Macro and Micro AI for Your Product

The term AI is often used as an umbrella term that encapsulates the idea that a machine, whether physical or not, is mimicking the way humans either think, work, speak, express, or understand. This is a pretty big concept to wrap up in one term. It’s hard to embody not just the diversity of models and use cases but the implementation of these models and use cases themselves. This chapter will serve as a handy recap of the various types of AI that products can absorb as you begin to explore the various ways you can leverage AI, as well as some of the most successful examples and common mistakes that can arise as PMs build AI products.

It’s important to note that the application of AI has seen many iterations and will continue to do so as we build into the next decade, but it’s hard to understate the importance of what this technological wave can offer PMs and companies. This is a substantial wave of innovation that’s...

Macro AI – Foundations and umbrellas

So far, we’ve talked a great deal about ML and DL models in the previous chapters. This was intentional because most of the time when we see AI advertised to us through various products, this is what the underlying technology employed is—for the most part. It’s either a DL or an ML algorithm that’s powering the products we’ve discussed. But as you’ve seen in the previous chapters, AI is a great umbrella term that can actually mean more than just an ML or a DL model is being used.

There are a number of major domains in AI that don’t involve ML and DL. We’ve minimally touched on the other areas but haven’t given them their due in their impact and contributions to the AI landscape. Focusing only on ML and DL makes sense from a practical perspective to offer AI technologists, entrepreneurs, and PMs the best chance to pitch their AI products to investors and users, but it also leaves...

ML

ML is a general concept that gives us the basic thinking and organizing power that gives machines the ability to reason with data in a way a human might. You can think of this as a way for a machine to mimic how a human can think, understand, and work. It dominates the market, and it serves as its own umbrella term of sorts because within it we can break down ML into the models we associate with traditional ML, along with their specializations such as computer vision, natural language processing (NLP), and DL as further subsets of ML.

We’ve gone over the specific models and algorithms that are used in ML in Chapters 2 and 3 of this book, so we won’t go into those in detail here, but suffice it to say this is where there are major categories in the kind of learning these machines do. To jog your memory, those learning types are supervised learning, unsupervised learning, reinforcement learning, and deep learning (neural networks). All these types of learning can...

Robotics

If AI is considered an umbrella term for machines mimicking how humans might work and reach conclusions, this is perhaps most visibly obvious with robotics, which is physically trying to reproduce the work of a human. The term robot is in and of itself a generic term that encompasses a lot of nuances, much like the term AI. ML is considered as powerful as it is because of the ability of machines to learn from past actions and behaviors. In that sense, perhaps they are considered more advanced than the underlying tech that powers robotics, but we felt it wasn’t fair to exclude robotics from AI because if a robot can make a meal, make a part for a car, or assist with surgery, it’s intelligent enough to be considered in the realm of AI.

The heart of innovations in this space will also come from ML being used in combination with robotics so that it can learn from the past. In Chapter 3, we briefly touched on the idea of Boston Dynamics robot dogs not using ML...

Expert systems

An expert system is a generic term meant to describe some sort of rule-based engine that’s been continuously refined and optimized over time. They were used heavily in the past in medical and legal use cases before ML became popularized. Although there are some companies that may still rely on them, their relevance has receded. They have a user interface (UI) and are powered by an inference engine that is connected to a knowledge base of some sort. It’s a more basic form of AI that’s made up of If Then statements. A rule-based engine means there are a set of pre-programmed instructions and algorithms that have been programmed into the backbone of how a product or system functions and there is an absence of self-learning. This means that ML models are not used and the system is not learning over time. Though this might sound like a dumb system, it’s still considered AI because it still might be functioning in a way that mimics how a human might...

Fuzzy logic/fuzzy matching

Fuzzy matching, also referred to as approximate string matching, uses some logic to find terms or phrases that are similar to each other. Perhaps you’re looking through your database to isolate anyone with a first name John but some entries are Jonathan or Johnny. Fuzzy matching would be an intelligent way of finding those alternative names. Fuzzy matching was used ubiquitously in translation software before machine translation came into the picture. Whether you’re looking for alternative naming conventions or mistakes, fuzzy logic and matching are able to offer us intelligent ways for machines to find the things we’re looking for.

As with robotics and other areas of AI, we can see an ensemble with fuzzy matching as well. We’re seeing ML applied to fuzzy matching in an effort to improve accuracy. But even without ML, fuzzy logic and fuzzy matching can stand on their own as a subset of AI that’s been relied on heavily...

