Machine Learning models and their applications

Written by: Alissa Machin | Edited by: Salar Komeyshi |

Machine Learning is a subset of Artificial Intelligence that works on data and algorithms to improve the capability of AI to imitate the way humans learn. Machine Learning intersects closely with Deep Learning, its subset, which we explore in another article. Over time, computer scientists can train Machine Learning models to increase their accuracy. Machine Learning algorithms form the backbone of Machine Learning systems, which in turn determine the kind of model they produce. In everyday life, Machine Learning has produced speech recognition technologies, online chatbots, and algorithmically determined recommendation engines. The potential of Machine Learning to shape a new paradigm of living and working in a technologically enabled society raises ethical questions about privacy, bias, and accountability, which we will explore in the following article. 

What is Machine Learning?

Machine Learning is a subset of AI and operates in a narrower context than general AI. In turn, Deep Learning in turn is a subset of Machine Learning. As we see in a related article on Deep Learning, what distinguishes Machine Learning and Deep Learning from each other is that Deep Learning involves at least three layers of neural networks, which are Machine Learning Models designed to mirror the human brain’s neuronal pathways and prediction-making processes. 

Machine Learning algorithms consist of three main parts. First of all, Machine Learning algorithms take input data and scan the patterns they make to inform a prediction. An error function of a Machine Learning algorithm can then compare the prediction it has generated with any known examples, thereby evaluating the accuracy of the model. Finally, the algorithm adjusts to the data points in the training set, in order to align the model prediction more closely with the known example. An iterative cycle can then result from this initial process, producing a model optimisation which will incrementally reach a threshold of accuracy. We will now consider various Machine Learning models involved in this process of model optimisation: 

1. Supervised learning methods
Supervised learning methods utilise labelled datasets, which consist of raw data assigned labels to supply context, in the process of training algorithms to classify data or predict outcomes. The supervised learning model receives input data and adjusts its weights accordingly to fit the data. In this part of the process, the adjustment step is crucial to enable the model to fit the data without relating too narrowly or broadly to the data and losing the capacity to accurately predict outcomes. Tools such as spam classification within an email inbox are an example of supervised learning. Supervised learning methods can also involve neural networks. In the case of focalx, our AI vehicle scanning software improves in accuracy the more vehicle data enters the system, enabling the algorithms to fine tune their scanning capabilities exponentially. 

2. Unsupervised learning methods
Unsupervised learning methods mirror supervised learning methods in the fundamental way they train algorithms to work on datasets. The difference with unsupervised learning is that the data they work on is unlabeled and form datasets also known as clusters. In the case of unsupervised learning, algorithms expose hidden patterns or clusters of data, revealing similarities and differences in information. As a tool, unsupervised learning lends itself well to exploratory data analysis (EDA), a useful research method for testing hypotheses and spotting anomalies. Here, principal component analysis (PCA) and singular value decomposition (SVD) are common EDA techniques. In a marketing context, unsupervised learning can inform cross-selling strategies and enable customer segmentation. Like supervised learning methods, unsupervised learning methods can involve algorithms such as neural networks, as well as domain-specific methods, such as probabilistic clustering algorithms.

3.  Semi-supervised learning methods
Semi-supervised learning methods are a hybrid of supervised and unsupervised learning methods. Semi-supervised learning algorithms train smaller labelled datasets to direct classification whilst extracting from a larger, unlabelled dataset. In this regard, semi-supervised learning methods can balance out the bias in labels which can emerge in supervised learning methods whilst also prioritising the higher degree of accuracy in using the clear target outcomes of supervised learning methods, in contrast to unsupervised learning methods. Moreover, an advantage of prioritising semi-supervised learning methods is adapting to contexts where not enough labelled data exists to train supervised learning algorithms. As a solution for economising on resources when labelling data becomes costly, semi-supervised learning methods can enable the embedding of AI in smaller businesses and organisations.

4. Reinforcement learning methods
Reinforcement learning methods are closely connected to supervised learning methods, the difference being that reinforcement learning methods do not train algorithms using sample data. Instead, reinforcement learning models learn by trial and error, generating the best solution to a given problem after encountering various mismatches along the way. A recent example of reinforcement learning is in the area of robot locomotion, where humanoid robots follow the direction of reinforcement learning inputs, successfully navigating indoor and outdoor environments. 

