Introduction
Artificial Intelligence enjoys quite a range of applications and capabilities. This module will explore their role and impact on business. The current crop of AI applications is causing business to take notice and deploy them to generate a competitive edge for their organisation and a better experience for their customers. AI has revolutionised and will continue to revolutionise not only business and how it operates, but our entire society.
In December 2022, the McKinsey Global Survey on AI (Chui et al. 2022) [^1] published the results of their survey of the respondents who indicated that AI is used in at least one of their products or services. The results are below:
Figure 1: McKinsey Global Survey results
According to Chui et al. (2022) [^1], the average growth in capabilities in AI has grown from1.9 in 2018 to 3.8 in 2022, or 100% growth. According to McKinsey (Chui et al. 2022) [^1], there are four groups of AI capabilities:
- Machine learning
- Language models
- Computer vision and
- Robotics
Bean (2019) cites Thomas and his description of the AI Ladder for business. The AI Ladder, according to Thomas, involves five steps:
- Start with the business problem that you are attempting to address
- Understand your data requirements – these are the foundation for AI success
- Develop the right skills to leverage AI capabilities
- Focus on algorithmic trust and data integrity to ensure credibility
- Recognize the need for cultural and business model change.
This ladder would appear very similar to the steps involved in strategic planning. It is the author’s opinion however that it lacks a step to define the intended outcome of the process.
Bean (2019) [^2] also suggests that many organisations are leading on the front foot with the establishment of labs and centres of excellence. Although this may well be the case for large corporations, SMEs will have to rely on the trickle down of the technology as they generally have neither the funding or expertise to indulge in such research and development.
According to a report by NewVantage Partners (2022) [^3], the number of AI initiatives rose from 12.1% in 2021 to 26% in 2022, a considerable increase. The advent of ChatGPT in November of 2022 may well see a further jump in initiatives as the public and business now become more aware of the general potential of AI and its ability to potentially improve customer experiences.
Machine Learning
AI, like some many other things enjoys a number of definitions of what Machine Learning is exactly. Fagella (2020) [^4]
“machine learning is the science of getting computers to learn and act like humans do, and improve their learning over time in autonomous fashion, by feeding them data and information in the form of observations and real-world interactions.”
The algorithms used by machine learning provide computers with the ability to perform tasks previously considered unimaginable for humans at the same level of speed. According to Heath (2018) [^5], the rapid advancement of these applications is allowing software to “make sense of the messy and unpredictable real world”.
The example of Covid 19 is valuable in that so much data was collected from around the world and processed through machine learning models to provide predictions and information on the virus. In turn, this informed health organisations and government with the data on which they could base decisions on policy for their communities. Without the speed with which this data became available, implementation measures could have taken much longer than what they did.
It is important to understand the difference between machine learning and traditional software programming. Traditional programming uses defined values in the program to complete defined tasks. Machine learning on the other hand actually learns to discern one thing from another through the use of large datasets. Consequently, it learns the difference between a tree and a cat or the image of a pedestrian compared to an open road ahead for a self driving car.
All of this is still dependent on the quality of the data used in the training of the model which is created by humans. As such, there is still bias in data and erroneous data can be fed to the model which will in turn affect its output.
Supervised Learning
This is the most common form of training an LLM. The model is fed data that has already been structured, categorised and labelled by humans. Consider this to be a two part concept where x is either equal to or unequal to y. Therefore, the model learns through examples in the data.
An article by Bajpai (2020) [^6] relates a most interesting use case of AI when applied to Covid. According to the article, Alibaba’s DAMO Academy (Academy of Discovery, Adventure, Momentum and Outlook) was tasked with the role of finding a way to speed up the diagnosis of Covid. They developed an AI model that could diagnose Covid within 20 to 30 seconds, and at 96% accuracy, compared with a trained doctor going through 300 – 400 scans who would take around 10 – 15 minutes. Such an increase in efficiency may well have saved many lives and released doctors for treatment purposes instead of reading scans.
With supervised learning, there are essentially two types which are the classification type as described above, and then there is regression. Regression involves comparing two variables to establish the relationship between them and then use the result for predictions and forecasts.
The bottom line with supervised learning is that the more quality data used to train the model, the greater the level of accuracy. In the Covid example above, it may have been possible to obtain 100% accuracy. The cost benefit analysis then needs to be considered for perhaps a 1% gain in accuracy. Such a debate will be case dependent and no doubt ethics would also be involved.
Unsupervised Learning
These algorithms are not based on providing examples to the model. Instead, the model is tasked with identifying patterns from the data it receives. If the model was provided with images of a million types of shoes, it would sort the shoes by comparing the size and features of each image and then sort them into groups. This is very similar to the manner in which humans think according to LeCun et al. (2015) [^7], “Human and animal learning is largely unsupervised: we discover the structure of the world by observing it, not by being told the name of every object.”. This statement is considered to be an important definition of the relationship between human and machine learning.
