hello everyone. I'm Kumar Madla. Today I'm going to present what are the advantages of auto ML and how auto ML is transforming. Predictive analytics with 60 to 80% of foster development. I'm currently working as a senior data scientist at PepsiCo, and I had ma I did my masters in artificial intelligence from University of EBA back in 20 at 13. I had a total of 11 plus years of experience working in data science, machine learning, and artificial intelligence. And I have implemented projects in multiple industries ranging from retail, healthcare, manufacturing, banking, and finance sectors. So based on all my experience and my expertise, I'm going to present this today. So my topic title is Democratizing Enterprise ai, how Auto ML is Transforming Predictive Analytics with 60 to 80% faster deployment Development. Before understanding this Auto ml in olden days we were using different machine learning models and different hyper parameter configurations, and it is taking more time whenever you take any problem and we have to, if you have to develop a predictive analytics solution being it classification, recreation or clustering we have multiple machine learning models and each algorithm has. And number of hyper parameters. And each hyper parameter takes many different values. Building multiple models and tuning those hyper parameters, multiple hyper parameters in the, coming up with the best model takes a lot of time. But with the advent of these. Auto ML techniques there is a paradigm shift in the way that machine learning engineers or data scientists are developing these machine learning models for different problems that they're working on. So these are the advantages, main advantages of auto ML techniques. So this automobile techniques encompasses and the tools that automate traditionally. Manual expertise dependent processes from data preprocessing and feature engineering tool algorithm selection and hyper parameter optimization. And another advantage of using these autos the bridge, the talent gap if you how to build a different machine learning models, if without having these auto ml, you need to hire a machine learning engineer or a data scientist to build. These different models and to come up with a solution. But by using these auto techniques, you can effectively bridge the talent gap that has constrained many organizations, analytical capabilities, enabling business analyst and domain experts to develop predictive models without extensive program expertise. So these are the evolving capabilities. As I said early systems focused primarily on AMIC selection and hyper parameter. And the recent developments in the ml focuses not only on algorithmic selection and the hyper parameter tuning, they also address the entire machine learning pipeline, including deployment and monitoring. So these are the theoretical foundations of, as we initially discussed in the previous slide, these are the common. Or basic foundations of auto. So one is the, so this optimization theory employs various strategies to navigate where as they discussed in the beginning of the session. So these machine learning models have different hyper parameter configurations and each configuration, each hyper parameter has multiple values. So this is related to this optimization in theory. So what algorithm to use and what hyper parameter tool. Tune and what are the best values for those hyper parameters comes under this optimization theory and the second foundation for auto ML is the meta learning. So this meta learning includes, I. How to learn across data sets and modeling tasks modeling tasks here, it can be regression classification or clustering. And how do we train multiple models to come up with the best model to use it in the real time and the comp. The next third foundation for this art M is the computational efficiency, balancing exploration and exploitation during the model training. Comes under this computational a, c, and C. So we should not focus only on the exploration or only on the exploitation. So we have to do a balance between the exploration and exploitation while training these machine learning models. So these automal systems leverage these foundations to efficiently identify promising solutions while transferring knowledge between the problems. The employee techniques such as base optimization, genetic algorithms, and gradient based approaches to navigate the complex space of possible model configurations significantly, and also the significantly improve the efficiency and performance of traditional systems. So this is how the auto ML frameworks have been about. So it initially started with auto which was introduced in 2013. So this auto. Pioneered combined algorithmic selection and hyper parameter optimization, and it established the foundation of subsequent auto ml frameworks. This is the first in the auto ml frameworks or auto ML techniques which was introduced by automaker and then it followed by autos, KN. So this autos, KN not only involves hyper parameter optimization and algorithmic selection. It also introduced the transfer learning concept. So what does it mean by transfer? Learning is so the, any model or algorithm that was trained on one task can be used to retrain on another task. Maybe we can take an example of any deep planning models or image recognition models. So the models that were trained to. Identify cats and dogs can be used to retrain from last couple of layers to identify other things other animals like horses of, or any other animals. So this comes under transfer learning and nowadays there are more