AI Frameworks

There are a myriad of AI frameworks, which provide end-to-end systems to build models, train them, and deploy them. By far the most popular is TensorFlow, which is backed by Google. The company started development of this framework in 2011, through its Google Brain division. The goal was to find a way to create neural networks faster so as to embed the technology across many Google applications

By 2015, Google decided to open source TensorFlow, primarily because the company wanted to accelerate the progress of AI. And no doubt, this is what happened. By open sourcing TensorFlow, Google made its technology an industry standard for development. The software has been downloaded over 41 million times, and there are more than 1,800 contributors. In fact, TensorFlow Lite (which is for embedded systems) is running on more than 2 billion mobile devices.¹²

The ubiquity of the platform has resulted in a large ecosystem. This means there are many add-ons like TensorFlow Federated (for decentralized data), TensorFlow Privacy, TensorFlow Probability, TensorFlow Agents (for reinforcement learning), and Mesh TensorFlow (for massive datasets).

To use TensorFlow, you have the option of a variety of languages to create your models, such as Swift, JavaScript, and R. Although, for the most part, the most common one is Python.

In terms of the basic structure, TensorFlow takes in input data as a multidimensional array, which is also known as a tensor. There is a flow to it, represented by a chart, as the data courses through the system.

When you enter commands into TensorFlow, they are processed using a sophisticated C++ kernel. This allows for much higher performance, which can be essential as some models can be massive.

TensorFlow can be used for just about anything when it comes to AI. Here are some of the models that it has powered:

• Researchers from NERSC (National Energy Research Scientific Computing Center) at the Lawrence Berkeley National Laboratory created a deep learning system to better predict extreme weather. It was the first such model that broke the expo (1 billion billion calculations) computing barrier. Because of this, the researchers won the Gordon Bell Prize.¹³
• Airbnb used TensorFlow to build a model that categorized millions of listing photos, which increased the guest experience and led to higher conversions.¹⁴
• Google used TensorFlow to analyze data from NASA’s Kepler space telescope. The result? By training a neural network, the model discovered two exoplanets. Google also made available the code to the public.¹⁵

Google has been working on TensorFlow 2.0, and a key focus is to make the API process simpler. There is also something called Datasets, which helps to streamline the preparation of data for AI models.

Then what are some of the other AI frameworks? Let’s take a look:

• PyTorch: Facebook is the developer of this platform, which was released in 2016. Like TensorFlow, the main language to program the system is Python. While PyTorch is still in the early phases, it is already considered the runner-up to TensorFlow in terms of usage. So what is different with this platform? PyTorch has a more intuitive interface. The platform also allows for dynamic computation of graphs. This means you can easily make changes to your models in runtime, which helps speed up development. PyTorch also makes it possible for having different types of back-end CPUs and GPUs.
• Keras: Even though TensorFlow and PyTorch are for experienced AI experts, Keras is for beginners. With a small amount of code—in Python—you can create neural networks. In the documentation, it notes: “Keras is an API designed for human beings, not machines. It puts user experience front and center. Keras follows best practices for reducing cognitive load: it offers consistent and simple APIs, it minimizes the number of user actions required for common use cases, and it provides clear and actionable feedback upon user error.”¹⁶ There is a “Getting Started” guide that takes only 30 seconds! Yet the simplicity does not mean that it is not powerful. The fact is that you can create sophisticated models with Keras. For example, TensorFlow has integrated Keras on its own platform. Even for those who are pros at AI, the system can be quite useful for doing initial experimentations with models.

With AI development, there is another common tool: Jupyter Notebook. It’s not a platform or development tool. Instead, Jupyter Notebook is a web app that makes it easy to code in Python and R to create visualizations and import AI systems. You can also easily share your work with other people, similar to what GitHub does.

During the past few years, there has also emerged a new category of AI Tools called automated machine learning or autoML. These systems help to deal with processes like data prep and feature selection. For the most part, the goal is to provide help for those organizations that do not have experienced data scientists and AI engineers. This is all about the fast-growing trend of the “citizen data scientist”—that is, a person who does not have a strong technical background who can still create useful models.

Some of the players in the autoML space include H2O.ai, DataRobot, and SaaS. The systems are intuitive and use drag-and-drop ease with the development of models. As should be no surprise, mega tech operators like Facebook and Google have created autoML systems for their own teams. In the case of Facebook, it has Asimo, which helps manage the training and testing of 300,000 models every month.¹⁷

For a use case of autoML, take a look at Lenovo Brazil. The company was having difficulty creating machine learning models to help predict and manage the supply chain. It had two people who coded 1,500 lines of R code each week—but this was not enough. The fact is that it would not be cost-effective to hire more data scientists.

Hence the company implemented DataRobot. By automating various processes, Lenovo Brazil was able to create models with more variables, which led to better results. Within a few months, the number of users of DataRobot went from two to ten.

Table 8-1 shows some other results.¹⁸

Table 8-1.

The results of implementing an autoML system

Tasks	Before	After
Model creation	4 weeks	3 days
Production models	2 days	5 minutes
Accuracy of predictions	<80%	87.5%

Pretty good, right? Absolutely. But there are still come caveats. With Lenovo Brazil, the company had the benefit of skilled data scientists, who understood the nuances of creating models.

However, if you use an autoML tool without such expertise, you could easily run into serious trouble. There’s a good chance that you may create models that have faulty assumptions or data. If anything, the results may ultimately prove far worse than not using AI! Because of this, DataRobot actually requires that a new customer have a dedicated field engineer and data scientist work with the company for the first year.¹⁹

Now there are also low-code platforms that have proven to be useful in accelerating the development of AI projects. One of the leaders in the space is Appian, which has the bold guarantee of “Idea to app in eight weeks.”

With this platform, you can easily set up the data structure that is clean. There are even systems in place to help guide the process, such as alerting for issues. No doubt, this provides a solid foundation for building a model. But low-code also helps in other ways. For example, you can test various AI platforms—say from Google, Amazon, or Microsoft—to see which one performs better. Then you can create the app with a modern interface and deploy it to the Web or mobile apps.

To get a sense of the power of low-code, take a look at what KPMG has done with the technology. The company was able to help its clients transition away from the use of LIBOR in loans. First of all, KPMG used its own AI platform, called Ignite, to ingest the unstructured data and use machine learning and Natural Language Processing to remediate the contracts. Next, the company used Appian to help with document sharing, customizable business rules, and real-time reporting.

Such a process—when done manually—could easily take thousands of hours, with the error rate of 10% to 15%. But when using Ignite/Appian, the accuracy was over 96%. Oh, and the time to process the documents was in seconds