Examples of unsupervised learning would include fraud detection in banking where normal transactions are scanned by the thousands so that unusual activity is flagged. Likewise, this process could also be useful in cybersecurity where any unusual activity on the network is immediately flagged for response by the security team. As machines increase in power and sophistication, this model of learning will potentially grow exponentially.
Semi-supervised learning
The tagging and labelling of data by humans is both time and cost intensive. Semi-supervised learning utilises the concept of training a model on a limited amount of structured data to establish its initial parameters and the allow it to analyse a much larger pool of data and assign its own labels based on the initial dataset. This process is known as psuedo-labelling.
Reinforcement learning
Reinforcement learning is the creation of models used for goal achievement or outcome. The machine plans which moves or actions will be best to achieve its goal. As various strategies are applied and failed, it reinforces to the model that this was an incorrect choice. It will then find the optimal action and “remember” this.
An extension of reinforcement learning is reinforcement learning with human feedback (RLHF). This type of learning is collaborative between the machine and a human where the human provides the machine with feedback on its actions. This type of learning builds human concepts, experience and processes into the machine learning process.
Language Models
A language model is a statistical dataset fed to a machine that learns the patterns and structures from the content. It can then use this data for predictive purposes and generate understandable language in its response to prompts. Language models consider word frequency, grammar rules and semantic relationships to learn and predict the words in a sentence given the context..
Natural Language Processing (NLP)
NLP is a branch of language processing that uses concepts from language and linguistics.. This can be in either the written or spoken form. NLP is complex in terms of a task because of the number of nuances and context surrounding human language.
A challenge around NLP is the understanding of context and emotion. Language is often variable with different meanings for words depending on context.. You also have issues of vernacular where words that are native to particular areas have specific meaning.
NLP has extremely useful application in the area of chatbots, summaries of text, virtual assistants and more. The increased usage of these types of applications will continue to grow as the technology advances.
Large Language Models (LLMs)
LLMs have revolutionised NLP due to their immense size. It is LLMs that have allowed the development of ChatGPT and progressing to GPT4 which is now available. No doubt further iterations are already in process.
LLMs are trained on huge amounts of data, however, there is no clear evidence on the number of parameters used. According to Bastian (2023), GPT3 was trained on 175 billion parameters and GPT4 on 1.76 trillion parameters. Even on those figures, that is a tenfold increase in parameters.
One aspect of LLMs is that they require large amounts of computing power to both train the models and then run them. This puts them out of the reach of small business of course, however, they can often benefit by using a SaaS provider who offers access. LLMs are going from strength to strength with technological advances which can only improve their abilities. A part of this advacnce will be to address the elimination or at least reduction of bias in the models and a greater level of awareness around ethics of deployment.
Computer Vision
Another branch of AI whereby algorithms are developed that allow computers to consider videos and images and then interpret them. The goal of computer vision is to replicate a human’s ability to see and comprehend the world though vision.
Object Recognition
A relatively standard task where a machine identifies and classifies the objects that are embedded in images and videos. The application of this is wide reaching in robotics on assembly lines, self driving cars and security systems.
Image Classification
Classification of images is extremely useful in such applications as search for example. You could search a database for photos of birds, or landscapes for example and have the AI return images matching the search parameters. Medical imaging is another area where the presence of objects that should not exist in a healthy organ are identified.
Facial Recognition
Security systems are already in use, such as your smartphone recognizing you.. Other examples include uploading a copy of your passport to the internet and a screenshot of your face so the AI on the other end can match the two photos for identification purpsoses.
Augmented Reality
Computer vision techniques are also applied in the area of augmented reality too. Virtual objects can be overlaid or superimposed onto the real world. This can create a fully integrated and immersive experience for the user. Apple’s Vision Pro headset is an example (Ritsos & Butcher 2023) [^8].
Robotics
Robots are machines that are designed to perform tasks by executing a series of movements. Robots can free humans from mundane tasks. The greater the advances in AI and robotics, the more complex tasks can be delegated to robots. A self driving car, for example, is nothing more than a very advanced robot fulfilling a role traditionally done by humans, or at least with human assistance.
Robotic organisations are working towards situations where robots can learn from their environment. This would involve the integration of artificial intelligence and machine learning techniques. Imagine, as an example, a robotic vacuum cleaner that moves all over the house bumping into walls and furniture. Over time, it develops a map of the house and all the obstacles and its performance improves. This is reinforcement learning integrated with robotics.
Conclusion
This concludes our discussion on the four most common types of aritificial intelligence capabilities. Discussion has been had around the development of the different types of AI and the enormous potential it has in the workplace, society and building competitive edge for business.
