The AI Technology Stack: Machine Learning, Deep Learning and beyond

In today’s tech-savvy world, the terms “Artificial Intelligence” (AI), “Machine Learning” (ML), and “Deep Learning” are buzzwords that often leave people scratching their heads. In the previous edition of Techsplained, we explored the foundation of AI and introduced the concepts of machine learning and deep learning through the lens of the AI evolution since the fifties.

These concepts may sound complex, but fundamentally, they’re designed to mimic human intelligence and solve problems more efficiently.

Imagine having a digital assistant that can understand your commands, predict your preferences, and even learn from its mistakes — that’s the power of AI!

This article will explore in further detail the technologies that are making AI possible today.

Machine Learning: Teaching Computers to Learn

Machine learning is like the super smart student in class who learns from examples and experiences. Rather than being programmed with instructions, ML algorithms learn from the data they receive to recognize patterns and make predictions and decisions without human instruction. Think of it as a fancy calculator that can solve complex equations by studying past solutions, rather than a calculator that’s been programmed with all possible equations.

Recommendation systems are one of the most popular use cases for ML. Did you ever think about how Netflix or Spotify suggest movies and songs that may interest you? That’s ML at work! These algorithms dissect your watch history or listening history and also incorporate data from millions of other people, attempting to tentatively produce content they think you may like.

Another fascinating use case is fraud detection in the financial sector. ML models can study patterns of legitimate and fraudulent transactions, enabling them to identify suspicious activities and flag them for further investigation.

The image below outlines the typical machine learning steps, which can be summarized in the 8-step process:

1. Gather and prepare relevant data

2. Clean and normalize the data

3. Select important features for model training

4. Choose appropriate ML algorithm

5. Train the model with prepared data.

6. Evaluate model performance

7. Fine-tune model parameters

8. Deploy model in production environment

Deep Learning: Mimicking the Human Brain

As an AI function, deep learning draws on a model of the brain to build its system. Similar to how our brain neurons connect and form intricate neuronal networks, in deep learning algorithms we use artificial neural networks instead of biological ones because using them further enables processing data by learning from experience.

Imagine a team of workers sorting objects based on their characteristics. In traditional Machine Learning, each worker would have a specific task, like identifying shapes or colors. However, in Deep Learning, the workers form a hierarchical network, with each layer building upon the previous one’s understanding. This allows the system to learn more abstract and complex representations, just like how our brain processes visual information.

A great impact area of deep learning is image and speech recognition, which some don’t consider AI, but when understanding its foundation, we know that the core lies in a deep neural network. Siri or perhaps Alexa sound like some virtual assistants that you might have used.These AI assistants used deep learning for better voice recognition as well.

The table below provides a visual comparison between traditional programming, machine learning, and deep learning.

Neural Networks: The Building Blocks of AI

Neural Networks are the backbone of both Machine Learning and Deep Learning.

They are based on the biological neural networks of our brain, which are interconnected nodes (artificial neurons) that process information and learn from data.

The pattern recognition work that neurons in the brain naturally do is something you can model, and so we have made headway by making programs called “neural networks” to employ this approach.

Training involves adjusting networks with data, requiring significant computer power and data for tasks like facial recognition.

Imagine a team of people trying to solve a puzzle collaboratively. Each person represents a node in the neural network, and their connections represent the weights or strengths of their relationships. As they work together, the team adjusts these weights based on their successes and failures, ultimately learning the most efficient way to solve the puzzle.

One real-world application of Neural Networks is self-driving cars. These vehicles use various sensors (cameras, radar, lidar) to perceive their surroundings and make decisions about steering, braking, and accelerating. Neural Networks process this data and learn from experience, enabling the car to navigate complex environments safely.

Large Language Models: Neural Networks at Scale

Large Language Models represent a leap forward in natural language processing, built upon the foundation of neural networks, particularly transformers.

Key Features:

• Massive Scale: Billions of parameters trained on vast amounts of text data
• Self-Attention Mechanism: Allows the model to weigh the importance of different words in context
• Transfer Learning: Pre-training on general language understanding, followed by fine-tuning for specific tasks

If a basic neural network is akin to a person learning a specific skill, an LLM is comparable to a polymath who has absorbed knowledge from millions of books and can apply this broad understanding to a wide range of tasks.

Conclusion:

While AI, Machine Learning, Deep Learning, and Neural Networks may seem like complex concepts, they’re ultimately tools designed to augment and enhance human capabilities.

As we’ve explored the how of artificial intelligence, from the foundational concepts of machine learning to the complexities of deep learning and large language models, it’s impossible not to marvel at the pace of innovation we’ve witnessed in the last decade.

The journey from simple neural networks to the sophisticated AI systems we interact with every day is nothing short of extraordinary. What’s driving this rapid evolution? The answer lies in the exponential growth of data and computing power. As we generate more data through our digital interactions, AI models become increasingly capable of understanding and mimicking human behaviour. It’s a virtuous cycle: the more we interact with AI, the smarter it becomes and the more useful it is in our daily lives.

Large Language Models, in particular, showcase the incredible potential of scale in AI. These models, trained on vast corpora of text, demonstrate an almost uncanny ability to understand and generate human-like text. They’re not just regurgitating information; they’re synthesizing knowledge in ways that can sometimes surprise even their creators.