Towards Democratizing AI: A Comparative Analysis of AI as a Service Platforms and the Open Space for Machine Learning Approach

Introduction

The democratization of AI is currently a trend in AI research, driven by the shortage of AI experts that hinders the use of AI in many areas, including stock marking trading [1] and personalized medicine [2], to name only a couple. The aim is to reduce the required expertise and make AI more accessible to a wider range of users. Significant progress has been made in the development and training of AI models, enabling the creation and training of models automatically from data, known as low-code AI. Examples of low-code tools include Ludwig [3], fast-ai [4], and Autogluon [5, 6]. Low-code AI tools can be coupled with user-friendly interfaces to transition to no-code AI. These advancements hold the potential to make AI accessible to people with little or no programming experience, allowing them to apply AI to various tasks without the need for technical expertise.

However, providing the infrastructure and setting up these frameworks and tools remains a hurdle that requires expert knowledge. To overcome this obstacle and further promote the democratization of AI, platforms are emerging that take care of these tasks for the user and offer AI as a service. These platforms provide a range of services, such as data storage, model training, and API access, allowing users to easily integrate AI into their workflow without the need for significant technical expertise. With AI-as-a-Service, individuals and organizations can benefit from the power of AI without the need to invest in expensive hardware, software, or personnel. As a result, AI-as-a-Service has the potential to accelerate the adoption of AI and make its benefits more widely available.

In this article, we aim to compare existing AI-as-a-Service platforms and examine why they have not yet led to a breakthrough in the democratization of AI. Based on this analysis, we aim to identify the requirements for a platform that can successfully achieve this goal and then discuss our implementation to meet these requirements. By exploring the strengths and weaknesses of existing AI-as-a-Service platforms, we hope to provide insights into the design and development of a platform that can truly democratize AI and make its benefits accessible to everyone. Through this analysis, we hope to contribute to the ongoing efforts to democratize AI and make its benefits available to a wider audience.

Study of Existing

Platforms

The availability of easy-to-use AutoML solutions as a platform is a crucial factor for the participation of non-AI experts in the technology of machine learning. We will now take a look at the various existing platforms for AutoML.

AutoML Platforms by major Cloud

Providers

The major cloud providers - Amazon SageMaker[https://aws.amazon.com/sagemaker/ ], Google Vertex AI[https://cloud.google.com/vertex-ai ], and Microsoft Azure Automated ML[https://azure.microsoft.com/en-us/products/machine-learning/automatedml/ ] - have invested significantly in AutoML and have developed powerful and scalable solutions. These machine learning platforms leverage cutting-edge algorithms and state-of-the-art technologies. Additionally, these platforms offer a high level of service, from prototyping to production, and provide a no-code environment that enables developers to easily build and deploy machine learning models.

One major disadvantage of these platform solutions is their lack of openness. Since these platforms are not open source, expert users are limited in their ability to customize and modify algorithms to fit their specific needs.

Additionally, these platforms are not self-hosted, meaning that sensitive data must be entrusted to a third-party provider.

Furthermore, the algorithms used by these platforms are often black boxes, making it difficult to understand how they arrived at their conclusions. This lack of transparency can be problematic, particularly when it comes to ethical considerations and ensuring that the algorithms are not biased or discriminatory.

Additionally, customers are often locked into using a single vendor, limiting their ability to switch providers or take their business elsewhere. Finally, while these platforms offer a no-code environment, they are often complex and difficult to navigate without a solid understanding of machine learning concepts and terminology. This can be a major barrier for non-experts trying to leverage machine learning in their work.

Standalone AutoML

Platforms

An established platform that has made it its mission to overcome this obstacle is DataRobot[https://www.datarobot.com/ ]. The platform stands out in terms of user-friendliness. Its no-code environment makes it easy for non-experts to build and deploy machine learning models, and its scalable AutoML solution covers the entire machine learning life cycle. However, like the major cloud providers, DataRobot’s platform is not open source, meaning that experts are limited in their ability to customize algorithms. Additionally, the platform has no option to self-host, and the algorithms used are often black boxes, limiting transparency and posing ethical concerns. Finally, users are locked into the DataRobot ecosystem, which can be a major drawback.

