Download presentation
Presentation is loading. Please wait.
Published byWilfrid Blake Modified over 5 years ago
1
AILA – Artificial Intelligence Life Cycle Assurance
Jan 2019 © 2019 Cognizant
2
Template for Agenda, use as required
Content Template for Agenda, use as required AILA- Overview QA for AI AI for QA | © 2019 Cognizant
3
Where AI fits in QA and QA fits in AI
Traditional App Lifecycle AI/ML based App Lifecycle 1 Requirements Analysis AI/ML Model Requirement AI/ML Model Design & Data Analysis 2 Design 3 Development Deployment & Maintenance 4 Testing QA for AI 5 Maintenance Frictionless lifecycle automation Model Testing Model Dev & Training AI for QA QI BOTs | © 2019 Cognizant
4
How is Quality Assurance for AI different
QA for Traditional Apps QA for AI based Apps There is no defined Output Output is defined Expected behavior/result significantly consistent across dataset Expected behavior/result significantly varies across dataset Low functionality drift overtime High functionality drift overtime Very high need to regress application Minimal regression of application When we talk about testing or QA for traditional apps. Output or expected behavior is pre-defined. While in AI or ML based apps there is no predefined output or expected behavior. You might see low functionality drift and hence minimal regression required for traditional apps. While AI or ML based apps will have high functional drift overtime and hence requires high need to regress the application. In short we have “moving target” to test in AI/ML apps while we have “predefined target” for traditional apps. Thus QA is totally different when it comes to AI/ML apps. | © 2019 Cognizant
5
Artificial Intelligence Lifecycle Assurance - AILA
AI for QA Quality insight Bots framework QA for AI Data& model “AI for QA“ Use AI to test application intelligently Accelerator to assure Data & Model fitment Improve Appln Quality with Analytics QA for AI Process Framework QI BOTS studio customization framework AILA Enhancing Quality or ML/AI Process Jumpstart Bots execution platform “QA for AI” Assuring Quality of AI/ML models QA for AI Performance & Security iDashboard- Exec. dashboard powered by AI/ML Assuring AI/ML business stability Intelligent enterprise level metrics dashboard AILA has 2 components. AI for QA is to use AI to augment and optimize testing of software applications. And QA for AI is to do testing for AI/ml apps. AI for QA is “quality insights” and it does use data from SDLC and optimize testing of software applications. It has workbench or studio which can be used to develop/deploy ML bots to augment testing of software apps. It displays the outcome or results in dashboard for end user to consume and act upon insights to improve quality of software apps. QA for AI has 4 main components. We will see them in detail in next slide. It broadly assures quality of data, model, process and performance/security of AI/ML apps. | © 2019 Cognizant
6
QA for AI Assessing AI/ML model from QA Standpoint | © 2019 Cognizant
7
What is QA for AI? Good question. QA for AI is also a testing process.
I know software testing is a process used to identify the correctness, completeness and quality of developed computer software. What is QA for AI? Good question. QA for AI is also a testing process. It will also identify correctness, completeness, reliability and quality of “data and AI/ML” application. As you know data is the new “oil or electricity” And AI/ML process has 2 parts: Train ML model with data and Test ML model with data. And as you can see “data” forms a critical role in AI/ML process. Hence we need to “validate” data for it’s quality too. Interesting, why validate data specifically? | © 2019 Cognizant
8
Provide comprehensive quality assurance for AI/ML solutions
360° QA for AI - Overview Provide comprehensive quality assurance for AI/ML solutions 1. Assure Data Quality 3. Assess ML Process Quality Test data/feature for correctness, completeness and quality. This will remove any bias or insufficient data to train/test the model. Helps the model to perform well in production as in development. Assess ML process and frameworks from quality assurance perspective. Validate whether industry best practices are adapted in AI/ML process. Ensure the process is lean and efficient too. 4. Assure ML Performance, Security Quality 2. Assure ML Model Quality Test/evaluate ML model for it’s correctness, completeness and quality. Ensure best algorithm selected for the problem. It is providing required outcome for data driven decisions. Assure ML solutions on non-functional aspects including security and performance. Ensure that model is trained and deployed within allowable batch window. It is resilient for expected data volume spikes or poor quality of data. 360° QA for AI This slide talks about 36 degree view when it comes to QA for AI. We are covering the entire testing life cycle of AI/ML apps. Assure data quality – The key to any ai/ml app is data. Data fuels and defines an ai/ml app. Hence it is vital to validate data before we move forward with ai/ml app development. This first component ensures that we have all required data needed and also the correct data required. This will identify any bias in data which might influence the outcome of the ai/ml system. This assurance of data ensures that we have right data during development as we might see it in production. If there are any data gaps like missing values, not right data to build