PaperlayerView exact revision on arXiv

Source-linked AI summary

An Artificial Intelligence Life Cycle: From Conception to Production

Daswin De Silva, Damminda Alahakoon

arXiv:2108.13861v1cs.SE

TL;DR

The paper addresses the need for a comprehensive process for designing, developing, and deploying AI systems from conception to production. It proposes the CDAC AI Life Cycle, a three-phase, 17-stage framework integrating problem contextualization, data and model development, evaluation, operationalization, and organizational alignment. The authors conclude that the framework and its algorithm-to-application and organizational mappings can support AI awareness, transparency, and stakeholder coordination.

  • Problem

    AI initiatives require a comprehensive life cycle covering their progression from conception to production.

  • Method

    The paper proposes the CDAC AI Life Cycle with Design, Develop, and Deploy phases, 17 stages, algorithm-to-application mapping, and organizational context.

  • Results

    The resulting framework covers problem contextualization, data preparation, model development and evaluation, deployment pipelines, and continuous monitoring across the AI life cycle.

  • Takeaways & Limitations

    The life cycle is intended to support AI awareness, knowledge, transparency, informed stakeholder discussions, and integration with organizational strategy.

  • Takeaways & Limitations

    Operationalizing AI involves cross-functional complexity and silent errors during debugging.

Abstract

from arXiv · show

Drawing on our experience of more than a decade of AI in academic research, technology development, industry engagement, postgraduate teaching, doctoral supervision and organisational consultancy, we present the 'CDAC AI Life Cycle', a comprehensive life cycle for the design, development and deployment of Artificial Intelligence (AI) systems and solutions. It consists of three phases, Design, Develop and Deploy, and 17 constituent stages across the three phases from conception to production of any AI initiative. The 'Design' phase highlights the importance of contextualising a problem description by reviewing public domain and service-based literature on state-of-the-art AI applications, algorithms, pre-trained models and equally importantly ethics guidelines and frameworks, which then informs the data, or Big Data, acquisition and preparation. The 'Develop' phase is technique-oriented, as it transforms data and algorithms into AI models that are benchmarked, evaluated and explained. The 'Deploy' phase evaluates computational performance, which then apprises pipelines for model operationalisation, culminating in the hyperautomation of a process or system as a complete AI solution, that is continuously monitored and evaluated to inform the next iteration of the life cycle. An ontological mapping of AI algorithms to applications, followed by an organisational context for the AI life cycle are further contributions of this article.

1 Introduction

The paper introduces the CDAC AI Life Cycle as a framework for designing, developing, and deploying AI systems, grounded in extensive AI practice and research experience.

  • 1 Introduction: The life cycle is motivated by AI’s expanding influence across national, social, economic, personal, and industry contexts.The paper situates its framework within AI’s increasing impact and importance.
  • 1 Introduction: The authors present the CDAC AI Life Cycle to characterise the design, development, and deployment of AI systems and solutions.The framework draws on more than a decade of academic research, technology development, industry engagement, teaching, supervision, and consultancy.
  • 1 Introduction: The authors aim to increase awareness, knowledge, discussion, and debate that can inform AI research, applications, and policy.

2 The CDAC AI Life Cycle

The CDAC AI Life Cycle organizes AI initiatives into three phases and 17 stages, beginning with contextual problem formulation and continuing through data preparation.

  • 2 The CDAC AI Life Cycle: The life cycle comprises Design, Develop, and Deploy phases containing 17 stages from conception to production.Figure 1 links the phases to AI/data scientists, AI/ML scientists, and AI/ML engineers, with team size corresponding to project size.
  • 2.1 Identify and formulate the problem: Problem formulation defines the environment, entities, data, current solution, alternative flows, rules, storage, and available historical data.
  • 2.2 Review AI literature - ethics, algorithms and pre-trained models: The design process reviews ethics guidance, state-of-the-art algorithms, models, and pre-trained models applicable to the problem.Examples of reusable pre-trained models include AlexNet, ResNet, BERT, and GPT.
  • 2 The CDAC AI Life Cycle: Data preparation addresses unified data representation, governance, quality, acquisition, transformation, missing or erroneous values, and formatting.The framework considers data ownership, stewardship, metadata, ethics, regulations, stakeholder trust, and transparency.

2.7 Data augmentation

The development stages augment and prepare data, map application capabilities to algorithms, build initial models, benchmark them, and iteratively refine complexity and generalization.

  • 2.7 Data augmentation: Data augmentation includes oversampling or undersampling, transfer learning from synthetic data, feature engineering, and feature representation techniques.Examples include dynamic time warping and vector symbolic architectures.
  • 2.7 Data augmentation: AI capabilities are mapped to algorithms, while parameters and architecture are selected to represent the problem domain.
  • 2.7 Data augmentation: The initial AI model is trained, validated, tested, and benchmarked using common-sense heuristics such as human expertise.The benchmark evaluates both the model and the dataset.
  • 2.7 Data augmentation: Subsequent models focus on what the initial model fails to capture, using tuning, regularization, preprocessing, and gradually increased complexity.Gradual complexity increases are presented as supporting model explainability, while regularization can support generalization after overfitting.

2.11 Evaluate primary metrics

The framework evaluates AI models with appropriate metrics across parameter settings and examines their trade-offs and explainability through intrinsic and extrinsic methods.

