Machine learning is rapidly transforming our world, powering everything from the chatbots we interact with daily to the predictive text that anticipates our thoughts, and the curated content that fills our social media feeds. It’s the intelligence behind autonomous vehicles navigating complex environments and the diagnostic tools that can identify medical conditions from images with increasing accuracy. In today’s business landscape, deploying artificial intelligence often means leveraging machine learning – so much so that the terms are frequently used interchangeably. Machine learning, as a core subfield of artificial intelligence, fundamentally provides computers with the ability to learn without being explicitly programmed for every single task.
“Over the past decade, machine learning has emerged as a critical and arguably the most vital methodology for advancing artificial intelligence,” notes Thomas Malone, an MIT Sloan professor and the founding director of the MIT Center for Collective Intelligence. “This prominence is why the terms AI and machine learning are now almost synonymous; the majority of recent breakthroughs in AI are deeply rooted in machine learning principles.”
The increasing pervasiveness of machine learning means that professionals across all sectors will likely encounter it and require a foundational understanding of this transformative field. A Deloitte survey from 2020 revealed that 67% of companies were already utilizing machine learning, with a striking 97% planning to adopt or expand its use within the following year.
From optimizing manufacturing processes to personalizing retail experiences, and enhancing banking security to even refining bakery operations, established companies are harnessing machine learning to unlock unprecedented value and drive efficiency gains. “Machine learning is currently reshaping, and will continue to revolutionize every industry. It’s imperative for leaders to grasp its basic principles, recognize its potential, and acknowledge its limitations,” emphasizes Aleksander Madry, an MIT computer science professor and director of the MIT Center for Deployable Machine Learning.
While deep technical expertise isn’t universally necessary, a fundamental understanding of what machine learning technology can achieve, and equally importantly, what it cannot, is crucial. “I believe it’s becoming essential for everyone to be aware of these advancements,” Madry adds. This awareness must also extend to the broader social, societal, and ethical implications of machine learning. Dr. Joan LaRovere, MBA ’16, a pediatric cardiac intensive care physician and co-founder of The Virtue Foundation, highlights this point: “It’s vital to engage with and understand these powerful tools, and then carefully consider how to apply them responsibly. We must ensure these tools are used for the collective good. AI holds immense potential for positive impact, and this should always be our guiding principle. How can we leverage this technology to improve our world and benefit everyone?”
Decoding Machine Learning: How Computers Learn
Machine learning is intrinsically linked to artificial intelligence, a broad field aiming to equip machines with the ability to mimic intelligent human behavior. AI systems are engineered to tackle complex tasks in ways that mirror human problem-solving approaches.
Boris Katz, a principal research scientist and head of the InfoLab Group at CSAIL, explains that the overarching goal of AI is to develop computer models that demonstrate “intelligent behaviors” akin to humans. This encompasses enabling machines to interpret visual scenes, comprehend natural language text, and execute actions within the physical world.
Machine learning represents a specific pathway to achieving AI. It was originally defined in the 1950s by AI pioneer Arthur Samuel as “the field of study that empowers computers with the capability to learn without being explicitly programmed.” This definition remains remarkably relevant today.
Shulman, a lecturer at MIT Sloan and head of machine learning at Kensho, a firm specializing in AI for finance and U.S. intelligence, draws a helpful analogy. He contrasts traditional computer programming, which he terms “software 1.0,” with baking. Traditional programming, like a precise baking recipe, demands detailed instructions, specifying exact ingredients and procedures.
However, in numerous scenarios, crafting explicit programs becomes exceedingly complex or even impossible. Consider training a computer to recognize faces – a task effortlessly performed by humans. Articulating the precise rules for facial recognition to a computer is incredibly challenging. Machine learning offers an alternative: it enables computers to learn to program themselves through experience.
The foundation of machine learning is data – whether numerical, visual, or textual. This data can range from financial transactions and images of people to bakery items, equipment maintenance logs, sensor readings, or sales figures. This data is meticulously collected and prepared to serve as training data, the information that will educate the machine learning model. Crucially, the volume of training data significantly impacts the program’s efficacy; generally, more data leads to better learning and performance.
