Beyond innovation: Leveraging the power of machine learning for business growth
GenAI and machine learning are no longer just R&D experiments. They're core drivers of business growth, and here's a framework for putting them to work.
What is machine learning (ML), genAI, and their role in AI?
Within the large umbrella of Artificial Intelligence (AI), Machine Learning (ML) has been the most common and effective paradigm for driving innovation. In this blog post, we’ll explore the relationship between AI and ML, debunk misconceptions, describe Generative AI (GenAI), and examine benefits and challenges that businesses face when harnessing the power of machine learning for business.
ERA 1: The early days: 1950s — 1990s
The scope of Artificial Intelligence (AI) is continually expanding and reshaping the way businesses operate, compete, and interact with customers. AI, at its core, aims to power machines with human-like intelligence. This includes enabling computers to learn, plan, reason, detect patterns, and make decisions based on data, ideally with competency comparable to or exceeding that of humans. This lofty goal has existed within the world of computation since the 1950s; however, achieving this level of sophistication is no small feat. It’s only in the past decade or so that the technology has become performant enough to serve a role in our daily lives: from driving directions, voice assistants in our home, face recognition at airports, smart search engines, to assisting in medical imaging diagnoses, and more.
ERA 2: Classical machine learning: 1980s — present
By the 1990s, many folks started to focus on a budding, data-centric paradigm within AI named machine learning. This subfield of AI was defined by its distinct departure from relying on humans to manually define every piece of knowledge and intelligence. The name itself signifies that models or machines are designed to learn on their own, to the extent that the latest techniques allow.
Specifically, machine learning aims to learn how to perform a task — in a fairly automated manner — by using statistical and probabilistic models. The humans simply provide:
Appropriate data that is useful for the given task of interest; and
The “blueprint” or key architectural specifications — encompassing details such as the type and size of the computational model, the method for measuring its inaccuracy (commonly known as “loss”), and the algorithm that specifies how the model optimizes itself.
Then, the model tries to learn from the data so that it can perform that task well in the future when it experiences new data. This paradigm has dominated the field of AI and provided unprecedented progress. Many tasks and problems were crafted to fit the style of ML.
As a concrete example, perhaps one wishes to design a system that can accurately recognize digits that are written in an image, like those found on hand-written zip codes of mail envelopes. The user can provide a dataset of images of digits, along with each image’s corresponding, correct value of 0 to 9. This constitutes having “labeled” or annotated data. The user needs to select a computational model (i.e. a a particular type of neural network) to train and make certain design decisions (e.g. what to optimize, how long to train the model, etc). One can view these collective decisions as providing a “blueprint”. Training the model allows it to learn from that data how to make accurate predictions for detecting hand-written digits. Technically, this would be a discriminative model, as it focuses on the model learning to distinguish between various categories or classes of data.
Deep learning enters the chat
In 2012, a particular family of machine learning models named Neural Networks started to provide unprecedented performance improvements over existing state-of-the-art systems. Although Neural Networks were invented decades prior, it was only until 2015–2017 that neural approaches demonstrated undeniably compelling results across an incredibly wide range of problems. Neural Networks encompass many different types of architectures, such as Feed-forward Neural Networks, Convolutional Neural Networks, Recurrent Neural Networks, Transformers, etc.
The technical details of these models are well beyond the intended scope of this blog, as each model could easily require 1–3 academic lectures to adequately describe such. However, it’s worth noting that all neural networks and models have atomic units named neurons. Neurons store numeric data and they are connected to and can impact many other, subsequent neurons within the model. The “formula” the model executes (i.e. the complex, non-linear interactions between all neurons) is essentially self-learned from the training data, based on the “blueprint” the developer designed.
By 2015, Deep Learning became the highest performant paradigm for all of AI. Its prowess has been revolutionary, and it has fueled essentially all significant aspects of the current era of AI. While the name Deep Learning may seem a bit mysterious, it’s simply any neural network that has multiple “hidden” layers of neurons. Further details are irrelevant for this article, but the key takeaways are that:
Machine learning (ML) is a subfield of AI.
ML aims for the model to learn how to perform a particular task well.
Deep Learning refers to neural networks, and it’s the highest performing paradigm within ML.
Deep Learning yields incredible performance across nearly all domains of data, including text in natural language processing, graphics in computer vision, and audio in signal and speech processing.
ERA 3: The genAI era
In contrast to discriminative models, generative AI models aim to model all the data at hand, as if to understand the underlying process by which such data could have come to exist. In doing so, they have the ability to generate new, realistic data.
Although generative AI models have existed since the inception of AI, it was only in 2017 that generative AI models started to become incredibly popular and impactful. This is primarily due to the advent of a particular Neural Network named the Transformer, which was introduced as a language model that could be used for various tasks. Transformers have the option of being used in a method known as Decoding, which can generate new text. In fact, all of your favorite ChatBot-like Large Language Models (LLMs), like ChatGPT, are said to be generative AI models, since they are trained to generate realistic-looking data.
