Demystifying AI: Neural Networks

Welcome back to another edition of Demystifying AI! Building on last week’s exploration of how machines learn, this edition dives into neural networks – really the backbone of today’s most powerful AI systems, where computers begin to mimic the way the human brain thinks. By the end, you should have a solid understanding of what neural networks are, how they work, and why they’re revolutionizing everything from image recognition to language translation. Enjoy the read!

Neural Networks: The Foundation of Deep Learning

Neural networks are at the heart of deep learning, a branch of AI that enables machines to tackle complex tasks like recognizing faces, understanding speech, and even driving cars. You know… all the stuff that some people are freaking out about, because they think AI is taking over the world! 😉 Just as the internet is the infrastructure behind all of our online communications, neural networks are the structural foundation for deep learning systems.

Just as you would expect, for machines to fully understand and handle complex tasks like these, it takes an incredible amount of data. That’s why you need something that acts similarly to the human brain.

🧠 Interesting Factoid: Were you aware of just how much data the human brain can hold? It’s staggering – estimated at 2.5 million gigabytes (GB). If your average cell phone holds around 128 GB, that’s almost 20,000 cell phones' worth of storage in your head!

If you need a quick refresher on Deep Learning, please refer back to the April 11 edition of my newsletter!

Mimicking the Human Brain

Again, think of a neural network as a simplified digital version of the human brain. In our brains, billions of neurons (nerve cells) connect and communicate, allowing us to process information, learn, and make decisions. In much the same way, a neural network is made up of artificial “neurons” (also called nodes), which connect and pass signals to each other, enabling the system to learn from data.

The Basic Structure: Layers of Learning

A typical neural network consists of three main types of layers:

• Input Layer: This is where data enters the network. Each neuron in this layer represents a feature or variable from the input data (e.g., pixels in an image or words in a sentence).
• Hidden Layers: These are the “thinking” layers. They process data passed to them from the input layer (or from other hidden layers) through a series of mathematical calculations, extracting patterns and features at each step. Complex networks may have dozens – or even hundreds – of hidden layers… which helps further solidify the term “deep” learning.
• Output Layer: This layer produces the final result, such as a prediction or classification (e.g., determining whether an image is either a “cat” or “not a cat”, or determining whether to slam the brakes on a self-driving car).

Data <-> Signals: How Information Flows

Imagine data moving through the network like signals traveling along a (really long) series of wires. Each input (i.e., a signal) is passed from one layer to the next, where it gets transformed and combined in various ways to help decide what to do next with that signal. Ultimately, by the time the signal reaches the output layer, the network has made a decision or prediction based on all the processing that occurred along the way. Just like the human brain, this all happens really fast!

Weights and Biases: The Network’s Memory

At every connection between neurons, there’s a number called a “weight.” This weight determines how strongly one neuron’s output affects the next neuron. Think of weights as adjustable knobs that the network tunes while it is learning. Each neuron also has a “bias,” which acts like a baseline adjustment, helping the network make more flexible decisions.

✅ Very over-simplified example: Imagine getting driving directions from your favorite GPS application on your phone (e.g., Apple Maps, Google Maps, Waze). The “node” or connection point is essentially a piece of data showing where you are at a given point in time (e.g., 38.66818° N, 77.26778° W, which just happens to be a random spot on I-95 in Virginia). Other pieces of data, of course, include your destination location, speed limits on the roads, the current traffic, and your driving preferences. All along the way, your GPS is calculating the best route for you, based on “biases” like speed, safety, and the likelihood of a crash scene clearing before you get to it. Each of these calculations will also have a “weight,” which helps determine the priority of each decision in the process, so your GPS can give you a recommendation that gets you from Point A to Point B in the fastest AND safest way possible.

Activation Functions: Decision-Making Gates

After combining the inputs and weights, each neuron uses an activation function to decide whether to “fire” or not. I won’t go too deep here, but activation functions determine the output of a particular neuron, to decide whether it should be considered in the overall decision/recommendation. If it is determined to be important enough in the decision, the neuron activates, and this influences where the “signal” goes next through the other layers and neurons. Suffice it to say, if we didn’t have these activation functions, neural networks would really be limited to solving only simple problems.

