The Next Applied
social share alt icon

IMPROVING TRUSTWORTHINESS OF AI MODELS

Know More

DEVELOPING UNDERSTANDABLE MODELS

 

AI/ML based solutions are becoming ubiquitous, solving many critical problems impacting people’s lives. Banks and financial institutions rely on AI based solutions to improve user experience and take lending decisions. Insurance companies use AI to identify policy claim frauds. Recruiters leverage them to identify the best candidates. With the pervasiveness and the black box nature of these complex models, there is a growing need for models and solutions to be responsible and accountable.

Responsible AI develops artificial intelligence in a way that is inclusive and understandable by users. It provides explanation to models that make AI trustworthy, accountable and bias-free. This improves user confidence and ensures regulatory compliance. It allows all stakeholders to understand how the model predictions change with different inputs. Regulators can check for biased and discriminatory practices and decision makers can harness the potential of AI while maintaining the risks.

OUR SOLUTION

 

Mphasis Responsible AI is an end-to-end framework that enables companies to develop and deploy robust, interpretable, explainable, bias-free, auditable, and privacy preserving AI through a unique user experience and design thinking engagement. This makes the system trustworthy, thereby improving customer experience, reducing liability risk and ensuring regulatory compliance.

The Responsible AI components are generic and modular, enhancing scalability and repeatability across several use cases. For example, global and local explanation module understands internal logic and model limitations; bias identification and mitigation module assures model fairness; PII redaction module preserves privacy and so on. The ability to log experiments and model versions allow for explanation-accuracy trade-off analysis and auditability. Mphasis Responsible AI framework is fully integrated with PACE-ML, Mphasis’ proprietary MLOps framework for easy deployment.

 

KEY TENETS

 

High performing: Machine learning models achieve high level of performance in terms of identified metrics (such as accuracy, sensitivity, specificity, F1 score, log-loss, etc.)

Interpretable: A cause and effect can be observed where a human is able to predict the change in output given a change in the models’ input

Explainable: The internal mechanics of models can be explained in human terms

Auditable: Models’ actions and the attributes driving them are recorded with integrity and readily available for scrutiny

Bias Free: The models are impartial; they work without giving undue advantage/disadvantage to any class

Privacy Preserving: The models are free from any PII to respect privacy of all users

KEY FUNCTIONALITIES

 

Solution Design

  • User Experience & Design Thinking Workshops: Mphasis’ proprietary client engagement model leverages unique user journey mapping and design thinking workshops involving all the stakeholders, to create the most effective AI solution to address the root causes.
  • Gap Analysis: Gap analysis / assessment is used to assess the maturity of business processes with respect to applicability of specific AI/ML solutions.

Solution Design

Model Performance

  • Model Debugging: Performance is measured across a range of inputs to identify if there are any anomalies. Comparison with competing models or new scenarios can also be performed.
  • Monitoring: Performance metrics are continuously monitored to check for any degradation. Change in model behavior from accepted standards results in alerts for model refresh.
  • Data and Model Drift: Changes in data distribution and processes/policies are captured though drift analysis and incorporated back into the models through re-training, to maintain their performance.

Model Performance

Privacy Preserving

  • PII Protector: Anonymize PII to prevent falling into the wrong hands. Secure the anonymized mapping using access privileges.

Privacy Preserving

Bias Check

  • Bias Detection: Detect if models are positively or negatively biased towards a class and check for data fairness before initiating model development.
  • Bias Mitigation: Human interventions and automated techniques are used to mitigate pre-known or identified biases. Bias mitigation can be done before, during and after initiating model development.

Bias Check

Interpretability & Explainability

  • Highlighting Key Features: Global explanation techniques explain how the model is doing predictions.
  • Decision Explanations: Local explanations techniques explain why a specific decision was taken for the given input.

Interpretability & Explainability

Auditability

  • Tracking: The complete lifecycle and components (data, code & models) of AI/ML projects are versioned and tracked.
  • Logging: All artifacts like metrics, explanations, predictions, etc., are logged for post-hoc scrutiny and analysis throughout the AI/ML project lifecycle (development, deployment and operationalization).
  • Governance: Methods that resist the tampering of audit trails are deployed to keep track of data usage and model development, for later scrutiny. This helps in comprehensive auditability and error debugging.

Auditability

BENEFITS

 

Improves model fairness: Allows developers to check if AI predictions discriminate against a class. This helps to mitigate biases and results in fair outcomes.

Ensures regulatory compliance: Increases transparency of AI models and helps understand reasons for a negative outcome.

Improves confidence in critical business decisions: Useful while the AI models take critical decisions that impact liabilities of the organization.

Simplifies model governance: Various steps in model development are available for scrutiny throughout its lifecycle. This includes post-deployment or even retired models.

Increases user trust: Produces human-centered intuitive explanation, thereby building user satisfaction and trust.

Eases deployment: Leverages pre-built components that offer fast deployment and customization to suit varied business contexts and requirements.

Cloud-agnostic and hybrid friendly: Can be deployed on-prem and any private or public cloud.

Flexible integration: Stand-alone deployment to explain AI models in production or integrated with other solutions.

SUCCESS STORIES

 

MPHASIS RESPONSIBLE AI SOLUTIONS
ON AWS MARKETPLACE

 

Through AWS Marketplace, Mphasis offers AI solutions that deliver immediate results and ROI in critical enterprise business processes and operations. These solutions can be deployed with the speed and security provided by AWS. Our presence in the AWS marketplace has facilitated the expanding global footprint of Mphasis AI offerings, helping developers get hands-on experience on AI algorithms and layer AI based solutions in enterprise applications. Combined with AWS services, these AI solutions help in simplifying data experimentation, extract deeper insights from data estates and improve productivity for a variety of use cases.

Our current Responsible AI listings are:

THOUGHT LEADERSHIP

 

Using Automated Machine Learning to Detect and Mitigate Demographic-oriented Biases

 

Demystifying Bias Detection and Mitigation in Machine Learning

 

Responsible AI in Large Scale Machine Learning System

 

Explainable AI for Mitigating Biases in Judgment Systems