the model. Then this phase will highlight the same. Assume ML model quality – Once we have assured data which is required to build the AI/ML model. The next step is to build the model. It is important to make sure that model performs correctly. For example if the model is to capture anomalies in data. It should not miss out any anomalies and highlight the right data as “anomaly”. It will not serve the business purpose for which the model is built. Hence the results from model has to be validated and verified before moving on to next step. This phase ensures that best algorithm is used and also data driven decisions made using this model fits the business need (validate that true positive and true negative outcome and their importance for the business). Assess ML process framework quality – Since AI/ML is in it’s nascent stage. We need to make sure that entire process has followed industry standard best practices and also process is lean and efficient. This phase ensures the same. Assure ML performance, security quality – While we have validated the functional aspects of AI/ML application till now. It is critical that non-functional aspects like ML performance and security are also validated. For example model should be re-trained within allowable batch window and it cannot take days to complete training. And also it should be robust enough to handle any spikes or poor quality of data. Security should be validated to ensure that ML app is not vulnerable. *ML – Machine Learning | © 2019 Cognizant
9
1. Assure Data Quality 01 03 02 04 Correlation Analysis
Validate Data quality to pick and train the model for more accurate predictions 01 03 Correlation Analysis Data Bias Analysis Validate and filter potential Data where systems draw improper conclusions about data sets, either because of human intervention or as a result of a lack of cognitive assessment of data. E.g. Root cause predictor trained only with values related to “coding” will be flawed when provided with other root causes like “data issues”, “environment issues” Validate that no variable is left out which are correlated. E.g. “Root cause” is correlated to “priority”, “severity”, “description” in TriageBot Outliers Analysis 02 04 Trend Analysis Identify and exclude outlier data from the data set and refine the model E.g. Out of 100 Defect root causes one or two non standard category such as local hardware issues Assure all necessary data trends across time windows are duly considered in the data set E.g. Defect Seasonality prior to holiday seasons in retail industry due to application rollout and holiday freeze Let us quickly look at the first pillar for QA for AI – assure data quality. Correlation analysis – The variables or columns of data used to train ML model is called as features. We nee to ensure that we have selected the right features which is going to add more value to the outcome. Hence it is vital to do correlation analysis. It will correlate features to the output label and make sure that right features are selected for the model. Outlier analysis – Usually data will have outliers and that might become a noise while training ML model and impact the quality of outcome from the model. This step will ensure that outlies are analyzed and removed from data to train the model. Data bias analysis – Similar to outlier. Bias in data might impact the outcome. For example root cause predictor trained just with values related to “coding” as root cause might impact how well the model performs when it is given data related to other root causes like “data issues”, “environment issues” This step will ensure that we are identifying, removing any such bias in the data. Trend analysis – In most cases the data might follow certain trend. To relate to real world example. If we have Retail floor planned ML model. It should consider trend of not only “summer data” but also “winter” and other seasons. So that it performs well during other seasons data given as input. | © 2019 Cognizant
10
2. Assure ML Model Quality
Validate the ML model built for prediction accuracy and right model selection 03 A/B testing / Accuracy comparison / Quality metrics 01 ML Algorithms Support all type of ML Algorithms based on the requirement (like Classification, Regression and Clustering) Dynamically compare model accuracy across multiple models and identify best performing model 02 Model Frameworks 04 Regression Testing Support ML/AI models generated in leading technologies such as Python, R and Spark Perform regression testing on models and qualify model performance. Good or bad against current production model. This phase is to assure that model built is “correct” for the purpose and gives “right” results. ML algorithms – There are many types of algorithms like regression, classification and clustering to name a few. QA for AI will support all type of major supervised/unsupervised ML algorithms. Model frameworks - ML models can be built in various technologies like Python, R and Spark. QA for AI can support all technologies. Quality metrics – It is key to validate whether model we have in hand is the best out of all other models. This will be objectively measured and selected using industry standard metrics like AUC, F1 score, Confusion matrix. This is automated using various utilities in QA for AI. Regression testing – While quality metrics will help with model in development. We need to regress new ML model out of development and qualify that it is ready for production. Hence it is critical to do regression testing of such a model built and ready for production. | © 2019 Cognizant