Programmers then select an appropriate machine learning model, feed it the prepared data, and initiate the training process. The model autonomously analyzes the data to identify patterns, relationships, and ultimately, to make predictions. The human element remains vital; programmers can refine the model over time, adjusting parameters to steer it towards enhanced accuracy. Janelle Shane’s website AI Weirdness offers an engaging and often humorous glimpse into the learning curves of machine learning algorithms, showcasing both their remarkable capabilities and their occasional amusing missteps, such as when an algorithm attempted to generate recipes and produced creations like “Chocolate Chicken Chicken Cake.”
To rigorously evaluate the model’s learning, a portion of the initial data is reserved as evaluation data. This unseen data tests the model’s ability to generalize its learning to new, unfamiliar inputs. The outcome of this process is a trained model ready for deployment with new datasets in real-world applications.
Successful machine learning algorithms can fulfill diverse functions. As outlined in a research brief on AI and the future of work co-authored by Malone, MIT professor and CSAIL director Daniela Rus, and Robert Laubacher, associate director of the MIT Center for Collective Intelligence, machine learning systems can be:
“Descriptive, using data to explain past events; predictive, forecasting future outcomes; or prescriptive, recommending optimal actions,” the researchers explain.
Machine learning is further categorized into three primary types, each representing a distinct approach to learning:
Supervised Learning: This is the most prevalent type today. Supervised learning models are trained using labeled datasets. Imagine training an algorithm to recognize dog breeds. It would be fed numerous images, each labeled with the correct dog breed by humans. Through this labeled data, the model learns to associate visual features with specific breeds, progressively improving its accuracy in identifying dogs on its own. The “learning” here is in mapping inputs (images) to outputs (dog breed labels) based on provided examples.
Unsupervised Learning: In contrast, unsupervised learning algorithms explore unlabeled data, seeking inherent patterns without explicit guidance. This approach excels at uncovering hidden structures or trends that humans might not readily identify. For example, analyzing online sales data using unsupervised learning could reveal distinct customer segments based on purchasing behavior, without pre-defined categories. The system “learns” to cluster and categorize data based on similarities and differences it discovers itself.
Reinforcement Learning: This method trains machines through a process of trial and error, guided by a reward system. Reinforcement learning is particularly effective for tasks where optimal actions need to be learned through interaction with an environment, such as game playing or autonomous driving. The algorithm learns by receiving feedback (rewards or penalties) for its actions, gradually refining its strategy to maximize rewards. The “learning” is about discovering sequences of actions that lead to the best outcomes in a given context.
Source: Thomas Malone | MIT Sloan. See: https://bit.ly/3gvRho2, Figure 2.
Malone’s Work of the Future brief emphasizes that machine learning thrives on large datasets – thousands or even millions of examples. Consider Google Translate, which became viable by “learning” from the vast multilingual content available on the internet.
Madry points out that machine learning can unlock insights and automate decisions in situations beyond human capabilities. “Algorithms can not only be more efficient and cost-effective but can also accomplish tasks that are simply beyond human reach,” he states.
Google Search exemplifies this. While humans can search, the scale and speed at which Google’s machine learning models process queries and deliver relevant results are unmatched. “This isn’t about computers replacing human jobs, but rather about computers undertaking tasks that would be economically infeasible if solely reliant on human effort,” Malone clarifies.
Machine learning is also closely intertwined with several other specialized subfields within artificial intelligence:
Natural Language Processing (NLP): This branch focuses on enabling machines to understand and process human language, both spoken and written, moving beyond traditional computer data formats. NLP empowers machines to recognize, interpret, and respond to language, generate new text, and facilitate translation between languages. Familiar technologies like chatbots and virtual assistants such as Siri and Alexa are powered by NLP. The “learning” in NLP involves understanding the nuances of human communication, including grammar, semantics, and context.
Neural Networks: These are a specific and widely used class of machine learning algorithms inspired by the structure of the human brain. Artificial neural networks consist of interconnected processing nodes organized in layers. In these networks, data flows through nodes, each performing computations and passing outputs to others. During training, labeled data adjusts the connections between nodes, allowing the network to “learn” complex patterns. For instance, in image recognition, different nodes might learn to identify edges, shapes, or textures, collectively leading to recognition of objects.
