Artificial Intelligence (AI) is a hot IT topic today, producing new and exciting results in academic research areas and across real-world deployments. So how do enterprises create value by implementing AI solutions? Many organizations are struggling to take the first step because the journey seems long and complicated. It doesn’t have to be. With some forethought, planning and good practices, you can implement a solution that will help you stay ahead of your competition without draining your expense budget.
Building production-ready AI models and solutions brings unique challenges during the development and operation lifecycles. The solutions need to run effectively and predictably for a long period of time, with simplified update procedures when new data becomes available. Additionally, they need to play well with your software stacks of other applications.
There are many approaches to incorporating AI technology in your business, from choosing automation tools built on AI technology to building a Data Science practice within your organization, and any mix in-between.
By incorporating customized AI solutions, leveraging Machine Learning (ML) and Deep Learning (DL) models in your applications and business processes, you can improve efficiencies and data driven decisions within your organization. We’ll take you through how many organizations today are getting started with AI by combing software development processes with ML and DL models in a consistent application platform.
Like any other business endeavor, starting with a solid plan is key to success. To help you start building a plan, we’ll take a look at a typical AI lifecycle from managing the datasets and the multiple stages of training, to deployment and integration into the application environment.
Before You Get Started
Here are some key questions to ask yourself:
- Have we implemented a big data solution in the past or do I have access to large amounts of structured or unstructured data?
- What skill sets do we have as it relates to AI? Take inventory of your internal skill sets and consider the scope of what you can manage with internal talent. You may consider contracting some technical skills to business partners or bringing on additional internal talent with the appropriate skills needed to be successful.
- Does your use case involve a specific business unit? If so, you will need subject matter experts (SMEs) from the business unit committed to the initiative, as well as AI experts to architect and develop your AI solutions.
Need to know more about what skills you may need? A good AI partner should be able to provide guidance based on the scope of your projects.
- Do we have budget and buy-in? Every worthwhile investment will have some initial cost. Ensure the projected cost fits within your available budget. Having executive and stakeholder buy-in is key for any initiative to be successful, and AI is no different. If you're working on an AI project that involves a specific business unit, you will need buy-in from all the stakeholders from the business unit as well.
Typical Stages in the AI Journey
Choosing the Right Use Case
Choosing the right use case for an AI application is critical to quickly prove the value to your stakeholders. It's important to choose a Proof-of-Concept (PoC) or project that can rapidly deliver business value and has few roadblocks to implementation. The three determining factors used to evaluate use cases for fastest business impact are:
Data Availability. Any AI application will only be as good as data used for training it. In the enterprise environment, data can be more important than the algorithms. Gathering and organizing a new collection of data can be a long and labor-intensive task. A project with readily available data, or easily collectable from existing processes, can save several months in the data preparation phase.
Clean Data. After “Data Scientist”, “Clean Data” may be the second most common term that AI has added to the IT vernacular. Clean data refers to the accuracy and completeness of the dataset you’re working with. Imagine a use case in which you’re using customer data to predict which customers will be interested in a specific product. For an accurate prediction, it will require that all of your customer information is complete and up-to-date. By spot checking a few accounts in your Customer Relationship Management (CRM) database, you may start to get a sense of the cleanliness of your dataset and, for most organizations, a better sense of the health of your data. Dirty data simply means that the data is too unstructured, contains too many errors, or is incomplete and will not produce accurate results when used in model training. “Data Wrangling” is the term used for the process of analyzing and evaluating which data stays, which data gets cleaned up, and which data gets tossed out .
Existing Technology. In order to deliver value in a short period of time, try to leverage already tested approaches that run well on your current technology. In many cases, you may have existing infrastructure to leverage for your PoC, but if you’re lacking the infrastructure to get started, there are plenty of options for you. Your hardware vendor may even provide access to infrastructure at no cost for an initial PoC. Everybody wants to work with cutting-edge technology and use the latest and greatest solutions, but wait for a later iteration when you have more resources and buy-in. You’ll save time and effort that can be used elsewhere.
What criteria should be used to select an initial AI PoC?
- Time Saving Potential: The best candidate is a simple task that is executed multiple times across the organization. For example, answering simple questions in a call center, filling expense reports for internal employees, etc....
- Data Intensive Tasks: Look for tasks which rely more on data versus experience or intuition. AI can better execute tasks which have clearly specified data inputs.