11
3. ML Process Quality Assurance
Assess maturity of Machine Learning process, frameworks of organizations and provide comprehensive recommendations on improvement areas. Assess ML process with regard to business value definition. Assess algorithms and data processing pipeline usage with regard to business problems. Process 01 04 Assess ML Frameworks on features, integration and scalability aspects. Assess support for processing of structured and unstructured data. 02 05 Technology Value Assess metrics and measurement techniques of model performance. Assess ML projects with regard to model deployment, governance, ethical assurance and robustness. 03 06 Process quality assurance will be driven by questionnaire and not utilities ML process quality is assurance is more of consulting objectively on existing ML process. Assess ML process with regard to business value definition. That whether required process in place to cater to the need of business like millisecond prediction can be validated or is it more of offline prediction which does not require more rigor. Assess ML Frameworks on features, integration and scalability aspects so that required validation is performed quickly and efficiently. Assess metrics and measurement techniques of model performance. Ensure that industry standard metrics are used for the purpose. Assess algorithms and data processing pipeline usage with regard to business problems. Validate whether pipeline in place can solve any algorithm or data related issues. Whether it is lean and efficient. Assess support for processing of structured and unstructured data. Data ingestion from unstructured and structured data sources has standard tools like talend, Logstash and processes. Assess ML projects with regard to model deployment, governance, ethical assurance and robustness. Validate frequency in which model is validated and whether it is sufficient for the governance of the model. | © 2019 Cognizant
12
4. Assure ML Security and Performance
Assure ML solutions on Non-Functional aspects including security and performance. Validate predictive response time of the models at varying user loads. Validate model with regards to train and retrain time during learning phase. Predictive Response Time Train-Retrain Time Validate scalability of models with regards to volume of data. Validate ML models on its robustness with regard to CPU usage, memory consumption etc., Scalable Models Model Robustness As you can see this step is make sure that ML models security and performance validated. 1. It ensures that we have model which has predictive response time. Depending on peak user loads or transactions, the model should yield the outcome within acceptable time window. 2. Similarly train and re-train time during learning phase should remain with the limits like it should not take weeks to train a model rather it should be within a nightly batch window. 3. Similar to user loads, the model should perform with influx of huge volume of data and entire stack should be scalable to handle the spike in volume of data. 4. Model robustness can be validated against standard performance tuning aspects like CPU usage, memory consumption. | © 2019 Cognizant
13
ML Assurance Workbench – Simplifying & Industrializing ML Assurance
“Enabling QA for AI through ML Assurance workbench” GUI based ML assurance platform Validate Model and Data Quality over industry defined metrics & measurements ML model comparison dashboard across varied data set & parameters Monitor model regression across different sectors and make informed decisions Automated feature impact assessment to determine model stability based on feature relationship | © 2019 Cognizant
14
AI for QA Use AI to test application intelligently | © 2019 Cognizant
15
Artificial Intelligence for QA activity through Quality Insights
Creating intelligent and agile systems that deliver results at a fraction of the cost DO LEARN Replicate repetitive human actions Learn to understand context, adapt to users and systems THINK Learn to understand context, adapt to users and systems Handle judgment-oriented tasks | © 2019 Cognizant
16
Quality Insights – Insights powered by BOTS
Eliminate Non Productive Testing Assure Immersive Customer experience Greater Business Requirements Coverage Close the feedback loop by learning from logs Proactively Plan the Test Process BizInsight DevInsight TestInsight OpsInsight CxInsight Requirements Coding & Unit Testing System &SIT Operations Customer Feedback | © 2019 Cognizant
17
BizInsight : Powered By BIZBots
Data Source User stories Defect logs Test Execution logs Defect BackLog User Story Points Defect predictor (biz) Predict defects early in the cycle by application of advanced machine learning techniques on user stories and defects and Predict count of defects in upcoming release using key parameters relating to defects in past releases TC Failure predictor (biz) Perform risk based testing by identifying the failure test cases from user stories based on pattern analysis. HOW DO I WORK ? Defect backlog pruner Prune and optimize defect backlog by identifying priority defects based on active user stories and invalidating stale defects that are no longer business relevant. Story Point Predictor Predict the user story point by application of clustering machine learning algorithm on user stories.
