AI Unboxed Series: Machine Learning 101

Have you ever felt lost in the sea of AI jargon and tech terms? Us too! Our Senior Investment Analyst, Lucy Law is someone trying to make sense of the complex world of artificial intelligence. Lucy understands how confusing it can be, and has embarked on this journey to demystify AI, and with the hope of encouraging discussions within our start-up ecosystem to shape our collective understanding of AI investments. 

This AI Unboxed blog series aims to break down the complexities of AI into digestible chunks, making it accessible to everyone. Together, we'll unravel the mysteries of AI and explore its potential impact on various industries.  

Series Overview: 
Unfolding over four content releases ...

1.   Machine Learning 101:  
     Understanding the fundamentals of machine learning and its applications. 
2.   AI applications + Refining investment thinking:  
     Exploring different AI applications and their implications for investment strategies. 
3.   Common tools for building AI: 
     A guide to popular tools and frameworks for developing AI solutions. 
4.   AI and Anxiety: Navigating ethics, governance, and security:  
     Delving into the ethical considerations and security challenges posed by AI technologies. 

MACHINE LEARNING 101 

What is AI? 

There is no one official definition of AI because it's a broad field with many perspectives. In general, AI refers to machines / systems that can do tasks humans do, like learning, reasoning, and problem-solving. 

It's worth noting that AI can be classified into two main categories: Narrow AI and General AI. Narrow AI, which you have today, is designed for specific tasks, while General AI aims to mimic human intelligence more closely. To keep things simple, we'll use AI and machine learning interchangeably, while recognising that machine learning (ML) is a subset of AI focused on enabling machines to learn from data and improve over time. 

For those curious to learn more, check out the free course "AI for Everyone" by Andrew Ng, a respected figure in the field. 

Now, time to jump into the foundational AI models! 

Foundational AI models: 

The AI models can be categorised into three main groups; 

• 1. Supervised Learning,
• 2. Unsupervised Learning,
• 3. Reinforcement Learning. 

The Venn diagram below, attempts to visualise the intersection of ML models, acknowledging the complex and evolving landscape. 

1.  Supervised Learning:

Algorithms trained on labelled data used for classification and regression tasks (aka predictive modeling). For example, a set of photos containing cat images are labelled as cat and are used to teach the machine to recognise cat photos.

Deep Learning:

A subset of ML using neural networks with multiple layers, effective in processing unstructured data like images and text. Neural networks are like neurons in your brain, designed to learn and make decisions by analysing patterns.

> Convolutional Neural Network (CNNs):
Imagine the CNN model is like a detective using a magnifying glass to look into images patch by patch, and then find a specific crime pattern and combinations of patterns. These are commonly used for tasks like image classification, object detection, and facial recognition (like the earlier cat example).
e.g. E-gates at airports, Waymo, Tesla Autopilot.

Here is a video explaining CNNs in simple way:

> Transformers and Recurrent Neural Networks (RNNs):
Imagine this model is like someone baking a new recipe with several steps. As it proceeds, each step depends on completing the one before it. You can’t bake a cake before mixing the ingredients. The transformers and RNNs operate in a similar manner. These are utilised for sequential data processing, such as language translation, sentiment analysis, and chatbots.
e.g. Google Translate, Open AI ChatGPT.

Here is a video explaining transformers:

Traditional Machine Learning:

Generally, most other algorithms outside neural networks would be classed in this category.
e.g. Email spam filtering, fraud detection systems in financial institutes.

2.  Unsupervised Learning:

Algorithms that work with unlabeled data to find patterns or groupings.
e.g. Detecting “abnormal” money transactions and flagging the potential fraud.

3.  Reinforcement Learning:

Algorithms that learn to make decisions by receiving rewards or penalties for actions taken. Think of it as training your dog. You give treats for good behaviour and skip the treats for the not-so-good actions. This “points” system teaches your dog what you are looking for, steering them towards better behaviour with a mix of rewards and penalties over time. 
e.g. AlphaGo, Spotify song recommendation, DeepMind AlphaStar

For a practical illustration, check out this interesting video from a US software engineer Peter Whidden who trained reinforcement learning model to play Pokémon Red!

While Supervised Learning dominates many AI applications, all models can potentially handle different types of data, each excelling in specific scenarios. We will dive deeper into the applications and nuances of these AI models in the next content piece of this series; AI applications + Refining investment thinking.

Before we wrap up this Machine Learning 101 content piece...

Please be aware that we may not cover every single model and detail here. With the AI field evolving so rapidly—almost by the week—we expect this classification and map to look quite different in the near future. Check back early next month for our AI applications + Refining investment thinking piece.

We welcome your feedback, insights, and suggestions on content relating to this topic as we navigate this ever-changing landscape together!  If you have any questions or thoughts to share, please get in touch with Lucy as she has a particular interest in this sector and would be happy to chat: lucy.law@nzgcp.co.nz