Deep Learning: Deep learning networks are essentially neural networks with multiple layers – often many more than traditional neural networks. These deep architectures enable the processing of enormous datasets and the identification of intricate patterns. In image recognition, for example, early layers might detect basic features like edges, while deeper layers learn to recognize complex objects like faces by combining these features. Deep learning’s ability to learn hierarchical representations of data has fueled breakthroughs in areas like autonomous driving, sophisticated chatbots, and advanced medical diagnostics. However, deep learning’s computational demands raise concerns about economic and environmental sustainability, as highlighted in MIT IDE research.
Machine Learning in Action: Business Applications
For some companies, like Netflix with its recommendation engine and Google with its search algorithm, machine learning is the very core of their business model. Other organizations are deeply integrating machine learning, even if it’s not their primary offering. However, many businesses are still exploring how to effectively leverage machine learning. Shulman observes, “One of the most significant challenges in machine learning is identifying the right problems that machine learning can actually solve. A gap still exists in understanding its applicability.”
[
Leading the AI-Driven Organization
In person at MIT Sloan
Register Now
A 2018 study from the MIT Initiative on the Digital Economy presented a 21-question framework to evaluate the suitability of a task for machine learning. Their findings suggest that while no occupation will remain untouched by machine learning, complete automation of entire occupations is unlikely. The key to successful machine learning implementation lies in reorganizing jobs into specific tasks, some of which are well-suited for machine learning, while others still require human expertise.
Companies are currently deploying machine learning in numerous ways, including:
Recommendation Algorithms: Powering personalized suggestions on platforms like Netflix, YouTube, and Facebook, as well as product recommendations in e-commerce, these algorithms learn user preferences to predict what content or products they are most likely to engage with. “[These algorithms] are continuously learning about our individual tastes,” explains Madry. “Platforms like Twitter and Facebook aim to learn what tweets or ads are most relevant to each user, optimizing content delivery for engagement.” The “learning” is in creating a model of user preferences based on past behavior and using it to predict future choices.
Image Analysis and Object Detection: Machine learning excels at extracting insights from images, including identifying people and differentiating between them. While facial recognition technology raises ethical concerns, its business applications are diverse. Shulman mentions hedge funds using machine learning to analyze parking lot images to gauge company performance based on car counts. The system “learns” to identify and count objects in images, providing data for business intelligence.
Fraud Detection: By analyzing patterns in transactions, spending habits, and login attempts, machine learning systems can learn to identify anomalies indicative of fraudulent activity, such as credit card fraud, unauthorized account access, or spam emails. The algorithm “learns” normal behavior patterns and flags deviations as potentially fraudulent.
Automatic Helplines and Chatbots: Companies are increasingly implementing chatbots for customer service, enabling automated interactions. These bots leverage machine learning and natural language processing to learn from past conversations and provide contextually relevant responses. The chatbot “learns” to understand user queries and provide helpful answers, improving over time with more interactions.
Self-Driving Cars: Autonomous vehicle technology heavily relies on machine learning, particularly deep learning, for tasks like perception, navigation, and decision-making in real-time driving scenarios. The car “learns” to drive by processing sensor data, recognizing objects, and making driving decisions through continuous learning and refinement.
Medical Imaging and Diagnostics: Machine learning algorithms can be trained to analyze medical images like mammograms or other patient data to detect subtle indicators of illness, aiding in early diagnosis and risk prediction, such as cancer risk assessment. The system “learns” to identify patterns in medical data associated with specific conditions, assisting medical professionals in diagnosis.
Read report: Artificial Intelligence and the Future of Work
Promises and Challenges: Understanding the Nuances of Machine Learning
While machine learning drives technological advancements that benefit businesses and workers, leaders must be aware of its inherent limitations and challenges.
Explainability:
A significant concern is the “explainability” of machine learning models – the ability to understand how they arrive at decisions. “It’s crucial to understand why a model behaves as it does,” Madry stresses. “Never treat it as a black box oracle. While utilizing it, strive to discern the underlying rules and validate them.”
