While practically every aspect of our world continues to be reshaped by Artificial Intelligence, Data Scientists and AI/ML enthusiasts are excited about the Responsible AI framework which ensures that the decisions taken by AI algorithms are accurate! With increasing dependence on AI to make human-like decisions – right from our social media feed to critical health care diagnosis, it becomes crucial that these predictive outcomes are explainable, secure, ethical & unbiased!
Importance of Responsible AI
Let me share an interesting story with you, in 2018, a large MNC was all set to roll out a new
AI-assisted resume screening application. This solution leveraged ML algorithms to train on past 10 years of multi-location recruitment data. In the testing phase, the team noticed a big flaw in the system – the solution was absolutely biased towards males & discriminated against the female candidates! Although unintentional, the model did train on historical data that had bias. They then decided to shelve this solution and go back to the drawing board, albeit this time, more Responsibly!
Key Components of Responsible AI
If you are a Data Scientist or an AI/ML enthusiast, here is a framework of 5 Key components of Responsible AI that you should know about:
#1 Reproducible AI: Helps recreate a specific ML workflow.
Experiment Tracking for Iterative Development – Data scientists experiment with multiple ML training algorithms. Version control for these experimental runs as well as model registry is recommended to standardize processes for tracking progress and reproducibility.
Automated System – Automated systems greatly help in faster iterative solution development. Adopt MLOps’ tools for building CI/CD pipelines for productionizing these training models as well as monitoring ML pipelines for data & model drift.
#2 Transparent AI: Makes ML models interpretable and answers ‘why’ for each prediction.
Build Trust for Critical Applications – ML models learn patterns from data and often provide predictions for critical tasks such as clinical decision support in healthcare applications. Experts would like to know ‘why’ the model predicted these results.
Ensuring Correct Patterns – Zooming out from one prediction to the other, over the complete dataset, helps experts understand if this model has picked up correct patterns for cancer and built trust in this solution over time. We would recommend adopting tools such as LIME and SHAP for explaining results right from the development to deployment phase.
#3 Accountable AI: Builds ethical AI solutions using tools for analysis of data and model input features.
Align Outcomes to Current/Future Goals – ML algorithms observe patterns in historical data of certain workflows or facts which may not align exactly when analyzed from an ethical POV. We recommend you form an ethics committee and involve the members in the iterative model building process.
Mitigate the Risk of Bias – As part of the Responsible AI framework, you can conduct regular analysis of the training data and trained model for quantitively measuring fairness.
#4 Secure AI: Ensures endpoint cyber and data security for interfaces that host the AI/ML models!
Adversarial Testing is Crucial – Organizations spend a lot of time and energy in building AI solutions and their artifacts. It’s important to guard this model as it’s the IP of the organization hosting it. Adversarial testing is key to discovering vulnerabilities and avert possible attacks.
Safe Handling of Client Data – ML models require client data to be sent across to the API endpoints for running inference. Client data needs to be handled with extreme caution and the integrity of execution is paramount. Hardware backed Confidential Computing solutions can balance the scenario of hosting models on cloud while protecting client data as well as the integrity of model inference.
#5 Private AI: AI Solutions should protect sensitive user data and should be compliant with privacy regulations such as GDPR.
Move ML Training to the Edge –Techniques such as Federated Learning push training to the edge such that data does not leave its source while providing a way to generate insights from private data and improve ML models.
Privacy Preserving Machine Learning (PPML) –The dichotomy between aggregating data for improving ML model capabilities while also meeting data privacy regulations has given rise to a new sub-domain of ML known as PPML. Federated Learning, Differential Privacy and Encrypted Training can help you adopt PPML faster.
As AI/ML solutions mature from being just experiments run by data scientists and AI/ML enthusiasts to more ‘production ready’ deployable solutions across domains, they must leverage a Responsible AI framework for accurate predictions. It’s important that we build secure endpoints for serving these models and the software building process needs to be updated to accommodate dynamic AI Solutions which will change over a period of time. With more and more organizations adopting Responsible AI framework; I am confident that AI solutions will continue to evolve to become highly trustworthy, transparent & auditable in the near future!
Authored by Amogh Kamat Tarcar, Lead Data Scientist, Persistent Systems