Getting to grips with gripes

‘AI’, ‘Supervised Learning’, ‘Machine Learning’… Phrases bandied around as flavor of the day: shots of AI in any software offering seems indispensable, despite sometimes referring to little more than filtering, or worse, black boxes churning out outputs as by magic.

A glance at the popularity of ‘Artificial Intelligence’ as a search query in Google since 2014 is a clear testimony of generated interest.

Google Trends

Source: Google Trends (2014-2019). 100 is the peak popularity for the term. 50 means that the term is half as popular

So misused are these terms that Apple’s Siri, and Google’s DeepMind’s developments are all frequently described as Artificial Intelligence, often used interchangeably with ‘Machine learning’ and other buzzwords.

How do you like your AI?

However, the financial industry amongst others, is under pressure to roll out such key technologies as AI, Machine Learning, and Natural Language Processing into products and outputs. Ubiquitous in our daily lives, from Alexa to our Spotify playlists, from Internet of Things (IoT) to WealthTech, it can be challenging for the laymen among us to tell apart the impostors from the geniuses working on cutting edge developments across these new tools. Relating to new concepts can be as tough as adopting new technology, but embracing the innovation is crucial for anyone willing to benefit from new advancements.

At HUBX, we are proud of our in-house expertise and our data-science team love of exacting standards. In our ongoing dialogue with clients, we always discuss our AI projects aiming to share our expertise. This being our initial AI-related post, we cover some basic concepts, in digestible, bite-size formats which will serve as reference for our upcoming, more complex publications.

Going back in time: AI wasn’t born yesterday

Earliest notion of AI we owe to the Greeks and ‘Automatons’ – self-operating metal robots created by the Greek God Hephaestus. However, the term ‘Artificial Intelligence’ was coined by John McCarthy, American computer scientist at the Dartmouth Summer Research Project on Artificial Intelligence in 1956.


“The question of whether a computer can think is no more interesting than the question of whether a submarine can swim.”


Edsger W. Dijkstra, Dutch Computer Scientist (1930 – 2002)

The operative word is ‘think’. Is ‘thinking machine’ equivalent to ‘human thinking’? In 1980 the philosopher John Searle has introduced notions of Weak AI and Strong AI in his thought experiment called The Chinese Room Argument, which aimed to answer the question.

Flexing its repetitive muscle: Weak Artificial Intelligence

Weak AI pertains a situation when a computer simulates thinking solely focused around a restricted task. Examples of applied Weak AI range from AlphaGo – a computer program – winning at Go against a human player or Apple’s Siri and Amazon’s Alexa, all based on pre-defined algorithms.

Taking over our souls? Not just yet. Strong Artificial Intelligence

Strong AI is usually referred to as AGI – Artificial General Intelligence. AGI as the Holy Grail of all artificial intelligence, is still the stuff of dreams and science-fiction, as well as Future Studies. AGI would think and benefits from ‘consciousness’ as humans do thanks to the combination of its cognitive ability and computing skills, thereby matching human performance. PathNet is the first step towards AGI.

Muscle at work: Definition of Machine Learning

Machine Learning has eaten up a large chunk of the AI research and development bandwidth; it refers to “the concept that a computer program can learn and adapt to new data without human interference” (Frankenfield, J. 2018).

Supervised and Unsupervised Learning, as well as Reinforcement Learning algorithms are three key forms of Machine Learning.

Subsets of Machine Learning
Subsets of Machine Learning

Food for thought: Supervised Learning

By feeding it large quantities of labelled training data and correct answers, algorithm learns patterns. Should algorithms produce incorrect results, i.e. completely different from the expected output, training data becomes crucial in guiding the algorithm back on the right path. As an example, let’s use a situation we have all encountered: Google’s email spam filter. The filter was created by feeding it spam and non-spam emails which have been pre-identified. As new emails come in, the model leverages its learnings from its previous meal of tagged emails to decide how to classify the new ones.

email spam filter
Email spam filter

The pattern detective: Unsupervised Learning

In an Unsupervised Learning scenario, data is without labels and the algorithm seeks to find patterns and extract important features from the data set, without the help of a human teacher. Unsupervised Learning is commonly used to cluster similar training data or detect patterns such as credit card fraud. It is important to note that no indication of truth is fed to the algorithm in the training data, which makes evaluating accuracy much more challenging.

The reward seeker: Reinforced Learning

Reinforced Learning rewards algorithms for steps taken towards an overall goal using predefined reward functions. Constantly seeking new ways to obtain rewards, algorithms then uses feedback to take new, better decisions. Reinforced Learning is widely used in the training of autonomous vehicles.

Do you speak my language? Natural Language Processing

Weak AI also covers Natural Language Processing (NLP): NLP investigates ways for computers to understand human language by using a two-layer neural network converting each original word into unique vectors within a vector space. Similar words to the originals (such as ‘car’, derived from ‘automobile’) are then identified as such in the vector space.

cosine similarity values with directions for different words

Cosine similarity values with directions for different words

Cosine similarity measures the cosine angle between the vectors, or the relationship between two words. Simply put, the closer the value is to 1, the more closely related the words.

Vectors in action

NLP also looks for extended relationships between words based on our existing knowledge of one relationship. Using vector addition on top of the existing connections will point to other relationships.

vector space of countries and capitals

Vector space of countries and capitals

China’s capital city is Beijing. In order to deduce Spain’s capital, using ‘China – Spain + Beijing’ as an equation will result in ‘Madrid’.

Understanding the Sentiments

Finally, not forgetting Sentiment Analysis: chatbots classification of a consumer’s message based on its tone. Negative sentiments (such as anger and displeasure) are assigned a negative value which ranges from 0 to -1, where -1 represents maximum displeasure. The opposite method is used for positive sentiments. Empowered with this knowledge, it is tempting to aim for a high negative, or high positive score when hoping for a human response from a customer service advisor!

Conclusion

AI is in its infancy, despite an already impressive number of real-world applications.

Financial services, healthcare, automotive, retail, oil & gas, and transportation are all benefiting from Machine Learning R&D. Financial services, for example, has evolved significantly through the application of process automation, fraud detection, financial monitoring, algorithmic trading, robo-advisory, underwriting and credit scoring.

Learning about AI in its various forms is essential to decode the extent of services on offer, their quality, and veracity of their claims.