Machine Learning – what is it and how is it used?
Recent growth in Artificial Intelligence (AI) takes advantage of Machine Learning (ML) to fuel rapid technological development in many ways.
One of those is within cybersecurity, especially when it comes to learning how to combat the ever-growing threat of malicious bots and cybercriminals deploying AI as a means of accessing personal data and hijacking business websites for financial gain.
History of Machine Learning
While AI and ML may seem like relatively new advances, the terminology for them both dates back to 1959.
Arthur Samuel, an IBM employee and pioneer in the field of computer gaming and artificial intelligence first used it in his article “Some Studies in Machine Learning Using the Game of Checkers” for the IBM Journal of Research and Development.
He described ML as “the field of study that gives computers the ability to learn without explicitly being programmed.”
The Samuel Checkers-playing Program was among the world’s first successful self-learning programs, and as such a very early demonstration of the fundamental concept of Artificial Intelligence (AI).
Modern ML is automating tasks and uncovering insights from complex data patterns that are beyond human capability to detect.
Categories of ML
Broadly speaking, ML is broken down into three sub-categories.
Supervised
Here ML models are trained with labelled data sets that allow them to learn and grow more accurate over time. An example would include an algorithm that is trained with images of a specific type – ie photos of dogs – plus images of other non-canine objects, which have been pre-identified by humans. The machine would then learn ways to identify images of dogs on its own. Supervised machine learning is the most common form in use.
Unsupervised
With unsupervised ML, a program looks for patterns in data which isn’t labelled. It can find patterns and trends that people often aren’t specifically looking for. One example could be looking through sales data and identifying the different types of clients who make purchases.
Reinforcement
This type of ML trains machines through a series of trial and error to take the best action by using a “reward system”. Using reinforcement ML can train models to play games or train driverless vehicles. Telling the machine it made a correct decision helps it to learn over time what actions it should be taking.
How does it work?
ML is a type of AI that “learns” as it recognises and analyses new patterns from the data it is given.
It uses algorithms to identify patterns and trends in data and then uses those to make predictions and decisions based on what it has “learnt”. ML is used to then build predictive models, classify data and recognise patterns and works as an essential component of many AI applications.
When you log into your streaming app and get recommendations for what to watch next, or suggestions for friends to connect with on social media, books you might like to read on your Kindle or other device? That’s all ML in action.
ML is behind chatbots and predictive text, language translation apps, being able to use your phone camera to identify a plant, and thousands of other applications.
MIT Sloan professor Thomas W. Malone, the founding director of the MIT Center for Collective Intelligence, said: “In just the last five or 10 years, machine learning has become a critical way, arguably the most important way, most parts of AI are done.
“So that’s why some people use the terms AI and machine learning almost as synonymous … most of the current advances in AI have involved machine learning.”
A 2020 Deloitte survey found that 67% of companies were using ML, and 97% were using or planning to use it in 2021. In the three years since then, ML has developed at a rapid pace with it even being used to diagnose some medical conditions based on image scans.
Malone, in a research paper co-authored by MIT professor and CSAIL director Daniela Rus and Robert Laubacher, the associate director of the MIT Center for Collective Intelligence, said:
“The function of a machine learning system can be descriptive, meaning that the system uses the data to explain what happened; predictive, meaning the system uses the data to predict what will happen; or prescriptive, meaning the system will use the data to make suggestions about what action to take.”
The researchers believe dramatic advances in automation and computation go hand in hand with improved opportunities and economic security for workers.
In the Work of the Future brief, Malone noted that machine learning is best suited for situations with lots of data — thousands or millions of examples, like recordings from previous conversations with customers, sensor logs from machines, or ATM transactions.
A great example for this is Google’s Search function. It is doing something a human could do – look up relevant links on an internet database – but at a speed and scale which could never be possible each time someone types in a query.
Veracity Trust Network’s Martin Barker, Head of IT (Global) & Data Protection Officer, said: “Machine learning has been around for decades, but recent advancements, particularly in Large Language Models (LLMs), have made AI more accessible and user-friendly.
“These models allow users to interact with AI in their native language, simplifying the process. While AI technologies like Photoshop’s auto-generative fill have existed for some time, their capabilities have significantly improved with the integration of AI.”
He added: “At its core, AI operates through complex decision trees. An AI system takes input parameters and processes them through a series of true or false decisions, ultimately producing an output. This process involves millions of binary questions that guide the AI to its final result. ML is the process of training this AI, the system learns from the data to make the decision tree.”
Who is using ML?
According to PwC’s 2024 AI Business Predictions:
“Seventy-three percent of US companies have already adopted AI in at least some areas of their business – according to our 2023 Emerging Technology Survey – and generative AI (GenAI) is leading the way.
However, PwC also warns that only a few companies are getting these technologies right. It refers to these companies – 7% of respondents in its survey – as EmTech Accelerators.
The survey data not only identifies these companies, it pinpoints four practices that likely explain their success:
- Think big: Use emerging tech for reinvention;
- Commit to it: Allocate the right resources;
- Think convergence: Integrate your emerging tech;
- Drive integration: Embed emerging tech into your business strategy.
According to the survey, EmTech Accelerators also stand apart because they’re already more likely to have already implemented some form of GenAI tech (88% versus 54%).
Although ML has many applications, the term AI has become a buzzword. For Veracity’s Barker its most common misuse comes in marketing.