On the other hand, Predibase[https://predibase.com/ ] attempts to achieve complete transparency using a declarative machine learning approach [7] and utilizing open-source components under the hood. However, minimal programming knowledge is required for the low-code platform. In addition to this, is not foreseeable that the platform will be made available as open source. Thus, it cannot be self-hosted for sensitive data.

For users looking for a more open approach, H2O[https://h2o.ai/ ] [8] is a powerful option. As an AutoML platform, H2O provides a scalable solution for the complete machine learning lifecycle that can be used either as a PaaS or self-hosted. However, H2O’s platform does have some limitations. Its low-code environment may not be accessible to users with no programming skills, limiting the democratization of machine learning. H2O knows about this drawback and introduced a commercial no-code offering, H2O Driverless AI Wizard, which leverages the full potential of AutoML for users with no programming. But this solution is not open-source and cannot be self-hosted.

Requirements

So far, we have gained insights into the various platforms that offer AI as a service, and we have concluded that they are not yet sufficient for achieving true AI democratization. Therefore, our next step is to derive the requirements for such a platform that can lead us towards the democratization of AI. These requirements will be crucial for creating a platform that can enable people from diverse backgrounds to access and benefit from AI technologies. We will focus on identifying the key features and functionalities that are needed to build a platform that can support a wide range of users, including those with limited technical expertise.

Platform

But first, we want to emphasize why such a platform that offers AI as a service is a requirement for achieving true AI democratization. One of the main reasons why such a platform is crucial is that it allows users to start using the software without the need for a complex setup process. Especially with AI applications, there may be a significant amount of expert knowledge required, as certain parts of the application need to be executed on specialized hardware, such as GPUs, for effective performance. By offering AI as a service, users can access the power of AI without needing to invest in expensive hardware or hire specialized personnel to configure and operate the system.

Easy-to-use

User-Interface

In order to enable users with limited technical expertise to use the platform, it is important to ensure that the user experience is optimized. The platform should have an intuitive user interface that is easy to navigate. With just a few interactions, users should be able to train an AI model based on their own data without needing any expert knowledge of AI or complex technical terms. By abstracting away the complexities of AI, the platform can become more accessible to a wider range of users, regardless of their level of technical expertise.

Scalability

As there are many areas where the potential of AI has not yet been fully exploited, a platform offering AI as a service could have a vast user base, making scalability essential. With the increasing amount of data being generated and stored, the platform must also be capable of handling large amounts of data. Additionally, the complexity of AI models is increasing, which requires more computing power for training, so the platform must be able to handle increasingly complex AI algorithms. In summary, the platform should be scalable in terms of its user base, data handling capacity, and ability to handle complex AI models.

Adaptability and

Extensibility

Beyond the requirement for scalability in handling more complex AI algorithms, there is another important aspect to consider: the platform’s adaptability and extensibility. This is crucial because the field of AI is constantly evolving, with ongoing research and new advancements being made. If the platform is designed in a way that enables easy integration of these new developments, it can benefit a wider range of users. By allowing for customization and flexibility in the development of new AI models and algorithms, the platform can continue to evolve alongside the latest advancements in AI, ensuring that users have access to the most up-to-date and relevant tools and technologies. It is important to note that this adaptability and extensibility need not necessarily be achieved through the user interface alone. As AI experts are required to implement and integrate these changes, they can be made directly at the code level.

Openness

Having open-source code is therefore a requirement for the platform so that these changes can be made by as many AI experts as possible. Generally, public access to the platform’s code offers further advantages. Firstly, it increases trust in the platform, as independent experts can verify the internal workings of the platform. Secondly, the platform also benefits from an open-source community around it. This way, it can be developed faster, more efficiently, and cost-effectively, and users can help each other with questions and problems.

Self-Hosting

Furthermore, in addition to the publicly available instance of the platform, it may be important for some users to be able to deploy their own instance. This allows for greater security considerations, as access can be heavily restricted. So, the platform can also be used for data that requires a higher level of security, expanding the use of AI in such areas. This is also another argument for making the code of the platform open-source, as it enables the tracking of what precisely occurs with the uploaded data.

