Unleashing Machine Learning: The Rationale Behind Smart Predictions
Network Science July 11, 2022
Machine Learning (ML) & Artificial Intelligence (AI) albeit recent technologies, are by no means new. Historically, machines have been used for several tasks and functions that were otherwise deemed as trivial and dirty for humans. Over time, these machines have evolved to take on more complex & sophisticated activities including decision making and strategy formulation. However, as complexities increase, there is a growing need for people to trust in the power of these machines rather than question their abilities.
A recent study by Accenture Labs has revealed that Machine Learning, especially deep learning, is quickly seeing an upsurge in its adoption in workplaces across industries. In healthcare, for instance, hundreds of companies are using Machine Learning algorithms and predictive analytics to reduce drug development time and diagnose ailments from medical images. Similarly, in the transportation sector, self-driving cars using ML are expected to become a norm within the next couple of years, with commercial applications of these automobiles being close behind.
What is Machine Learning?
These intelligent systems take on low-level pattern recognition tasks like image recognition, speech recognition and natural language processing to help companies churn large volumes of data for making specific recommendations. ML allows software systems to provide users with accurate predictions with minimal uncertainty.
The internal algorithms involved in this decision-making process, however, are often not visible to company personnel, making ML systems operate as ostracized “black boxes”. This makes organizations unwilling to allocate core competencies to machines due to higher risks of poor decision making and related costs.
Research indicates that in the upcoming years, machines will be compelled to explain their reasonings and recommendations in a deeper manner. As the next stage of human augmentation by machines, this interaction will enable people to understand and act responsibly. It will work towards creating an effective team between humans and machines.
Machine Learning Synergy
Intelligent systems powered by ML are now here to work alongside their human counterparts. By utilizing smart machines for responsibility, fairness, and transparency, organizations can enforce collaboration & efficiency within their workplaces. These advanced intelligent systems of the future, however, will not replace people. They will complement and support humans in a manner that allow businesses to make smarter, better and more accurate decisions.
There are 3 main market drivers for advanced ML-led systems. First, the growing need for transparency, as required by laws such as the EU’s GDPR, makes it essential for companies to disclose how personal data is being used for selection and other decision-making. Second, a growing need for trust between AI and human beings mandate that systems are able to effectively explain the rationale behind their decision-making. Third, the need for better machine-human synergy. With machines being better at recognizing minute patterns in large volumes of data and people being more efficient at connecting the dots among high-level patterns, the businesses of tomorrow are going to increasingly need both resources working hand-in-hand.
So, how exactly can Machine Learning convey the rationale behind outputs better? Research has identified three different methods to achieve this:
- Data-level explanation – Through this method, ML-based systems can provide evidence of the modeling and its results using comparisons made with other examples. This allows the system to justify the decision taken around any particular issue or targeted prediction.
- Model-level explanation – This approach focuses more on the Machine Learning algorithms. Through this method, the explanation provided makes the logic more understandable to humans by adding a layer of domain knowledge on top. Compared to the other methods, model-level explanation abstracts most from the data through rules or by combining it with semantics.
- Hybrid-level explanation – This approach works the best and is most useful if the data being studied is particularly large, complex or packed. The method uses a high level of abstraction by refactoring data at a metadata level. Rather than using the data as an evidence as in the case of other methods, the hybrid-level explanation offers explanation for every feature at a metadata level.
Enhanced ML will allow sophisticated systems to:
- Explain the reasoning behind their results and how they arrived at them.
- Characterize the system’s strengths & weaknesses.
- Compare their performance & output with those of other intelligent machines.
- Convey results in a comprehensive manner that showcases the potential of future technologies.
- Make the decision-making process in businesses smarter.