18
DevInsight: Powered By DevBots
Build verification predictor Predict success percentage of build based on SCM & build logs TC Failure predictor (dev) Helps to identify impacted test cases from file changes during the release from SCM logs HOW DO I WORK ? Defect predictor (dev) Predict defects based on file modification patterns from the SCM logs Predict defect based on duplicate code analysis Data Source SCM logs Code Build logs Defect Logs Test Execution Logs
19
TestInsight: Powered By TestBots
Regression Test Optimizer Identify test scripts that are similar by applying text analytics and clustering algorithms on http, https logs from client during test case execution Duplicate Defect Identifier Identify duplicate defects by using keyword matching techniques HOW DO I WORK ? Defect Resolution Time Predictor Predict defect resolution time by analyzing historic defects and resolution time Data Source Defect logs Test cases Test execution logs Test plans Schedule Variance Predictor Predict schedule variance using the historical data. Traceability Mapper Automatic Traceability mapper that generates traceability between Defects and Test cases using advance text analytics. Defect AI Assist Identifies frequent failures and to whom defect can be assigned.
20
OpsInsight: Powered By OpsBots
Testing Gap Analyzer Using App Server Logs Identify gaps in testing process by establishing the correlation between production logs and test execution logs. TriageBot Predict root causes of the defects using ML based solution that learns from historic defect data. Testing gap analyser using db foot print Identify testing gap in test data by comparing patterns of data from production and test databases key columns. HOW DO I WORK ? Data Source Application Logs Database Logs Server Logs Event Logs Error Logs Performance Logs Test Mgmt. data Database records Truck Roll Predictor Predict whether truck roll can be avoided using predictive analytics model Firmware update Predictor Predict whether firmware upgrade is required using predictive analytics model Test data distribution analyzer Algorithm based solution that plots data distribution patterns and identifies data mining attributes such as unique combination of data.
21
CXInsight: Powered By Customer Experience Bots
Customer Feedback Analysis BOT Helps capturing the voice of the customer and the underlying human emotions with sentiment analysis and clustering end-user feedback into actionable QA attributes like Functionality, Performance, Usability etc HOW DO I WORK ? Data Source Mobile App reviews (Playstore & iTunes AppStore) Social Media Comments/Tweets End User/ Survey Feedback in Excel format Conversational BOT Help user in selecting suitable bot in a conversational way/manner.
22
BOTs in a DevOps workflow
Req. Dev. Test Build Prod. Production Environment Req. Mgmt Test Design Design Test Plan Develop Unit Testing Build Version Control Deployment (QA) Test Execution Test Results Defect Backlog Pruner Defect Prediction Test Case Failure Prediction Reduced Defect leakage Faster QA cycles Improved visibility BIZ INSIGHT ADPART Defect Prediction TC Failure Prediction Improved visibility Increased speed Lower cost ADPART inDev. DEV INSIGHT Build Verification Predictor DEV INSIGHT Reduced Defect leakage Increased speed Continuous Feedback / Improvement Increased speed Quicker Defect resolution Improved visibility OBJECT PROBE DEV INSIGHT ADPART TEST INSIGHT OPS INSIGHT OPS INSIGHT Defect Prediction TC Failure Prediction Duplicate Defect Identifier; Test Data Distribution Analyzer; Regression Test Sequencer & Optimizer; Testing Gap Analyzer; Traceability Mapper Testing Gap Analyzer; Triage BOT; DashBot Survey BOT; Chat BOT Solution Selector BOT Customer Insight Increased visibility Better Quality Lower cost iDashboard CX INSIGHT BizInsight DevInsight TestInsight OpsInsight CXInsight Cognizant IPs | © 2019 Cognizant
23
Bots Studio –Industrializing BOTs Development and Deployment
“Simplifying AI for QA through a single platform” * Drag & Drop enabled visual workflow designer platform for the analytics solution development Simplifies development effort and the need for niche team composition Pre-built catalog of ready to use BoT for common use cases Ability to qualify and validate analytics model built with comprehensive metrics and measurement techniques Enables faster deployment to production through one click deployment BUILD & CONFIGURE TRAIN & TEST DEPLOY | © 2019 Cognizant * H Roadmap item
24
Thank you Digital Assurance CoE © 2019 Cognizant
Similar presentations
© 2025 SlidePlayer.com Inc.
All rights reserved.