  • 2.11 Evaluate primary metrics: Model evaluation requires selecting metrics that are accurate, robust, agnostic, scalable, and interpretable.Candidate metrics include accuracy, loss, error, precision, recall, and convergence.
  • 2.11 Evaluate primary metrics: The framework interprets outcome as model + error when assessing evaluation metrics.
  • 2.11 Evaluate primary metrics: All models are compared under default and fine-tuned parameter settings, including investigation of the bias-variance trade-off.
  • 2.11 Evaluate primary metrics: Model outputs are explained using intrinsic or extrinsic techniques, including PDP, ICE, LIME, and SHAP.The framework also considers algorithm extensions that provide explainability by design.

2.13 Evaluate secondary metrics

Deployment assessment extends beyond intelligence performance to computational effectiveness and operational fit. The evaluation considers resource use, execution demands, deployment context, and expected service requirements.

  • 2.13 Evaluate secondary metrics: Computational effectiveness must be assessed alongside intelligence performance before operationalisation.The proposed metrics include CPU, memory, complexity, and convergence.
  • 2.13 Evaluate secondary metrics: Deployment planning should weigh model compression and whether execution occurs on-premises, on-device, or in the cloud.
  • 2.13 Evaluate secondary metrics: Model deployment is also termed model serving, model scoring, or production.
  • 2.13 Evaluate secondary metrics: The deployment approach should be selected for time efficiency and adaptability, including real-time versus batch prediction.
  • 2.13 Evaluate secondary metrics: Operational requirements include users, applications, output formats, turnaround time, and frequency of use.

2.15 Operationalise using AI pipelines (MLOps, AIOps)

AI operationalisation moves from standard deployment toward containerised, microservice-based pipelines. It requires coordinated service management, data and AI pipeline activities, monitoring, and attention to cross-functional debugging complexity.

  • 2.15 Operationalise using AI pipelines (MLOps, AIOps): Operationalisation moves from standard deployment to AI pipelines built with containers and microservices.
  • 2.15 Operationalise using AI pipelines (MLOps, AIOps): A microservice performs an atomic function, while containers package code and dependencies for reusable and reliable installation.
  • 2.15 Operationalise using AI pipelines (MLOps, AIOps): Containers support technology-independent hyperautomation by abstracting diverse technologies so they can work together.
  • 2.15 Operationalise using AI pipelines (MLOps, AIOps): Operationalisation includes service definitions, versioning, auditing, retraining, maintenance, and monitoring.
  • 2.15 Operationalise using AI pipelines (MLOps, AIOps): Data pipelines address availability, collection, storage, preprocessing, versioning, and ethics, while AI pipelines address compression, device compatibility, and CI/CD.
  • 2.15 Operationalise using AI pipelines (MLOps, AIOps): Debugging AI involves cross-functional complexity and the possibility of silent errors, while revision time should be considered for major and minor changes.

2.16 Hyperautomation processes and systems

Hyperautomation integrates AI capabilities with automated process or system components and is evaluated through piloting, workflow outcomes, user activity, model condition, and return on investment.

  • 2.16 Hyperautomation processes and systems: Hyperautomation integrates AI capabilities with automated components in processes and systems.
  • 2.16 Hyperautomation processes and systems: A pilot should demonstrate the AI solution to process owners and stakeholders before evaluating workflow outcomes.
  • 2.16 Hyperautomation processes and systems: Hyperautomation outcomes can be assessed for efficiency, effectiveness, and innovation gains.
  • 2.16 Hyperautomation processes and systems: Model drift is addressed through retraining, whereas model staleness is addressed by revising the model.
  • 2.16 Hyperautomation processes and systems: Evaluation includes end-user adoption, questions, usage frequency, documentation, feedback, and returns from costs, revenue, errors, productivity, and turnover.

3 AI Algorithms to Applications

The paper maps AI algorithms to applications through four intermediate capabilities: Prediction, Classification, Association, and Optimisation. This ontology condenses algorithmic choices into a structure for connecting applications, capabilities, and algorithms.

  • 3 AI Algorithms to Applications: Four primary capabilities—Prediction, Classification, Association, and Optimisation—form the intermediate layer between algorithms and applications.
  • 3 AI Algorithms to Applications: Prediction covers regression, classification, time series, and sequence applications.
  • 3 AI Algorithms to Applications: Classification includes detection and covers object, anomaly, and outlier concepts, while Association covers clustering, feature selection, and dimensionality reduction.
  • 3 AI Algorithms to Applications: Optimisation covers scheduling, planning, control, generation, and simulation.

4 AI Applications to Strategy

The organisational context positions the AI Life Cycle within the flow from organisational strategy to decision-making. It condenses the 17 stages into five technical functions aligned with organisational functions.

  • The organisational context places the AI Life Cycle between organisational strategy and decision-making.
  • The 17 stages are condensed into five technical functions aligned with organisational functions.

5 Conclusion

The article presents the CDAC AI Life Cycle as a three-phase, 17-stage framework spanning AI initiatives from conception to production. It also contributes algorithm-to-application mapping and an organisational context connecting AI teams and solutions with stakeholders and strategy.

  • The CDAC AI Life Cycle comprises Design, Develop and Deploy phases with 17 stages spanning conception to production.
  • The article maps AI algorithms to applications through four primary capabilities to support discussions among AI scientists, engineers and stakeholders.
  • Its organisational context integrates AI teams and solutions with senior management and executives working toward organisational strategy.
Loading 2108.13861v1…