Other generative AI models are highly effective at working with image data, or audio data, as they can generate realistic photos and songs or speech audio. These technologies are now ubiquitous, and they’ll continue to get increasingly better at creating realistic data and dialog interactions. Generative AI models have become a household name, and they’re casually referred to as “GenAI.” They are synonymous. GenAI is simply any AI system that uses a generative model. We are now in a new era of AI: The GenAI era.
As businesses gravitate towards the thrilling new era of computation, it’s worth addressing some common issues and misconceptions around GenAI:
Human-level intelligence: Many falsely believe that AI systems can currently achieve human-like reasoning and understanding of the world. While AI systems include models that can far exceed expert humans on some tasks — such as playing chess or translating thousands of human sentences from different languages — computers generally still severely struggle with demonstrating any significant ability to reason.
Reasoning:This represents a serious limitation in AI. Thus, there’s an entire sub-field of Natural Language Processing (NLP) named reasoning, and it in turn includes many large areas of problems including commonsense reasoning and quantitative reasoning. Commonsense reasoning aims to measure and improve models’ ability to demonstrate skills or knowledge that is common sense to humans. While reasoning is now an incredibly popular area of research, currently, state-of-the-art systems largely still fail to understand many basic concepts.
AI is not equal to ML: Many falsely equate AI and ML. Rather, as we discussed, ML is a subfield of AI.
Biases: Any model is only as good as the data that it was trained on. Truly, machine learning models fully rely on data to train the models. And, because one will never have access to the full set of possible data, a model is always biased by only representing a subset of the entire population of possible data. Thus, by definition, every single computational model is biased. It is thus critically important to understand these biases and minimize them as much as one can.
Understanding these distinctions is a crucial prerequisite for businesses seeking to effectively leverage the benefits of ML.
The time is now: Implementing applications of machine learning in business
The fact of the matter is that ML serves as the crux of many modem computer systems in our daily lives. All of the proceeding eras of ML have led us to a place where it can enhance decision-making processes, optimize operations, and reduce costs. It allows businesses to automate repetitive tasks, detect anomalies, and gain deeper insights from their data. A few industry examples include:
In healthcare, ML has transformed patient care by predicting diseases, optimizing treatment plans, and improving diagnostics.
In finance, it’s the driving force behind automated trading algorithms and fraud detection systems.
In e-commerce, ML personalizes a consumer’s shopping experience, recommending products and improving satisfaction when interacting with the brand.
While the benefits are clear, there are challenges businesses must recognize when adopting ML solutions:
Data quality vs. quantity: High-quality data is essential for ML success. It’s not just about the volume of data but also its accuracy and relevance. Businesses must invest in data quality processes.
Business integrity: Bias in ML models is a significant concern. Ensuring fairness and avoiding discriminatory outcomes is an important part of understanding your algorithms and mitigating bias if your data is not holistic.
Scalability: As businesses grow, their ML infrastructure must scale accordingly. Scalability challenges can hinder growth, making it crucial to design ML systems with expansion in mind.
Talent: Finding and retaining skilled ML professionals can be a hurdle, and often an expensive business unit to invest in. For example, the demand for skills in ML engineering and research has greatly exceeded the supply of high-quality talent. Each year, most top-research universities only graduate 3–10 PhD students who strictly research ML or NLP for their dissertation topic. Businesses should invest in talent development and recruitment strategies to build their ML capabilities.
ML is an indispensable field of AI. It empowers businesses to make data-driven decisions, achieve remarkable efficiencies, and deliver personalized experiences to customers. However, it’s not without its challenges. By understanding the nuances of AI and ML, embracing data quality, ensuring fairness, planning for scalability, and investing in talent, businesses can harness the transformative potential of ML for their success.
p.s. For business leaders eager to explore ML solutions for business beyond this blog post, we recommend diving into additional resources and references. One recommended book is Automate This: How Algorithms Took Over Our Markets, Our Jobs, and the World. In terms of educational content, Harvard IACS, MIT, Stanford, CMU, UW, and many other prominent institutions have a plethora of free and paid online courses and bootcamps.
FAQ — generated by AI
What is the core goal of Artificial Intelligence (AI)?AI aims to imbue machines with human-like intelligence, enabling them to learn, reason, detect patterns, and make decisions based on data.
How did Machine Learning (ML) evolve from the 1950s to the present?ML transitioned from Classical AI, relying on manually programmed rules, to a data-centric paradigm in the 1990s, allowing machines to learn tasks autonomously.
What characterizes the GenAI era in AI development?The GenAI era, marked by generative AI models like Transformers, focuses on understanding and generating new, realistic data, leading to advancements in natural language processing and image/audio generation.