advanced auto ml. Frameworks are coming up into the market. They not only focus on hyper parameter optimization and algorithmic selection, they also focus on data preprocessing and other data preparation and other techniques. Some of the notable players in this end-to-end platforms are Google Cloud, R ml, and H2O is driverless a and they extend automation across the entire ML lifecycle. This is an extension to the previous slide. These are the prominent auto ML framework analysis. Now I'm comparing the differences between Autor, IBM's Auto, aa, and Microsoft. And I, as we discussed it in the previous slide auto was introduced in 2013, and it focuses more mostly on the algorithmic selection and hyper. Tuning. And the technique is called sequential model based algorithmic configuration, and it uses patient optimization to navigate these complex set spaces. And in 2023 KDD has awarded auto. As a test time award in 2013. 20, sorry. 2023. IBM's next one is IBM's Auto A. It extends traditional capabilities by addressing the entire machine learning lifecycle within enterprise context. It also includes automated data preparation, feature engineering, model selection, and high parameter optimization. The main difference with respect to IBM's Auto AAEs, it has a strong integration with IBM's Watson Ecosystem and it enterprise focused features like built in fairness checks and explainability tools. The third one. In this comparison is a Microsoft's NNI. It takes a specialized approach focusing on neural network optimization and automated deep learning. It provides infrastructure and a neural architecture search, hyper parameter tuning and model compression through a highly modular framework. It also supports deep learning frameworks such as tensor flow by, sorry. It's such as the tensor flow. By touch, et cetera. It offers flexible deployment options across diverse computing environments. This is the framework comparison, strengths and limitations. These are the different strengths and limitations of the three frameworks that we have discussed. Auto, IBM's Auto, a I, and Microsoft. And I let us first start with the auto. So the primary strength of Otca is in pioneering. Cash solution. And it also integrates with established ML toolkit and it has a strong academic foundation. And the key limitations with AKA is it has a limited end-to-end automation. It has less support for deep learning, minimal enterprise integration features. And the best enterprise use cases for using auto is academic research. For developing proof of concept projects, and it'll be suitable for organizations with existing WCA investments. The next one is auto EI which is from IBM. The primary sense of auto from IBM is its end to end ML lifecycle Automation, built in fairness and explainability tools strong, and it has a strong enterprise integration. The key limitations, it can be limitation for some of the users. We will have a vendor lock in with IBM, and it's a proprietary ecosystem. It may be expensive for some clients, and it is higher implementation complexity. The complexity of implementing IBM's auto, difficult and the best enterprise use case for IBM's Auto AEs regulated industries like finance and healthcare organizations since existing IBM infrastructure application requiring explainability. The third one in this comparison is and I from Microsoft. The primary strength of NI is it's deep learning specialization, multi framework support, flexible deployment options, and the key limitations. It has a steep learning curve, less focus on classical ML algorithms and requires more technical expertise. Best enterprise use cases are day planning applications, optimizations with teams with strong technical capabilities. These are the enterprise integration considerations. So IT infrastructure integration organizations must evaluate compatibility with current data, current storage solutions, processing frameworks, and more and model deployment infrastructure. A PA based integration approaches have emerge as preferred. Allowing auto ML systems to interact with existing data pipelines while minimizing disruption. And these are the deployment options. Cloud-based deployments offer scalability, but may introduce data movement challenges. So when you have to implement a solution in the cloud-based infrastructure, you may have to move your data from in-house systems to cloud-based systems, which comes with some cost and time. On-premises implementations provide greater control, but require significant computational resources and the middle ground like hybrid approach. Leveraging containerization and orchestration technologies present a promising middle ground data governance requirements. Robust mechanisms for data quality assessment, version control. Linear tracking and access control are essential. Auto assistance typically require substantial volumes of training data. Amplifying the importance of governance frameworks that ensure data accuracy and appropriate use. Organizational adoption for auto MR. Traditional boundaries between engineering, data science and ation analysis roles often blur as auto ML democratizes. Model development capabilities and these auto ML establishes cross-functional governance. Successful implementations typically establish cross-functional teams responsible for model governance with clearly defined hands of between automated and human LED processes, implement change management address potential. Resistance from technical specialists while encouraging responsible adoption by business users through targeted training and communication. Create