- Monotonous Tasks: There are tasks which are not possible due to costs or diminishing effectiveness of people doing it. A common example is a worker visually inspecting every part entering a factory for flaws. AI will never grow tired of monotonous tasks and will deliver consistent results, improving the overall quality and consistency of the end product.
- Scalability: The AI application is usually high initial investment, with low operational cost. The best task to start with is the one which can be implemented once, and run multiple times to spread the initial investment across multiple tasks or departments.
The right use case is critical to the success of the PoC. Remember to pick one that will deliver results quickly. In order to do that, the project must have readily available, clean data that requires little data wrangling and can run on your existing infrastructure. Try to target tasks that are repetitive, require only data to accomplish, not experience, and are difficult for people to do for any length of time. If the task is repeated across multiple departments, spreading the initial cost becomes easier.
Collect Your Data
Many AI algorithms being developed today are based on supervised learning and require a significant amount of data wrangling and pre-processing to ensure model accuracy. In supervised learning , the model is trained with labeled data, mapping a specific input with the expected output. For example, if you take a dataset of images of various vehicles, label all car images in your training, and, with proper tuning and enough data, you should be able to predict whether a new image is of a car or not. This of course requires clean and properly labeled data, which can be the most time-consuming stages in developing an AI model. For unsupervised learning, we do not have the data labeled for an expected outcome and the model is tasked to find patterns in the data.
Once you have selected your use case and collected your data, you will need to organize, clean, and process your data. For supervised learning models, your training data will need to be labeled, which can be a time consuming task for your data wranglers
Proving Your Use Case (Pilot)
Next, you will work through model development where your data scientist will build and test multiple models to identify the appropriate data science methodology that provides the highest level of accuracy. Once your data scientist has identified the appropriate model, they will perform hyper-parameter tuning to achieve the highest level of accuracy possible.
There are a broad range of tools supporting this process, spanning from open-source machine and deep learning frameworks like Spark* ML, sk-learn*, Keras*, TensorFlow*, CNTK*, or BigDL to specialized training platforms like LICO* from Lenovo.
Typically, it’s best to start with small chunks of data for initial model development. Once the accuracy of the model reaches an acceptable level, you can increase the amount of data you're working with which will, in turn, increase the accuracy of the model. Of course, these larger iterations will require greater processing power, storage, and run times.
Implementing AI into a production environment can be as challenging as the development process. AI inference may require changing business processes or modifying existing applications. The scope of operationalizing an AI model can be affected by the complexity of the use-case.
To implement the inference of your AI model, it could be as simple as incorporating it into your BI tool for predictive or prescriptive use cases, or as complex as modifying your DevOps process to incorporate it into your business applications. With your first PoC, start small with a light application or plugin to an existing tool.
The work is not over after the solution is implemented and running in a production environment. To improve the accuracy of your AI application, you may need to re-train your model. AI applications generate new data which can be used for additional training to improve accuracy. Also, changes in your environment (new customers, new products, etc...) may require modifications of the model to ensure accuracy.
In an ideal lifecycle, a fully automated pipeline, from data generation to deployment, is ideal, but may not be feasible. It's important to find the right level of automation in your AI deployment that ensures the best accuracy for your model, without being cost prohibitive. This may require changes in your DevOps process to ensure you're getting the best value from your AI models.
Is it Time to Get Started with AI?
Once you have completed a successful PoC, it's time to get serious about your AI strategy. Keep in mind, your competitors are looking at AI technology today to increase efficiency, reduce cost, and, ultimately, become more competitive in the marketplace.
Things to Consider
Consider working with a technology partner that can provide end-to-end solutions from the edge to the data center, providing hardware, software and (most importantly) expertise to help you get started.
At Lenovo, we're taking a prescriptive approach to AI and working with clients to build solutions that address their unique business goals. We have expertise in Big Data, AI, and IoT, realizing the dependencies of AI technology within the data center. We provide access to AI Innovation Centers for our clients to experience the power of AI and build initial PoCs prior to investing in new infrastructure. We have made significant investments in developing new software platforms to accelerate AI development and provide AI consultants to assist customers along the journey. By partnering with leading technology providers, such as Intel and select business partners that have deep industry and AI expertise, we're providing holistic solutions for our clients. Learn more at lenovo.com/AI.
Intel technologies’ features and benefits depend on system configuration and may require enabled hardware, software or service activation. Performance varies depending on system configuration. No computer system can be absolutely secure. Check with your system manufacturer or retailer.