What business leaders need to know about AI 7 lessons for successful machine learning projects Why finance is deploying natural language processing
This is particularly important as machine learning systems can be deceived or fail in unexpected ways, even in tasks humans find simple. For instance, subtly altering image metadata can mislead computers, causing a machine to mistakenly identify a dog as an ostrich.
Madry cites an example where a machine learning algorithm analyzing X-rays appeared to outperform physicians in diagnosing tuberculosis. However, the algorithm was actually correlating diagnoses with the X-ray machines themselves, not the images. Older machines, more common in developing countries where tuberculosis is prevalent, became inadvertently associated with positive diagnoses. The algorithm “learned” a spurious correlation, achieving the task in a way that was technically correct but clinically misleading.
The importance of explainability and accuracy varies depending on the application, Shulman notes. While machine learning can solve many well-defined problems, current models typically achieve around 95% human accuracy. This might be acceptable for movie recommendations but insufficient for critical applications like self-driving cars or detecting flaws in machinery.
Bias and Unintended Outcomes:
Machine learning models are trained on data created by humans, and human biases can inadvertently be embedded in algorithms. If biased data reflecting societal inequities is used for training, the algorithm will learn and perpetuate these biases. Chatbots trained on Twitter conversations, for example, can learn and exhibit offensive and racist language.
In some cases, machine learning can exacerbate social issues. Facebook’s use of machine learning to personalize content feeds, intended to increase user engagement, has been linked to algorithms showing users increasingly extreme content, contributing to polarization and the spread of misinformation.
Counteracting bias in machine learning requires careful data vetting and organizational commitment to ethical AI practices. This includes embracing human-centered AI, actively seeking diverse perspectives during AI system design. Initiatives like the Algorithmic Justice League and The Moral Machine project are at the forefront of addressing these critical issues.
Implementing Machine Learning Effectively
Shulman observes that business leaders often struggle to identify where machine learning can genuinely add value to their organizations. What might be a superficial application for one company could be core to another’s strategy. Businesses should avoid simply following trends and instead focus on identifying relevant and impactful use cases.
The successful machine learning strategies of companies like Amazon may not directly translate to a car manufacturer, Shulman points out. While voice assistants are valuable for Amazon, car companies might find greater ROI in applying machine learning to optimize their factory production lines.
“The rapid pace of advancement in machine learning is exciting but also challenging for executives making resource allocation decisions,” Shulman notes.
It’s crucial to avoid viewing machine learning as a solution searching for a problem. Companies should resist the urge to force-fit machine learning into existing processes. Instead, focusing on specific business problems or customer needs that machine learning can address is the most effective approach.
While in-depth technical knowledge isn’t always required, a basic understanding of machine learning is essential. LaRovere emphasizes the importance of interdisciplinary collaboration: “While I’m not a data scientist, I understand machine learning well enough to collaborate effectively with data science teams. This collaboration allows us to address critical questions and achieve meaningful impact. Success in machine learning is truly a team effort.”
Learn More:
Sign-up for a Machine Learning in Business Course.
Watch an Introduction to Machine Learning through MIT OpenCourseWare.
Read about how an AI pioneer thinks companies can use machine learning to transform.
Watch a discussion with two AI experts about machine learning strides and limitations.
Take a look at the seven steps of machine learning.
Read next: 7 lessons for successful machine learning projects
For more info Sara Brown Senior News Editor and Writer [[email protected]](mailto:[email protected] "email at [email protected]")[
Ideas Made to Matter The relationship between machine learning and climate change ](/ideas-made-to-matter/climate-change-and-machine-learning-good-bad-and-unknown “The relationship between machine learning and climate change”)
[
Ideas Made to Matter Finance taps generative AI to streamline accounting, hiring tasks ](/ideas-made-to-matter/finance-taps-generative-ai-to-streamline-accounting-hiring-tasks “Finance taps generative AI to streamline accounting, hiring tasks”)
[
Ideas Made to Matter When humans and AI work best together ](/ideas-made-to-matter/when-humans-and-ai-work-best-together-and-when-each-better-alone “When humans and AI work best together”)