Learning by Adjusting

Neural networks “learn” by adjusting their weights and biases. During training, the network makes predictions, compares them to the “correct” answers, and tweaks the weights and biases to reduce errors. It’s a continuous loop. This process, often called backpropagation, is repeated over and over, potentially millions of times, until the network becomes highly skilled and “nails” the task every time.

Another Example: Image Recognition

Let’s say you want to use a neural network to recognize handwritten digits from a scanned document like an invoice. Of course, all of us write numbers exactly the same way, with identical handwriting, so this is really easy, right? Yeah…

• The input layer receives pixel values from an image.
• Early hidden layers might start to detect basic shapes like lines or curves, which start to form the shape of certain numbers – if the shape is curvy, for example, the number is more likely to be something like a 3 or an 8, and less likely to be a 1 or a 7.
• Deeper layers combine these simple features to recognize more complex patterns – they may see a loop in a 2 if someone writes it that way, or lines intersecting in a 4.
• Ultimately, the output layer confidently predicts the digit in the image. This ability to automatically extract and combine features is what makes neural networks so powerful for tasks like image and speech recognition.

[Side note: this is one of the many things my company does when we’re doing intelligent automation and processing images and handwritten documents!]

Challenges: Overfitting and Data Hunger

Neural networks can be so flexible that they sometimes “memorize” the training data instead of learning general patterns – a problem called overfitting. If you recall, I mentioned overfitting a couple weeks ago. It’s similar to when a student memorizes the answers for a test question based only on the key terms; but if the question is worded just a bit differently, the correct answer may be “B” instead of “C.” To avoid this problem in the neural network world, machines need vast and diverse datasets. Gathering and labeling enough data can be a significant challenge, especially for highly specialized applications. Which leads us to…

Power and Cost: Deep Networks Are Expensive

Deep neural networks can solve incredibly complex problems, but this power comes at a high price. As we all know, everything has a price. I had an economics professor back in grad school who liked to use the term “TANSTAAFL” (pronounced “tans-TAFF-el”). It stood for “There ain’t no such thing as a free lunch.” 😊

Training neural networks requires enormous amounts of computing power and energy, as well as access to extremely large datasets. This is why companies invest in specialized hardware and cloud computing resources to build and deploy cutting-edge AI models. If you ever have the opportunity to tour a cloud data center that supports AI operations and these massive amounts of data, you should definitely do it. I got to see some of these data centers while I was with AWS, and they are fascinating!

Summary – 3 Key Points:

➡️ Neural networks are digital “brains” that learn from data by mimicking how the human brain processes information.

➡️ Neural networks use a layered structure, adjustable weights, and activation functions to allow them to handle complex tasks that once seemed impossible for machines.

➡️ These networks require massive amounts of data and computing power, but their ability to extract features and learn complex patterns is driving the AI revolution forward.

Coming Up:

If you’ve been reading my newsletter over the past 6 weeks, you’ll know that I frequently talk about the fact that AI is “all about the data.” Well, next week, we’ll dive into just that – the data! Data is the fuel that flies this AI fighter jet. We’ll talk about structured and unstructured data, data quality, and ways to collect this massive amount of data. We’ll also discuss the privacy and compliance considerations around data usage, because yes, I am a security and compliance guy at heart. Stay tuned!

Thanks for reading, and as always, feel free to share your thoughts or questions. See you next week!

🚀 What aspects of AI intrigue you the most? What aspects scare you the most? Drop a comment in the thread!

🚀 Follow me and subscribe to the newsletter for more weekly insights into AI, digital transformation, and cybersecurity.

🚀 Book a meeting with me: I'd love to chat more about AI and intelligent automation, and explore ways to collaborate – no pressure!

#ArtificialIntelligence #IntelligentAutomation #DigitalTransformation