tiered approaches. Many organizations implement tiered approaches where straightforward predictive tasks leverage auto ml, while complex mission critical applications maintain greater human oversight. These are the enterprise benefits. Of using auto, 60 to 80% of the development time will be reduced by using these auto ML techniques or strategies, and there will be a return on investment within 12 to 18 months of adopting these auto ML strategies. And there will be 30% reduction in downtime when using these optimal strategies or techniques. Democratization of predict two analytics. The first one is expanded access. Auto ML enables business analysts, domain experts, and non-specialists to develop sophisticated predictive models with less machine learning knowledge. Breaking down silos allows those closest to business problems to direct end engine and solution development. Citizen data science. Many enterprises establish citizen data scientist programs with auto ML as the technological foundation. Multiply the capabilities effectively multiplies analytical capabilities without proportional increases in specialized hiring. These are the implementation challenges, interpretability concerns. Automated approaches often generate complex models resistant to straightforward human understanding. Data quality dependencies systems remain foundationally, fundamentally dependent on data quality, potentially amplifying underlying issues, domain specific requirements. Generic solutions often struggle with highly specialized domain problems requiring industry specific approaches. Technical debt, cremation, unmanaged adoption can accelerate technical debt. Through model proliferation without adequate governance. These are some of the strategies to overcome those challenges. Implement explainable a address data quality, enable domain customization. So the first one is implement explainable ai. So integrated explainable AI techniques alongside auto ml establish interpretability thresholds for different application categories and. Maintain simpler model adoption for high transparency requirements. Address data quality, implement systematic data quality assessment prior to auto ML deployment. Automated data quality monitoring clear processes for handling quality exceptions and realistic expectation setting. Establish data quality thresholds that must be satisfied before automated modeling. Proceeds enable domain customization. Develop customized autoAML frameworks that incorporate domain knowledge via specialized pre-processing pipelines, custom algorithm implementations, domain specific object, two functions and expert guided constraints on model exploration. These are the future directions. For using auto explainable a integration, next generation auto ML frameworks are incorporating explainability by design rather than as in afterthought. This includes optimization objectives that balance predictable performance with interpretability metrics and automatic generation of explanation artifacts, edge computing applications, auto resistance, specifically designed to optimize models for edge development. Considering resource constraints, power limitations, and specialized hardware accelerators, these systems automatically balance model complexity against inference speed and memory footprint. Federated approaches converging with auto ml to address data privacy and sovereignty challenges, enabling prediction model development across decentralized data sources. Without requiring data centralization, regulatory aware, evolving regulatory frameworks are driving innovation in automated compliance verification and documentation generation with systems incorporating regulatory awareness directly into optimization. Objectives. Enterprise implementation roadmap. Assessment and planning. Identify high value use cases, evaluate data readiness, select appropriate framework and define governance structure. Success metrics include prioritize, use case portfolio, and. Established success criteria. Common challenges include unrealistic expectations and insufficient data quality. Assessment, initial implementation, deploy optimal for targeted use cases, established and monitoring and evaluation. Train initial user cohort, and document early learnings. Success metrics include model performance metrics and development time reduction. Challenges include. Technical integration issues and resistance from data scientists, scaling and optimization, expand use case governance, enhance governance frameworks, establish model arts practices, and integrate with business processes. Success metrics include enterprise web model inventory, and the business impact measurements. Challenges include model proliferation and technical data accumulation and advance integration. Incorporate emerging auto ML capabilities. Integrate with. Complementary AI systems develop custom domain adoptions and establish continuous improvement cycles. Success metrics include hybrid, a, system effectiveness and competitive differentiation metrics. Strategic recommendations For enterprise adoption, we have to first start with phased implementation and then proceed balance to governance model, strategic workforce integration with use case boundaries. Organizations should adopt phased implementation. Approaches that balance ambition with paradigm initial deployment should target well-defined use cases with clear success criteria and model complexity. Moderate complexity. Effective governance structures typically combine centralized oversight with distributed execution capabilities. Thank you.