Common applications of Machine Learning in everyday life

Common applications of Machine Learning in everyday life are in language-based settings, where language models can transform speech signals into commands, including in speech recognition technologies like Siri, or Virtual personal assistants like Amazon’s Alexa. In both of these contexts, natural language processing (NLP) combines a rule-based model of human language with Machine Learning models. Another real-life example of Machine Learning is the use of data-driven recommendation engines which uncover consumer behaviour to drive predictive models.

1. Speech recognition technologies

Speech recognition, also known as Automatic Speech Recognition (ASR), speech-to-text, or computer speech recognition harnesses NLP capabilities to transform human speech into a written format. Mobile devices and tablets often embed ASR into their systems, such as Siri or Google Assistant. ASR functions allow for improved accessibility for texting. In the context of the automotive industry, speech recognisers can utilise voice-activated navigation systems and search capabilities embedded in car radios to improve driver safety.

2. Online chatbots and virtual agents

Online chatbots and virtual agents are another everyday application of Machine Learning. In customer service areas such as online banking, AI-powered chatbots can synchronise with Customer Relationship Management (CRM) systems. Here, these chatbots can integrate with customer data platforms in the prediction of common issues and provide customers with a 24/7 personalised assistance service. In addition, online chatbots serve marketers in promoting their products through customer engagement on websites and social media platforms. One of the most useful functions of chatbots is that they can respond to FAQs (Frequently Asked Questions), providing relevant support for customers related to advice, shipping, and other relevant product recommendations. By training AI models to work on customer datasets, companies can use prediction to improve the relevance and usefulness of their messaging for customers. Some of the most common messaging bots include Slack or Facebook Messenger. Conversational AI therefore represents the frontier of how AI can mimic human speech and conversation.

3. Consumer data-driven recommendation engines 

Consumer data-driven recommendation engines are a useful means of using AI algorithms to identify patterns in data to improve cross-selling strategies. By drawing on data which reflects the history of a consumer’s behaviour and their interaction with products, these algorithms can make targeted product recommendations to  enhance overall customer experience, fostering loyalty and repeat business. An example of such a tool is Rosetta, an AI-powered tool which uses Machine Learning prediction functions to transform customer engagement. In this example, brands can use eCommerce recommendations which Rosetta generates to better understand customer needs and improve retention. Finally, a common example of consumer-centric recommendation engines is eds and improve retention. Finally, a common example of consumer-centric recommendation engines is Amazon’s data-informed predictive analytics function.

By leveraging Machine Learning algorithms to make relevant product recommendations to customers at checkout, Amazon’s cross-selling approaches target customers at a moment where they will be most likely to consider making an additional purchase. On a similar note, Amazon’s recommendation engines remember customers’ past purchases which also improve the quality of prediction for future purchase recommendations.

Pros and cons of Machine Learning: new opportunities and unforeseen impacts

1. Disruptions to the current labour market

Disruptions to the current labour market is the area of AI which critics of AI mention in debates about the cons of Machine Learning algorithms. Machine Learning has already created a new paradigm for working, bringing unforeseen benefits from automation. While there are concerns regarding automation, this disruption creates new avenues of employment. One such example is the automotive industry, where many manufacturers like General Motors are making the shift to EV production to meet greener standards for sustainability targets.

As vehicles transition from gas to electricity, there is ample opportunity for AI to empower the transition to a new paradigm. In a similar vein, AI will create a demand for jobs in other areas, including needing human hands to support the management of complex AI systems. Furthermore, AI will shape the creation of new jobs which address emerging technical problems related to the industries it affects. Customer service is an example of an industry where AI will harness Machine Learning to shift the way companies serve the needs of their customers, from the promotion of products to enabling customer loyalty and retention. In this area, there is the possibility for companies to use the capabilities of AI to automate product marketing campaigns and use data prediction to create a better product experience. 

Overall, the greatest challenge of integrating AI into the world of work will be supporting people’s transition into new roles that are in demand due to AI’s impact on industry demand shifts. 

