What is the role of machine learning in AI?
The machine learning branch of AI provides devices, operating systems, and applications the ability to learn without being explicitly programmed beyond the initial training phase. Engineers develop algorithms and models in ML-driven tools so that they evolve as they're exposed to new data and improve the performance of the systems, apps, and devices they power.
Machine learning algorithms and the applications that use them are able to access big data directly. These algorithms then classify data and analyze it to unearth patterns, discover critical insights, make predictions, recommend or take actions, and so on.
ML models come in several types:
This is the most common form of machine learning. Supervised machine learning models are trained using labeled data sets—e.g., images with different subjects that include labels denoting the difference. With time and exposure, the ML model learns and becomes more accurate until it can regularly recognize differences in images without labels.
Models of this kind are "taught" using unlabeled training data sets that include predetermined possible outcomes. They can't make unique predictions or recommendations. However, unsupervised machine learning models can easily spot patterns without intervention—including relationships people wouldn't necessarily know to look for—in vast enterprise databases and similarly large data sets.
Semi-supervised ML models learn from labeled and unlabeled data, in cases where there isn't enough labeled data available for the supervised approach. They will eventually be able to draw unique conclusions from their data sets.
In this system, data scientists and engineers use a trial and error system to teach machine learning models. The algorithm is programmed to execute a function according to defined rules. Correct execution generates a reward signal, while mistakes trigger signals telling the ML what went wrong. Over time, the model learns to complete the task based on the programmed rules—deciding the best order in which to follow those rules on its own.
Neural networks and deep learning
An artificial neural network (ANN) is a complex set of interconnected nodes, each node having its own function. Nodes produce outputs—based on automated inputs of labeled data—that move back and forth between one another, like neurons in the human brain.
If a neural network has one or more layers of processing nodes between its input and output node layers, it's classified as a deep neural network. A deep learning model can process massive, raw unstructured data sets, categorize and analyze the sets, and initiate actions based on the system's conclusions, all with no human intervention. The full potential of deep learning algorithms has yet to be realized, but experts are excited by the prospects.
Key subsets of machine learning
While machine learning has numerous specific uses, the majority of them fall into several distinct categories.
The biggest way ML improves on a "basic" AI system is offering scale and speed. ML models can handle the massive amounts of data necessary to automate time-consuming and/or complex tasks. The way these models are programmed allows them to uncover insights or make predictions faster because it requires virtually no human intervention after the programming and learning stages.
For example, in a field like analytics, the complexity and capabilities of ML algorithms are what allow data teams to run predictive analytics and prescriptive analytics in real or near-real time, and do so at scale. This is because the algorithms are always learning and improving as they're exposed to more—and more types of—data.
Natural language processing
This term refers to the ability of a machine learning platform or application to understand and respond to human language. A natural language processing (NLP) machine learning algorithm is what makes truly accurate speech-to-text writing tools possible. NLP is also the backbone of digital assistants like Siri and Alexa.
Computer vision uses machine learning algorithms to process digital images, video, and other visual data and analyze them. But analysis is just the beginning, as a computer vision platform can also recommend or even initiate actions. As such, it's essential for self-driving cars. More commonly, it's seen in newer smartphones with cameras capable of scanning text and translating it or copying it for other text applications, like messaging.
5 practical machine learning applications
As theoretical as some machine learning model use cases can seem—like driverless cars—plenty of them are commonplace. Here are some of the most well-known examples:
1. Web search
Tools like Google and Bing are able to deliver more than a billion results for a search term in less than a second—because of machine learning. ML ensures that search engines can process misspelled or grammatically incorrect queries and still deliver relevant results. Also, through NLP, voice-activated search can be just as accurate.
2. Customer service
Early chatbots could only answer basic questions with a selection of preset responses before requiring a human customer service agent to intervene. But many of the chatbots seen on today's websites are far more intuitive and can resolve plenty of different issues. This is because they run on ML models.
These advanced chatbots can cue up historical conversation data—not just from a querying individual's record, but also chats about similar issues with other customers—to help ensure an appropriate response every time. This helps boost customer satisfaction, which in turn can maintain brand loyalty.
