Akshata Moharir
Machine learning and data science are transforming decision-making in safety-critical domains such as aerospace, automotive, health care, and industry manufacturing. Though these technological advances have many advantages, they do present issues with transparency and reliability. This is where human-centered machine learning plays a crucial role. HCML ensures ML models are not just powerful, but also interpretable, trustworthy, and aligned with human oversight—a necessity in high-stakes situations in which errors can have dire consequences.
Why Explainability Is Necessary in the Context of ML for Safety-Critical Systems
One of the significant issues with the safety-critical applications of ML is the lack of explainability. In other words, an ML model is expected to be both accurate and able to explain the reasons to be accurate. As an example, an ML system in the health care industry is supposed to provide a diagnosis and explain the reasons for choosing the diagnosis to the doctors. In a similar way, ML should be able to provide the reasons for its actions in, say, an autonomous vehicle.
Explainability/interpretability also boosts the confidence of domain experts, regulators, operators, and consumers who interact with the ML systems. When interpretability is taken into consideration while training the ML models, professionals can make better decisions and stay away from automation bias—the tendency of relying too much on ML recommendations without challenging them.
Automation Versus Human Judgment: How to Combine the Two Approaches?
Another essential principle of HCML is to recognize when to employ automation and when to use human input. Although ML can analyze a large amount of data and make predictions quickly, the human element is needed, especially in safety-critical systems. For example, ML systems in the aviation industry are used to monitor various sensors that can give early signs of equipment failure, but human engineers/domain experts confirm these findings before any action is taken. The combination of both ML prediction and human judgment not only improves safety but also assists in reducing uncertainties in the form of false positives or negatives.
How to Combat Bias and Make ML Fairer?
It is important to note that ML models are not bias free, and this can be uncertain in high-risk domains such as fraud detection, vehicle safety, or finance. If the training data is biased or partially missing, the ML will make decisions that are worse for some people than others. For this reason, HCML requires the use of large and diverse data sets, the watching of models in real-time, and the inclusion of fairness into the algorithms. We can prevent bias from evolving in the real world when human oversight is applied during the training and validation of the ML models.
Future of HCML in Safety Critical Systems: What’s Coming Up?
In the process of ML development, HCML will continue to increase reliability, transparency, and trustworthiness in safety-critical applications. By integrating human-centered principles into ML development, we can create systems that are safe, ethical, and fair in addition to being effective. She has been working on developing Human in loop interpretable machine learning models as lead data scientist at Microsoft.
Akshata Moharir
Akshata Moharir has more than 12 years of experience addressing complex business challenges across diverse industries such as support, gaming, retail, aviation, and education. She offers expertise in generative AI, responsible AI practices, predictive modeling, explainable AI, natural language processing, fraud detection, and predictive maintenance. She also has extensive experience with large language models, including fine-tuning and optimizing them for specialized applications. Moharir has contributed to machine learning and AI with seven granted patents, one trade secret, and seven scholarly publications. Skilled at leading cross-functional teams and delivering results that exceed expectations, she has a keen eye for detail and a commitment to excellence. She is also passionate about leveraging AI algorithms to drive innovation and deliver value to businesses and customers.
Correctly said