Machine Learning and Deep Learning Deep Dive

In the age of AI implementation, the current period of AI we find ourselves in, we must understand the pros and cons of both machine learning (ML) and deep learning (DL) in order to best navigate when to use either technology. Some other terms you might have come across with respect to AI/ML tools are applied AI or deep tech. As we’ve mentioned a few times over the course of this book, the underlying tech that will, for the most part, power AI products will be ML or DL. That’s because expert- or rule-based systems are slowly being powered by ML or not evolving at all. So, let’s dive a bit further into these technologies and understand how they differ.

In this chapter, we will explore the relationship between ML and DL and the way in which they bring their own sets of expectations, explanations, and elucidations to builders and users alike. Whether you work with products that incorporate ML models that have been around...

The old – exploring ML

ML models attempt to create some representation of reality in order to help us make some sort of data-driven decision. Essentially, we use mathematics to represent some phenomenon that’s happening in the real world. ML essentially takes mathematics and statistics to predict or classify some future state. The paths diverge in one of two ways. The first group lies with the emergence of models that continue to progress through statistical models and the second group lies with the emergence of models that try to mimic our own natural neural intelligence. Colloquially, these are referred to as traditional ML and DL models.

You can think of all the models we covered in the Model types – from linear regression to neural networks section of Chapter 2 as ML models, but we didn’t cover ANNs in great depth. We’ll discuss those further in the Types of neural networks section later on in this chapter. In this section, we will take a look...

The new – exploring DL

Part of our intention with separating ML and DL conceptually in this book is really to create associations in the reader’s mind. For most technical folks in the field, there are specific models and algorithms that come up when you see “ML” versus “DL” as a descriptor on a product. Quick reminder here that DL is a subset of ML. If you ever get confused by the two terms, just remember that DL is a form of ML that’s grown and evolved to form its own ecosystem. Our aim is to demystify that ecosystem as much as possible so that you can confidently understand the dynamics at play with DL products as a product manager.

The foundational idea of DL is centered around our own biological neural networks and DL uses what’s often the umbrella term of ANNs to solve complex problems. As we will see in the next section, much of the ecosystem that’s been formed in DL has been inspired by our own brains, where the...

Emerging technologies – ancillary and related tech

ML and DL have been used heavily in applications related to natural language processing (natural language generation (NLG), as well as natural language understanding (NLU)), speech recognition, chatbots, virtual agents and assistants, decision management, process automation, text analytics, biometrics, cybersecurity, content creation, image and emotion recognition, and marketing automation. It’s important to remember, particularly from a product manager’s perspective, that AI will increasingly work its way into more of how we live our lives and do our work. This is doubly true if you work in an innovative capacity as a product manager where you’re involved with the ideation and creation of new use cases and MVPs for future products.

Over the passage of time, we’ll see AI continue to augment our workforce both through the process of internal automation as well as through the adoption of AI-based...

Explainability – optimizing for ethics, caveats, and responsibility

Ethics and responsibility play a foundational role in dealing with your customers’ data and behavior and because most of you will build products that help assist humans to make decisions, eventually someone is going to ask you how your product arrives at conclusions. Critical thinking is one of the foundational cornerstones of human reasoning and if your product is rooted in DL, your answer won’t be able to truly satisfy anyone’s skepticism. Our heartfelt advice is this: don’t create a product that will harm people, get you sued, or pose a risk to your business.

If you’re leveraging ML or DL in a capacity that has even the potential to cause harm to others, if there’s a clear bias that affects underrepresented or minority groups (in terms of race, gender, or culture), go back to the ideation phase. This is true whether that’s immediate or downstream harm. This...

Accuracy – optimizing for success

When it comes to DL, we can only truly grapple with its performance. Even from a performance perspective, a lot of DL projects fail to give the results their own engineers are hoping for, so it’s important to manage expectations. This is doubly true if you’re managing the expectations of your leadership team as well. If you’re a product manager or entrepreneur and you’re thinking of incorporating DL, do so in the spirit of science and curiosity. Remain open about your expectations.

But make sure you’re setting your team up for success. A big part of your ANN’s performance also lies in the data preparation you take before you start training your models. Passing your data through an ANN is the last step in your pipeline. If you don’t have good validation or if the quality of your data is poor, you’re not going to see positive results. Then, once you feel confident that you have enough...

Summary

We got the chance to go deep into DL in this chapter and understand some of the major social and historical influences that impact this subsection of ML. We also got the chance to look at some of the specific ANNs that are most commonly used in products powered by DL in order to get more familiar with the actual models we might come across as we build with DL. We ended the chapter with a look at some of the other emerging technologies that collaborate with DL, as well as getting further into some of the concepts that impact DL most: explainability and accuracy.

DL ANNs are super powerful and exhibit great performance, but if you need to explain them, you will run into more issues than you would if you stick to more traditional ML models. We’ve now spent the first three chapters of the book getting familiar with the more technical side of AI product management. Now that we’ve got that foundation covered, we can spend some time contextualizing all this tech.

...

References

• Some Methods for Classification and Analysis of Multivariate Observations: https://books.google.com/
• K-Nearest Neighbors Algorithm: Classification and Regression Star https://www.historyofdatascience.com/k-nearest-neighbors-algorithm-classification-and-regression-star/
• Random Forests https://www.stat.berkeley.edu/~breiman/randomforest2001.pdf
• Decision Trees https://www.cse.unr.edu/~bebis/CS479/PaperPresentations/DecisionTrees.pdf
• Support Vector Machine: The most popular machine learning algorithm https://cml.rhul.ac.uk/svm.html
• Logistic Regression https://uc-r.github.io/logistic_regression#:~:text=Logistic%20regression%20(aka%20logit%20regression,more%20predictor%20variables%20(X)
• Logistic Regression History https://holypython.com/log-reg/logistic-regression-history/
• Bayes Classifier https://www.sciencedirect.com/topics/computer-science/bayes-classifier#:~:text=Na%C3%AFve%20Bayes%20classifier%20(also%20known,use%20Na%C3%AFve%20Bayes%20since...