What is Transparent Machine Learning?

Machine Learning (ML) is an involved science, its models complex, not easy to understand. Transparent ML explains itself – its working, its prediction, its insights – so that the user understands it. There are various mechanisms used for Transparency – Shapleys and surrogate models, being two important ones. The sequel explains this in more detail.

How is Transparent Machine Learning Useful?

Any enterprise, especially financial institutions, have reams of data, some of it for regulatory compliance, some other for operations, and so on so forth. ML uses this data for two important reasons:

• Supervisory training of a statistical model to make predictions. This automates the enterprise, implements consistency in its business processes.
• Extract the intelligence from the data. The data has buried in it, critical intelligence about the dynamics of your business. This intelligence has considerable value about your business, the rules it follows, many of which you would well know, while some others may still surprise you – data has a way of objectively busting some popular myths.

Together, the two are a formidable tool for your business: the latter exposes bottlenecks in your business, that if undetected, could cost your business many millions; the former allows you to do what-if experiments to fine-tune and optimize your business.

Specifically, there are many use cases that rely on Transparent ML: regulatory compliance, rich customer experience with actionable intelligence, model enhanceability, training data bias, prescriptive analytics, and importantly, trusted man-machine collaboration.

What are the Details of Transparency?

Before we discuss the guidelines for Transparent ML, let us examine its three ways in which it manifests itself in ML platforms.

1. Explainable AI: Explain the reasons behind a ML model’s prediction. A reason is expressed as a rule of various predictors, each with its threshold, to make it actionable.
2. Augmented Intelligence: Unlike Explainable AI that explains a particular prediction for a record, Augmented Intelligence extracts the intelligence – the insights about a business – from the entire data set, that help improve decision making.
3. Traceability for Data Processing: Explains at each step, in context of your data, how it is processing it, why it chose that particular way – so that you can feel comfortable in the results, or override it, in case you don’t.

EazyML supports all the three ways.

As is evident from the three ways, the key theme for Transparency is trusted man-machine collaboration in the enterprise.

For purposes of this report, we will restrict the discussion on Transparency to points (1) and (2) only.

What must the Customer do?

1. Select the right ML platform: First, select a platform that does Explainable AI and Augmented Intelligence correctly. This means that the ML platform must associate a confidence score with each prediction or insight; you only select those predictions or insights that have confidence scores above a configured threshold – the higher the score, the more confident is the ML platform of its correctness. EazyML allows you to do that. Verify these results against the data, filtering it to carefully study whether the results were correct.

   Equally importantly, the criteria for platform selection should include the technical capability of the vendor to support your project for Transparency.

2. Define the scope of the project: Propose a well-defined and narrow scope of the trial for Transparent ML. Preferably, select a project that would benefit considerably from the output of Transparency. It is desirable that the project address an important function (please refer some of the use cases in section 2, for instance), but to alleviate project risk, address it in small incremental phases, each phase with well-defined deliverables, and each phase building on the previous, as you assemble them for achieving the big vision.

3. Get support from professional services: Work with the professional services that has experience on the platform; they can share knowledge with your organization on how to interpret, generalize and refine the results of Transparency, making you self-sufficient. As you experiment with the platform, configure the various parameters, their thresholds, so that the platform performs with your tolerance of error, your precision for automation. Please study the platform guide to study the various parameters, their recommended values and ranges.

   EazyML is ideally suited for this; the team at EazyML comes with the platform to ensure that you hit the KPIs for success of the project.

What must the Vendor do?

1. Training session for the users: EazyML has two interfaces – Graphical User Interface (GUI) and Application Programming Interface (API). GUI helps you get started quickly, without considerable investment of time. It allows you to configure parameters related to Transparency. The results of Explainable AI and Augmented Intelligence are displayed in simple and easy-to-understand tables and charts. Suppose you want to stitch your own data pipeline: for instance, you have your favorite model for prediction, or a tool you like for visualization, and want to use EazyML only to explain your model’s predictions transparently. You would have to use APIs to access EazyML’s Transparency functions.

   Very importantly, for Explainable AI to explain the correct reasons, requires sophisticated analytics and considerable compute power – for instance, it will have to determine the probability density function of the training data. This can only be done via API, not GUI which is not as powerful an interface.

2. Help scope the project: The sequel discusses some guidelines for scoping. Transparency can get involved. It’s important to limit its scope, start off with a smaller project, which is well understood by subject matter experts so that they can reason it well. It must have adequate training data (because buried therein lies the reasons) and a model for prediction that is reasonably accurate (incorrect predictions will be explained incorrectly as Transparency does get influenced by the original black-box model).

   Preferably, the scope will select a project that’s narrow in its impact – alleviates project risk, and importantly, helps us learn the concepts about Transparency by easily observing them in action. At the same time, we encourage you to select an important project, where Transparency eliminates a key bottleneck, making the decision makers clearly see the power of Transparent ML.

   We urge you to engage the vendor, it channel partner, to help with this important step. Their experience will help you prune the list of potential projects. Additionally, Transparency can be compute-intensive; EazyML vendor will help you size the compute resources.

3. Help execute the project: The previous sections have discussed how professional services – expert data scientists who have worked on EazyML – can assist you:

   • Define the scope, provide training, be available as a team member, take feedback from you to influence EazyML’s roadmap.
   • Execute specific tasks, especially related to APIs and its demand on compute resources
   • Define the metrics, its KPIs, to measure the success of the project in implementing Transparency.