clear result. For example, for identity documents from individual countries, there is not enough initial data for the AI to evaluate effectively. These cases, when they fall below a confidence threshold, are automatically forwarded to staff for manual review, feeding the algorithm with more valuable information.
In this model, the human is systematically integrated into the cycle of the AI, which thus systematically produces better results and learns faster than a model without a human in the feedback loop.
Conclusion: No AI without humans
Artificial Intelligence will play an increasingly important role in the company in the future. However, this will not make humans superfluous – on the contrary: The preparation and processing of data will become essential for the successful development of intelligent algorithms. Only through the systematic integration of human feedback can numerous applications be realized with Artificial Intelligence: from predictive analytics to the automation of customer-related processes.