3. Recommendation engines
Every time an e-commerce shopper using a site like Amazon or Target buys a product or even spends a certain amount of time on a certain product's page, that data eventually makes its way to an ML-powered recommendation engine. The same goes for video streaming services like Netflix and HBO Max.
All of those buying or watching choices form profiles of users' preferences. From there, the engine generates product or viewing recommendations for similar material—usually in a customized webpage or as email and mobile app notifications. The recommendations grow more accurate as user activity increases.
4. Fraud detection
Credit card companies use machine learning algorithms to analyze the records of cardholders' transactions, looking for patterns in where, when, and how they shop. When patterns of odd or out-of-place transactions emerge in real time, the ML may either recommend that the card be locked or initiate the lock itself, requiring the holder to verify the activity.
5. Medical diagnostics
Cutting-edge medical tools can uncover unprecedented insights from raw image data. For example, MIT researchers in 2021 developed an ML model capable of predicting cancer risk with a high degree of accuracy. Developments like these are especially promising because of the rapidly aging population and growing prevalence of late-life diseases. Such advancements may improve the odds that these life-threatening conditions could be discovered and treated earlier.
Advantages—and current limitations—of machine learning
Given that machine learning has become relatively common, most of its notable advantages are clear. The technology's capability for continuous improvement is arguably the biggest of these, as it's an advantage only shared by human intelligence.
Coupled with ML's ability to do things human beings simply cannot—e.g., analyzing hundreds of gigabytes of raw unstructured data in minutes or even seconds—it has clear and present business value in a world where enterprise data generation is only increasing. Furthermore, ML's utility in our most vital sectors, such as healthcare, all but guarantees it can have a positive impact on people's lives as well.
But ML isn't perfect, and its potential limitations must be addressed.
Key challenges and limitations
The most readily apparent challenge facing ML is the concerns over ML and AI ethics. For example, the efforts of data privacy advocates and lawmakers have led to bans on facial recognition technology in two dozen U.S. states and cities.
Fears of such technology play into another issue relevant to all forms of ML and AI—bias. Any ML model can develop racial, ethnic, gender, or other biases if it ingests data that reflects systemic inequality—e.g., average household income data in a city with major wealth disparity. In a stranger example, a recent study of a visual semantic labeling tool analyzing images of kitchen implements found the ML assumed the hands holding them belonged to women, even when that wasn't true.
Social media has unveiled further ML dangers. The algorithms of Facebook, YouTube, Twitter, and other platforms recommend engagement-generating content—and until fairly recently, they had no safeguards to check if the engagement was for hateful, prejudiced, or hyper-partisan material.
Other ML challenges aren't as shocking, but still require consideration. At the programming stage, the technology is still only as effective as the data it's given. Even small glitches in the metadata of a data set can create major errors—even in the deep learning algorithms that are the most advanced form of ML. Ensuring high data quality is a must, and data scientists should be prepared for that task—e.g., by pre-tagging and categorizing data sets for supervised learning.
The role of cloud analytics in the future of ML
Though the future of machine learning remains unwritten, it's not unreasonable to see promise in it based on where the technology currently stands. ML's ability to efficiently process data at a scale ideal for the biggest enterprise analytics is a major example of its present and future possibilities. Using the cloud as the setting for such analytics operations means organizations can take the task of data management, processing, and storage away from on-premises data infrastructure. As cloud technologies become more advanced, using ML for analytics within cloud deployments will mean reaching more valuable insights at a faster pace.
Careful oversight is necessary to gain as much business value from ML as possible while minimizing the effects of its limitations. In addition to being vigilant for bias and data quality, and ensuring models are carefully programmed at the outset of ML initiatives, your enterprise will need to carefully manage the data associated with—and generated by—these projects.
A cloud-ready analytics platform like Teradata Vantage is integral to the integration and management of ML project data. The scalability and elasticity of cloud resources and the availability of cost-efficient object storage are ideal for the data volumes involved in ML. The control and visibility Vantage allows can help you put together a single source of analytics truth—even if some project data remains on-premises.
To learn more about how Vantage can benefit your machine learning projects, check out the case study on Teradata customer Saudi Telecom Company (stc).
AI & Machine Learning: Lessons and Opportunities