If you like what we're working on, please  star us on GitHub. This enables us to continue to give back to the community.

Why is Explainable AI Important for Predictive Models?

If you would like to contribute your own blog post, feel free to reach out to us via blog@deepchecks.com. We typically pay a symbolic fee for content that’s accepted by our reviewers.


As Artificial Intelligence (AI) algorithms grow more complex every day, it becomes challenging to fully understand the inner workings of our AI applications. The expanded use of AI for automation has increased the complexity and scalability of intelligent systems and, consequently, the requirement for clearness and explainability of AI systems. A human can understand a model that makes predictions by looking at the model parameters without much assistance. An interpretable model provides its own explanation in comparison.

Explainable AI (XAI) is a collection of instruments and frameworks used to help you analyze and understand the results generated by your AI models. XAI supplies techniques to demystify the predictions of your employed AI so that any further development can help humans understand the inner workings of these “black boxes,” helping us build trust in our model.

Knowing how the AI models reach their decisions is vital for their validity, clarity, and enhancement. XAI assists clients in trusting information that differs from their expectations, strengthening a company’s performance, and improving bottom-line effects.

Principles

The National Institute of Standards (NIST) determines four XAI principles:

Explanation obligates AI systems to provide appropriate evidence or logic for all outputs. This principle does not demand proof that the evidence is correct, informative or intelligible, but only that a system is capable of explaining.

The Meaningful principle is fulfilled if a user can understand the explanation and/or if the explanation is helpful in achieving an assignment.

Explanation Accuracy delivers explanations that are meaningful to users. These two principles do not demand that a system produces an answer that accurately reflects a process of a system for developing its outcome. It requires precision from the system’s explanations.

The last is Knowledge Limits. The prior principles implicitly consider a system working within its knowledge limitations. This, however, discerns whether or not systems recognise cases that were not designed or approved to operate or if their answers are unreliable. This practice identifies and declares knowledge limits so a decision is not provided when improper. This principle boosts trust in a system by controlling misleading, destructive, and unjust conclusions.

Techniques

The XAI techniques below are suitable for most AI consumers.

  • Data Visualization

Here, different but simple data visualization techniques can be used for data explanation and understanding, such as a radar plot, the tool used for this specific AI technique.

Radar plot

Radar plot (Source)

  • Local Interpretable Model-agnostic Explanations (LIME)

LIME is an XAI technique that approximates any black box AI model with a local, interpretable model to explain each individual prediction. It can present any model without specifying a value, hence model-agnostic.

Explaining with counterfactual method

Explaining with Counterfactual Method is used to describe the particular predictions of an AI model. The predicted value of a sample is computed from the feature values of the model, where the values of those features affect the prediction. Counterfactual explanations “interrogate” a model to demonstrate how much particular feature values would have to be modified to reverse the overall prediction.

  • Shapley Additive Explanations (SHAP)

SHAP is a game theoretical technique for describing the output of any AI model. It combines optimal values distribution with local explanations using the traditional Shapley values from game theory and their corresponding extensions. In the figure below, there are two vertical dotted lines. The gray line represents the expected probability that a stroke can happen to any patient represented in the dataset – approximately 50%. Red corresponds to the experimental data from a patient.

SHAP for stroke prediction

SHAP for stroke prediction (Source)

Open source package for ml validation

Build Test Suites for ML Models & Data with Deepchecks

Get StartedOur GithubOur Github

Solutions

XAI solutions have allowed us to understand AI models better. The most used and well-known are the solutions from Google and IBM.

Google’s XAI solution is a service pack of tools and frameworks integrated with many other Google products and services, such as Contact Center AI (human-like AI-powered contact center) and Document AI (automated data capture for document processing). The AI model’s performance can be debugged and improved using them. The What-If Tool, a tool for visually probing the conduct of trained AI models with minimal coding, can generate feature attributions for model predictions in AutoML Tables, BigQuery ML, and Vertex AI. AutoML Tables help build high-performing models from the tabular data.

BigQuery ML allows data scientists and data analysts to create and operationalize models on planet-scale structured or semi-structured data using SQL. Using Vertex AI, one can efficiently train and compare models using AutoML or custom code training, and all models are stored in one central model repository. Google’s solution provides the creation of interpretable and inclusive AI systems from the bottom up using mechanisms developed to notice and determine bias, drift, and other gaps in data and the XAI model. It also supplies end-user trust and improves transparency with human-interpretable descriptions of models. It simplifies the company’s requirements to handle and enhance AI models with streamlined implementation monitoring and training.

Google’s XAI

Google’s XAI (Source)

IBM’s AI Explainability 360 is a comprehensive open-source toolkit of state-of-the-art algorithms that reinforce the interpretability and explainability of AI models. They provide various explanations using algorithms for case-based reasoning, directly interpretable rules, post hoc local explanations, and post hoc global explanations, to name a few. It provides several thorough tutorials to familiarize practitioners on how to infiltrate explainability in other high-stakes applications. It also includes documentation that teaches the practitioner to select a suitable explanation method.

AI Explainability 360

AI Explainability 360 (Source)

Conclusion

XAI is established with an active community that has designed very successful procedures to describe and interpret predictions of complex AI models. These models are used in different industries to attain greater transparency.

Although briefly discussed, newcomers can use and explore the aforementioned XAI solutions to better understand its handy significance. You may find more information concerning XAI with the links below.

Subscribe to Our Newsletter

Do you want to stay informed? Keep up-to-date with industry news, the latest trends in MLOps, and observability of ML systems.

Related articles

How to Choose the Right Metrics to Analyze Model Data Drift
How to Choose the Right Metrics to Analyze Model Data Drift
What to Look for in an AI Governance Solution
What to Look for in an AI Governance Solution
×

Event
Identifying and Preventing Key ML PitfallsDec 5th, 2022    06:00 PM PST

Days
:
Hours
:
Minutes
:
Seconds
Register NowRegister Now