Technologies

Now that we have identified the necessary requirements for a platform that offers AI as a service, we will take a look at the technologies that we will use to implement our own platform.

Kubernetes

In order to address the significant scaling requirements of our platform, we will be utilizing the open-source container orchestration framework Kubernetes [9]. By deploying Kubernetes in the cloud, we can take advantage of the auto-scaling mechanisms provided by cloud providers.

Kubernetes provides a high level of abstraction by organizing available computers as nodes and grouping them into a Kubernetes cluster. The platform’s Control Plane API enables users to manage the state of their containers, while the Control Plane itself distributes the workload in the form of pods to the nodes. Pods are the smallest unit in Kubernetes and consist of one or more containers that run on the nodes [10]. This approach allows us to efficiently manage the scaling of our platform and ensure optimal resource utilization.

Service Mesh

A service mesh is a software infrastructure that enables managing and operating microservices-based applications. It provides a range of benefits, such as improved scalability, reliability, and security. By using a service mesh, application developers and operators can reduce the complexity of their architecture while improving flexibility and agility. A service mesh allows developers to focus on application development without having to worry about the underlying infrastructure, while providing operators with better control over application traffic and security.

One of the main advantages of using a service mesh is its ability to provide advanced traffic management capabilities. With a service mesh, operators can manage and control the flow of traffic between services, allowing them to balance the load and ensure optimal performance. Additionally, a service mesh can also provide improved security features such as authentication and authorization, allowing only authorized traffic to pass through.

Istio [11] is a popular open-source service mesh that provides a variety of features such as load balancing, traffic management, security features like authentication and authorization, as well as troubleshooting and monitoring. By using Istio, developers and operators can reduce the complexity of microservices-based applications by managing traffic between services within the application. Overall, Istio is a powerful and flexible service mesh that helps developers and operators operate microservices-based applications more effectively and securely.

Kubeflow Pipelines

While simple microservices can be deployed as native Kubernetes deployments, our machine learning tasks require special treatment due to their need for significant computing power and specialized hardware. To address this, we will be using a workflow engine, specifically Kubeflow Pipelines [12], which is built on top of Argo Workflow [13] - a workflow engine designed specifically for Kubernetes environments. With Argo Workflow, users can easily create, manage, and run complex workflows, including creating workflow templates and executing them with different parameters.

Kubeflow Pipelines provides additional ML functionality such as an artifact store for data and models, visualizations for executions and metrics, and robust features that increase the reliability of our workflows, such as automatic retries and error handling. By using Kubeflow, we can ensure that our ML workflows can handle errors and unexpected events gracefully and achieve greater efficiency and reliability in our development process.

Kubeflow Pipelines is a component of the Kubeflow project, which is designed to facilitate the development, deployment, and management of machine learning models on Kubernetes. In addition to the pipelines component, it includes support for Jupyter Notebooks to explore data and an interface to TensorFlow to train and serve models. Kubeflow also offers model versioning and monitoring, enabling users to keep track of their models’ performance over time.

Ludwig

For our No-Code AI component, we will make use of Ludwig [3], a declarative ML framework [7] that creates and trains ML models using a configuration file as input. This configuration file specifies the input and target data types, as well as other parameters needed for training, such as the optimizer, the loss function, and the number of epochs. Ludwig’s declarative approach eliminates the need to write low-level code, making it ideal for non-technical users who want to build and train ML models easily.

This can best be demonstrated through an example, using the PetFinder dataset [14]. This dataset contains information about animals in a shelter, such as their name, age, or an image of the animal. The goal is to predict how quickly these animals will be adopted, which is represented as categorical values in the dataset. An AdoptionSpeed of 4, for example, indicates that the animal has not been adopted within 100 days. An example is shown in the following figure fig-pet, and a Ludwig configuration for this task is shown in the next figure . The configuration file defines the input and output columns and sets some additional training parameters.

By providing intuitive selection options in a user interface, we eliminate the need for manual writing of the configuration file. Additionally, we remove additional training parameters from the configuration process in favor of employing techniques of hyperparameter optimization. This approach expands the search space for AI models, reducing the potential for bias caused by an expert’s experience. Using a user interface, we can thus expand this low-code ML tool into a no-code tool. This enables us to allow users with no technical experience to interact with our platform.

