The field of artificial intelligence (AI) has made significant advancements in recent years. However, one aspect that many AI systems fail to fully comprehend is human error and uncertainty. In many systems where humans provide feedback to a machine learning model, these systems assume that humans are always certain and correct in their decisions. Real-world decision-making, however, often includes occasional mistakes and uncertainty. Recognizing this gap between human behavior and machine learning, researchers from the University of Cambridge, in collaboration with The Alan Turing Institute, Princeton, and Google DeepMind, have been working on a solution. Their aim is to incorporate uncertainty into AI applications where humans and machines work together, with the potential to reduce risk and improve trust and reliability, particularly in critical areas such as medical diagnosis.
Humans constantly make decisions based on the balance of probabilities, often without conscious consideration. It is common for humans to make mistakes, such as waving at a stranger who resembles a friend. While these mistakes are typically harmless, certain applications, particularly safety-critical ones, carry real risks related to uncertainty. Unfortunately, many human-AI systems assume that humans are always certain in their decisions. This is not an accurate reflection of human nature, as humans are prone to errors. Therefore, addressing uncertainty from the human perspective is crucial in safety-critical settings. Katherine Collins, the first author of the study from Cambridge’s Department of Engineering, emphasizes the need to develop tools that empower individuals working with AI models to express their uncertainty. It is important to recalibrate these models to account for uncertainty from human input, as machine learning models struggle with uncertainty when humans can’t provide complete confidence.
The researchers conducted their study using benchmark machine learning datasets. They used one dataset for digit classification, another for classifying chest X-rays, and a third for classifying images of birds. While the researchers simulated uncertainty for the first two datasets, they incorporated human input in the bird dataset. Human participants were asked to indicate their level of certainty in identifying whether a bird is red or orange, for example. These “soft labels” provided by the human participants allowed the researchers to examine how uncertainty affected the final output. However, they discovered that when machines were replaced with humans, performance significantly degraded. This finding highlights the challenge of incorporating human uncertainty into machine learning models.
Challenges and Future Research
The researchers acknowledge that their study has identified several open challenges in incorporating humans into machine learning models. They plan to release their datasets to facilitate further research and exploration of incorporating uncertainty into machine learning systems. Addressing uncertainty is crucial for developing trustworthy and reliable human-in-the-loop systems. By accounting for human behavior and understanding the mis-calibration of uncertainty in humans, the trustworthiness and reliability of these systems can be improved.
The incorporation of uncertainty into machine learning models is not only about accuracy but also transparency and trust. Uncertainty serves as a form of transparency, allowing humans to understand the reliability of a model and when to trust it. In certain applications, where probabilities are considered, such as the rise of chatbots, it is essential to develop models that better incorporate the language of possibility. This integration of uncertainty and possibility can lead to a more natural and safe experience for users.
The research conducted by the University of Cambridge, The Alan Turing Institute, Princeton, and Google DeepMind highlights the importance of incorporating uncertainty into machine learning systems. Recognizing that human error and uncertainty play significant roles in decision-making processes, the researchers aimed to bridge the gap between human behavior and machine learning. While their study revealed challenges and degraded performance when humans were involved, it also outlined the potential for improving the trustworthiness and reliability of human-in-the-loop systems by accounting for human uncertainty. This research opens up avenues for further investigation and development of tools and techniques that empower individuals to express their uncertainty and enable machine learning models to adapt to human behavior. Ultimately, incorporating uncertainty into AI applications has the potential to revolutionize decision-making processes and improve safety in critical areas such as medical diagnosis.