2. Data security

Data security is another aspect of Machine Learning where disruptions to the current paradigm of digital interaction have led to changes in recent policy. An example of a resulting policy change is the 2016 General Data Protection Regulation (GDPR) legislation to protect people’s personal data within the European Union and European Economic Area. Moreover, in the State of California, U.S, the authorities introduced a consumer-specific data protection act, the California Consumer Privacy Act (CCPA). The CCPA demands that companies inform consumers about data collection of their Personally Identifiable Information (PII). 

Despite the cons of adapting to a new paradigm, where the impact of Machine Learning algorithms on data protection and privacy poses complex challenges, there are pros to the emergence of AI and  Machine Learning in this area. For example, while AI adds a layer of complexity and vulnerability to existing models of data security, there are equally several opportunities for using AI algorithms and prediction models to solve cybersecurity challenges. 

When used skilfully, Machine Learning algorithms can solve the same challenges that they pose to security. While cybercriminals could manipulate ChatGBT to their own ends and disrupt internal business systems, AI developers, businesses, and policymakers can collaborate to train a generation of new cybersecurity professionals to improve global cybersecurity infrastructure. In this way, a new sector of roles within the cybersecurity sector can emerge and become an attractive career path for problem-solvers, with an estimated global shortage of 4 million cybersecurity professionals. 

According to a 2023 article from the World Economic Forum, the African continent has the greatest demand for training to prevent threats from AI-related cybercrime. This statistic is important because it reflects the importance of cybersecurity as a global issue which has structural implications on local and international economies in a hyperconnected world. 

3. Bias, discrimination, and ethical grey areas

Bias, discrimination, and ethical grey areas are a third concern relating to the widespread use of Machine Learning algorithms in today’s world. The key concern with Machine Learning-related discrimination is that Machine Learning models can inherit bias from human influence and amplify it across major aspects of society [1]

Bias is particularly relevant in the area of supervised learning models, where AI professionals train Machine Learning models using labelled datasets, which presupposes a selection bias before the models have analysed the data. Another example of the cons of Machine Learning models is in recruitment hiring processes for transnational corporations. One study showed that Amazon had to discard an experimental Machine Learning tool used to screen applicants’ CVs and identify top talent, as the tool proved discriminatory, penalising CVs which included the word “women’s”, as in phrases like “women’s chess club captain”. 

Moreover, critics have further pointed out the murky waters of using Machine Learning models in hiring practices, including a lack of clear policy on how much data about a candidate an organisation can access. In another context, IBM discontinued its facial recognition and analysis products in the light of the risks of unethical applications of AI facial recognition tools for mass surveillance and racial profiling which violate basic human rights. 

On the other hand, some organisations support the argument that Machine Learning models enable their recruiters to get beyond the usual networks of candidates and appeal to a larger talent pool. For example, Goldman Sachs created a CV analysis tool to funnel candidates into the division where they would be a best fit. 

As is the case for data security challenges which Machine Learning algorithms pose, the potentially discriminatory functions of AI-supported hiring practices requires a collaboration between policymakers, business management systems, and researchers to ensure hiring practices remain as fair as possible. According to Reuters, the legal landscape is adapting with new legislation to these challenges, including the US Algorithmic Accountability Act and the Artificial Intelligence Act in the EU, providing a framework to ensure accountability and neutrality across AI applications. 

Conclusion 

Machine Learning Models are a fundamental aspect of Artificial Intelligence which can revolutionise the way humans interact in everyday life, automating a variety of functions. As we have explored in this article, the pros and cons of Machine Learning algorithms interact in a complex way, which means that lawyers, policymakers, and organisations are quickly adapting to the ethical grey areas posed by their impact on key areas of life, such as the labour market.

 Our takeaway from this analysis is that the impact of Machine Learning Models and their pros and cons for human domains is not limited to a specific area. Instead, the generalised implications of Machine Learning models for data security, workforce opportunities, security software for facial recognition, and social media algorithms mean that AI ethics and values are an essential topic for discussion and collaboration. It is therefore paramount that ethicists, researchers, and legal specialists team up to devise adequate legislation to regulate AI practices and harness their benefits.