Architecture

We will now explain how we combine various technologies to create our AI-as-a-Service platform called Open Space for Machine Learning (Os4ML). Figure fig-architecture illustrates the architecture of our platform.

Our platform is designed using microservices as its architectural pattern. We have structured these microservices into a mesh topology that is accessible through an API gateway, which we have implemented using Istio. This architecture provides a flexible and scalable platform that enables developers to access different microservices conveniently. The frontend and other components can easily access the microservices through the API gateway, which simplifies development and maintenance tasks.

One of the most crucial microservices in our platform is the Os4ml-Service. It is responsible for managing our domain models and storing critical information such as metrics and execution times. This service plays a vital role in providing efficient and reliable machine learning solutions.

In addition, we have implemented the objectstore service, which offers a dependable way to store and retrieve models and data. Furthermore, we have the workflow service, which handles communication with the workflow engine. For our platform, we use Kubeflow Pipelines as our workflow engine. Both the workflow engine and the objectstore service can share the same objectstore, which ensures consistency and reliability in our platform.

Our frontend allows the user to upload data. In the background, the data is transformed into an interface that is compatible with Ludwig (e. g., into Pandas dataframes), with data types detected and saved. We provide templates for the workflow engine to create a Ludwig configuration and train a machine learning model using the dataframe. These templates can be quickly and easily customized to create tailored machine learning models. Our platform thus provides a user-friendly and effective way to train and deploy machine learning models.

To ensure scalability, we deploy the entire system on Kubernetes, which provides a robust and reliable platform for container orchestration. With Kubernetes, our system can handle large volumes of data and easily adapt to changes in demand. By breaking down the system into smaller, independent components, we can maintain better control and reduce the risk of system-wide failures. This architecture enables a flexible and scalable system that can efficiently handle complex machine learning workloads.

Deployment

The requirements section has highlighted the need to allow the advanced users to deploy the platform themselves. To achieve this, we have adopted a three-part Infrastructure-as-Code approach. In general, Infrastructure-as-Code involves provisioning the necessary infrastructure using code so that it can be reproduced or restored in a predictable manner. To implement this, we use the open-source tool Terraform [15]. Our deployment process is divided into three steps: cloud setup, Kubernetes cluster setup, and the actual deployment of our platform. If steps 1 and/or 2 are not required because they already exist, the user can skip directly to step 3.

We will be publishing the Infrastructure-as-Code scripts along with our application code. This means that deploying the platform will only require a single command, in our case, ’terraform apply’. This will greatly reduce the time and effort required for the deployment process, making it more efficient and user-friendly. Additionally, this approach ensures consistency in the infrastructure setup, reducing the possibility of errors caused by manual configuration.

In order to realize step 3 of our deployment process, we utilize the declarative GitOps continuous delivery tool ArgoCD [16]. This provides the administrators of individual instances with access to updates to the platform as well as a pre-packaged monitoring solution. By utilizing ArgoCD, we can automate the deployment process, allowing for seamless updates and streamlined management of our platform. Additionally, the use of declarative configurations ensures that our infrastructure remains consistent across all environments, reducing the risk of configuration drift and ensuring high reliability.

accomplishments-and-contributions

Accomplishments and

Contributions

We are thrilled to report significant progress made on our AI-as-a-Service platform, “Open Space for Machine Learning”[https://www.os4ml.com/ ], and are eager to share the goals we’ve achieved and contributions we’ve made.

Transforming Ludwig from Low-Code to

No-Code

Our most noteworthy accomplishment is the creation of a user-friendly frontend that has transformed Ludwig, our low-code ML tool, into a no-code tool. With our efforts, we have opened up the platform to a broader range of users, including those with little to no experience in machine learning, who can now effortlessly train and deploy AI models without the need to write any code. Our innovation has revolutionized the industry, making machine learning accessible to anyone.

Integrating Multiple Open-Source

Tools

Another proud achievement is our integration of several powerful open-source tools, such as Istio, Kubernetes, and Kubeflow, into our platform, creating a blueprint for highly scalable AI applications. This has opened up new avenues for exploring the potential of AI.

