Model Interpretability Solution

Model Interpretability Solution

Model Interpretability

Treasure Mapping

ML Operations

System Control

Model Development

Forecasting & Prediction

  1. Build a model.
  2. Is my model biased or unfair?
  3. What’s the potential ROI impact? What’s the potential PR impact?
  4. How can we fix the model?
  5. Fix the model.

Phase 1: detection

Phase 2: reconciliation

Executive Summary

Bias in quantitative models leads to unethical business practices and reduced ROI. If your business relies on machine learning to make business decisions or serve answers to customers, you must perform model bias investigation and correction. Especially in the application of deep learning models, but applicable for any model type, the discovery of model bias is critical. Don’t find out when your summons arrives in the mail. Let Expero’s expert team of machine learning professionals perform an audit of your models to discover biases and correct them. We work with your existing data science team to uncover hidden irregularities in your models, and work together to correct those inconsistencies. Work in an industry regulated by oversight? No problem! Expero works with regulatory oversight bodies to enable your business to deploy sophisticated, modern techniques even in industries with sensitive data.

Solution Modules

Engagements

Solution Engagements

2-5 Day Assessment

3-5 Week Experiment

3-6 Month Project

Data assessment focused on business goals and user enablement.
Working data pipeline prototype for data/user subset, from experimentation to deployment.
Working data pipeline, scalable to your organization from experimentation to deployment.

Deliverables may include

  • Report on suggested data and model analysis
  • Exploratory data analysis notebooks and findings

Prerequisites requested

  • Sample data
  • Detailed documentation of model application space

Customer team members requested

  • Project manager - business goals
  • Technical lead - technology scope

Deliverables may include

  • Report including detailed model biases and correction strategy
  • Measurements of bias magnitude w.r.t. inputs

Prerequisites requested

  • Access to full data: inputs, model outputs, and groundtruth
  • Trained model(s)

Customer team members requested

  • Technical lead - technology scope
  • Technical implementer - specific model requirements

Deliverables may include

  • A corrected model with corresponding I/O data
  • Measurements of bias magnitude w.r.t. inputs

Prerequisites requested

  • Access to full data: inputs, model outputs, and groundtruth
  • Trained model(s)

Customer team members requested

  • Technical lead - technology scope
  • Technical implementer - specific model requirements

No items found.
  1. Build a model.
  2. Is my model biased or unfair?
  3. What’s the potential ROI impact? What’s the potential PR impact?
  4. How can we fix the model?
  5. Fix the model.

Phase 1: detection

  • Is my model biased?
  • Common methods: feature factor analysis, feature correlation, model I/O perturbation, and more
  • We require: input data, output predictions, output groundtruth, and a trained model
  • We can deliver: report on a model’s bias and sensitivity, evaluation of potential magnitude of impact on business, possible strategies for fixing the model bias, a corrected model.

Phase 2: reconciliation

  • How can I fix my model?
  • Common methods: latent space analysis, concept activation vectors, neuron factor analysis, and more
  • We require: input data, output predictions, output groundtruth, and a trained model
  • We can deliver: model’s current interpretability score, a system to perturb model latent space and observe outputs, a strategy for fixing model, a corrected model.

View our Gallery

Included Technologies

Preferred Technologies

Why Our Solution is Different

Model interpretability

  1. Build a model.
  2. Is my model biased or unfair?
  3. What’s the potential ROI impact? What’s the potential PR impact?
  4. How can we fix the model?
  5. Fix the model.

How Does It Work?

Phase 1: detection

  • Is my model biased?
  • Common methods: feature factor analysis, feature correlation, model I/O perturbation, and more
  • We require: input data, output predictions, output groundtruth, and a trained model
  • We can deliver: report on a model’s bias and sensitivity, evaluation of potential magnitude of impact on business, possible strategies for fixing the model bias, a corrected model.

Phase 2: reconciliation

  • How can I fix my model?
  • Common methods: latent space analysis, concept activation vectors, neuron factor analysis, and more
  • We require: input data, output predictions, output groundtruth, and a trained model
  • We can deliver: model’s current interpretability score, a system to perturb model latent space and observe outputs, a strategy for fixing model, a corrected model.