Micro AI – Feature level

It can be daunting to understand how various categories of AI fit together, and the reality is that in real-world AI product applications, many of these are working in concert. Seeing various examples of how that happens, particularly when we get to the later sections in this chapter, will offer us a way to see how much opportunity and potential AI really offers us!

We will consolidate ML, DL, computer vision, and NLP into their own section because these models are often used collaboratively as well. That collaboration can then bleed into the other subsets of AI. Robotics, expert systems, and fuzzy logic can remain in their own sections because their applications are so specialized in and of themselves. Seeing how subsets of AI can work together further results in the greater complexity and growth that powers innovation for our markets and brings to market products that serve, delight, and capture our hearts.

ML (traditional/DL/computer vision/NLP)

Which type of model you’re using will depend on your use case and goals for your product. As we’ve covered in varying chapters, the exact model you go with will depend on the data you have, how you’re able to tune your hyperparameters, and what level of explainability and transparency you’ll need for your use case. We’re focusing on AI/ML native products in this section of the book and, as such, identifying which ML model(s) you will use for the foundation of your product will be an important decision, and all features you add onto your core product will also be an act of doing a cost-benefit analysis of the models you’re adding to power those features.

Most products that are out there right now are not AI/ML native in that they are existing software programs and packages that are incrementally adding new AI features and then rebranding their products as AI products. This isn’t exactly true...

Successes – Examples that inspire

In this section, we will be looking at examples of complex, collaborative AI products that use a number of models and build a product intuition of how they can be inspired by examples where companies had a considerable amount of commercial success. The purpose of this section is to show real-life examples where a product used an assortment of AI/ML specializations to deliver a product that gave value to its end users and market.

The product examples we will be covering in the following section include Lensa (a generative AI selfie app) and PeriWatch Vigilance (a health app for mothers and babies made by PeriGen).

Lensa

Given the current sensation of the Lensa app, we thought this would serve as a great first example. Lensa took the internet by storm with its fantasy AI selfie-generation app. The idea is you feed it between 10 and 20 images for the neural networks to learn from, and based on that training, it will generate 50 or more...

Challenges – Common pitfalls

We’ve spent a considerable amount of time talking about how to build AI/ML products and use models in a way that empowers your products. We’ve also discussed the hype and commercial excitement about AI. In this section, we’ll temper this hype by understanding why certain AI/ML products fail. We’ll be looking at a few real-world examples that highlight some of the common reasons why AI deployments have received controversy. We will also look into some of the underlying themes within that controversy for new AI products and their creators to try to avoid.

In the following sections, we will focus on challenges associated with ethics, performance, and safety and their accompanying examples.

Ethics

Companies have long struggled with maintaining the quality and ethics of consumer-facing conversational AIs. If you recall back in 2016 when Microsoft unleashed its AI named Tay onto the Twittersphere, it took less than...

Summary

In this chapter, we’ve covered a lot of ground. We’ve discussed the various areas of AI at a high level, giving us a macro landscape of the variety of options we can take when building AI products. We’ve also brought those options down to the feature level, giving us a micro view of applied AI features. We were then able to look at a few examples of collaborative AI products that have received positive feedback and acclaim, along with a few examples that highlight the challenges of AI products. Building AI products is still new. We’re still building out new use cases, and with every new AI product that comes to market, we’re able to discover new pathways to use these algorithms.

This means that every newly applied use case has the potential to show the world what AI can do, and that’s what makes the current phase we’re in so exciting. In order to uncover innovative new uses for AI/ML we must be willing to make mistakes and...

References

• https://www.hansonrobotics.com/sophia/
• https://www.theverge.com/2016/3/24/11297050/tay-microsoft-chatbot-racist
• https://theintercept.com/2022/12/08/openai-chatgpt-ai-bias-ethics/
• https://www.thedailybeast.com/openais-impressive-chatgpt-chatbot-is-not-immune-to-racism
• https://www.theverge.com/tldr/2020/11/3/21547392/ai-camera-operator-football-bald-head-soccer-mistakes
• Stories of AI Failure and How to Avoid Similar AI Fails: https://www.lexalytics.com/blog/stories-ai-failure-avoid-ai-fails-2020/
• Three Insights From Google’s ‘Failed’ Field Test To Use AI For Medical Diagnosis: https://www.forbes.com/sites/forbestechcouncil/2020/06/09/three-insights-from-googles-failed-field-test-to-use-ai-for-medical-diagnosis/?sh=2b4233c4bac4
• https://www.theverge.com/2021/9/9/22660467/irobot-roomba-ai-dog-poop-avoidance-j7-specs-price