Cloud-Enabled Platform with Self-Hosting

Option

Furthermore, we have made our platform cloud-agnostic and easy to deploy for the community. This approach allows us to take advantage of the benefits of cloud computing, such as scalability and cost-effectiveness, while still providing the option for users to self-host our platform on-premises. By enabling easy deployment, we have made it simpler for the community to take advantage of our AI-as-a-Service platform, including those dealing with sensitive data.

Overall, we are proud of our contributions to the AI community and look forward to continuing to push the boundaries of what is possible in the world of machine learning.

Summary & Outlook

In this article, we have demonstrated that the next step in AI democratization is the provision of AI as a service through a platform. However, current solutions have not been satisfactory, leading us to collect requirements for such a platform. We have shown how we use existing open-source solutions and add our own components to meet the collected requirements. Our platform is also available as open source[https://github.com/WOGRA-AG/Os4ML ], enabling further collaboration and development towards more accessible AI solutions. Now that the foundational work is complete, the next step is to enhance the AI component. One area of improvement would be to incorporate an explainable AI mechanism. This would provide valuable insights into the trained model’s decision-making process, thereby increasing trust and transparency.

Acknowledgments

Os4ML is a project of the WOGRA AG research group in cooperation with the German Aerospace Center and is funded by the Ministry of Economic Affairs, Regional Development, and Energy as part of the High Tech Agenda of the Free State of Bavaria.

References

enumi.

• Jan Juricek. 2014. The Use of Artificial Intelligence in Building Automated Trading Systems. International Journal of Computer Theory and Engineering 6, (January 2014), 326–329.DOI: https://doi.org/10.7763/IJCTE.2014.V6.883

• Jamilu Awwalu, Ali Garba, Anahita Ghazvini, and Rose Atuah. 2015. Artificial Intelligence in Personalized Medicine Application of AI Algorithms in Solving Personalized Medicine Problems. International Journal of Computer Theory and Engineering 7, (December 2015), 439–443. DOI:https://doi.org/10.7763/IJCTE.2015.V7.999

• Piero Molino, Yaroslav Dudin, and Sai Sumanth Miryala. 2019. Ludwig: a type-based declarative deep learning toolbox .

• Jeremy Howard and Sylvain Gugger. 2020. Fastai: A Layered API for Deep Learning. Information 11, 2 (2020). DOI:https://doi.org/10.3390/info11020108

• Nick Erickson, Jonas Mueller, Alexander Shirkov, Hang Zhang, Pedro Larroy, Mu Li, and Alexander Smola. 2020. AutoGluon-Tabular: Robust and Accurate AutoML for Structured Data.

• Xingjian Shi, Jonas Mueller, Nick Erickson, Mu Li, and Alexander J. Smola. 2021. Benchmarking Multimodal AutoML for Tabular Data with Text Fields .

• Piero Molino and Christopher Ré. 2021. Declarative Machine Learning Systems .

• H2O.ai. Retrieved March 29, 2023 from https://github.com/h2oai/h2o-3

• Kubernetes Components. Retrieved March 29, 2023 from https://kubernetes.io/docs/concepts/overview/components/

• Shazibul Islam Shamim, Jonathan Alexander Gibson, Patrick Morrison, and Akond Rahman. 2022. Benefits, Challenges, and Research Topics: A Multi-vocal Literature Review of Kubernetes .

• The Istio service mesh. Retrieved March 29, 2023 from https://istio.io/latest/about/service-mesh/

• Kubeflow Pipelines. Retrieved March 29, 2023 from https://www.kubeflow.org/docs/components/pipelines/

• Argo Workflows. Retrieved March 29, 2023 from https://argoproj.github.io/argo-workflows/

• Sherine Zhang and K. X. Zhang. 2019. PetFinder Challenge: Predicting Pet Adoption Speed.

• Terraform. Retrieved March 29, 2023 from https://developer.hashicorp.com/terraform/docs

• ArgoCD. Retrieved March 29, 2023 from https://argo-cd.readthedocs.io/en/stable/

Attribution

arXiv:2311.04518v1 [cs.LG]
License: cc-by-4.0