“Products like ‘AI Thermal Paste’ misuse the term, as AI has no role in their creation. Legally, anything produced by AI is considered public domain, as copyright and patent laws require human origin.
“Unfortunately, AI is also exploited for fraud, scams, and other malicious activities, much like any other human invention.”
Challenges with AI and ML
AI faces significant issues related to overuse and misuse. One major concern is the phenomenon of AI learning from other AIs, leading to an “echo chamber” effect. This can result in the rapid propagation of errors, as one AI’s hallucinations can be learned by others, compounding inaccuracies.
Despite these challenges, AI principles have been in use for decades. For example, since the 1990s, antivirus software has utilised pattern recognition to identify viruses.
MIT computer science professor Aleksander Madry, director of the MIT Center for Deployable Machine Learning, said there are limits for what ML can do at the present time and systems can be deliberately fooled.
He said: “Understanding why a model does what it does is actually a very difficult question, and you always have to ask yourself that.”
In an article for Popular Science published in 2016, titled Fooling the machine Alex Kantchelian, a researcher at Berkeley University who studies adversarial machine learning attacks, said: “Any system that uses machine learning for making security-critical decisions is potentially vulnerable to these kinds of attacks.”
These attacks use adversarial examples: images, sounds, or potentially text that seems normal to human viewers, but are perceived as something else entirely by machines. Small changes made by attackers can force a deep neural network to draw incorrect conclusions about what it’s being shown.
Machines are trained by humans, and human biases can be incorporated into algorithms – if biased information, or data that reflects existing inequities, is fed to a machine learning program, the program will learn to replicate it and perpetuate forms of discrimination.
Chatbots trained on how people converse on X (formerly Twitter) can pick up on offensive and racist language, for example and Facebook came under fire in 2016 after it was revealed that personal data from American users had been harvested by Cambridge Analytica to try and influence their votes and Russian hackers had been using the platform to disseminate hate and speech and fake news.
Nicholas Barrett, Technology reporter at the BBC, in an article looking at how the way people interact has been affected by social media algorithms, wrote: “This year alone, governments around the world have attempted to limit the impacts of harmful content and disinformation on social media – effects that are amplified by algorithms.”
And a bipartisan group of 14 attorneys general from across the USA have this month (October 2024) launched a lawsuit accusing the social media app TikTok of helping to drive a mental health crisis among teenagers.
The suit alleges: “…that the company uses addictive features to hook children to the app and that it has intentionally misled the public about the safety of prolonged use.”
For MIT Sloan senior lecturer Renée Richardson Gosline, with human-centred AI, algorithms and humans can work together to compensate for blind spots and create clearer outcomes.
“There has been a tremendous amount of research pointing out issues of algorithmic bias and the threat this poses systemically,” Gosline says in a new MIT Sloan Experts Series talk.
“This is a massive issue — one that I don’t think we can take seriously enough.”
Ethical considerations for its use
There are a number of different lawsuits going through the courts at the moment regarding claims of copyright over the use of GenAI.
A group of 11 nonfiction authors joined a lawsuit in Manhattan federal court that accuses OpenAI and Microsoft (MSFT.O), of misusing books the authors have written to train the models behind OpenAI’s popular chatbot ChatGPT and other AI- based software.
AI company Anthropic is also the subject of a class-action lawsuit in California federal court by three authors who say it misused their books and hundreds of thousands of others to train its AI-powered chatbot Claude, which generates texts in response to users’ prompts.
And artists suing GenAI art generators have cleared a major hurdle in a first-of-its-kind lawsuit over the uncompensated and unauthorised use of billions of images downloaded from the internet to train AI systems, with a federal judge allowing key claims to move forward earlier this year.
US District Judge William Orrick ruled that the companies Stability AI, Midjourney, DeviantArt and Runway AI were violating artists’ rights by illegally storing their works in their image generation systems.
Karla Ortiz, who brought the lawsuit, has worked on projects like Black Panther, Avengers: Infinity War and Thor: Ragnarok and is credited with coming up with the main character design for Doctor Strange.
Getty Images is also suing Stability AI in the UK for illegally copying and processing millions of its copyrighted images. That case is expected to reach the High Court in 2025.
Veracity’s Barker said: “Ethically, AI and ML is no different from other technologies; it can be abused. Companies that use copyrighted works to train AI models without proper licensing are committing copyright theft and should be held accountable.”
The long-term effects of AL and ML
It’s a concerning thought that – if it were a country – the internet would be the seventh largest polluter in the world.
There are also worries over the amount of power needed to operate AI centres and many are turning to nuclear options, with Google this week (October 2024) becoming the latest. The internet giant signed a deal with Kairos Power to use small nuclear reactors to generate the vast amounts of energy needed to power its artificial intelligence (AI) data centres.
Machine learning is changing, or will change, every industry, and leaders need to understand the basic principles, the potential, and the limitations said MIT’s Madry.
In a 2018 paper, researchers from the MIT Initiative on the Digital Economy outlined a 21-question rubric to determine whether a task is suitable for machine learning.
It concluded that the way to ML success was to reorganise jobs into discrete tasks, some which could be carried out by a computer using an algorithm, and others that would always require human input.
While AI has the potential to replace certain jobs, this is part of technological progress. Historically, advancements like conveyor belts have replaced manual labour, and AI is no different.
Barker added: “In programming, for example, AI can assist but not replace human creativity and innovation. Junior programmers can adapt and grow, ensuring their roles remain relevant.”
