Understanding Convolutional Neural Networks (CNNs)

One of the techniques I learned when writing “Deming’s Journey to Profound Knowledge” is to create two target readers from the beginning. Most advice on this subject says you should pick one target reader. However, I aspire to write books like Michael Lewis, and I think this technique will get me closer. For example, in Profound, I needed to try to explain Dr. Deming’s physics background and how it related to management theories. My two target readers for "Profound" were Ben Rockwood and my mother-in-law Dixie.? Ben is an operational research expert, and Dixie has no technical background but reads 2 to 3 books monthly. My "white whale" in Profound was trying to explain "Schr?dinger's cat" to them, and both enjoyed the story. In my new book “Rebels of Reason - The Heroes of ChatGPT and Modern AI,” I need to explain MNIST or, more generically, neural networks - my new "white whale" for Dixie and target two, Joshep Enochs. However, I thought I’d let some LinkedIn followers tag along. Below are some of my first cut notes. I’d love to hear any comments or thoughts.

Convolutional Neural Networks, or CNNs, are specialized types of neural networks, particularly well-suited for analyzing visual data like images. They’re structured like a digital brain, containing layers of interconnected artificial neurons that work together to identify patterns and recognize objects.

Let’s explore CNNs by imagining one trained to recognize handwritten numbers. When you show this network a picture of a handwritten "2," it doesn’t simply analyze each pixel in isolation. Instead, CNNs use convolution, which involves scanning small, overlapping sections of the image called filters or kernels. Each filter detects specific features, such as edges or curves, forming the foundational elements of the image.

In the case of our handwritten "2," the CNN might first detect more minor features, like lines or corners, in the initial layers. As the image passes through multiple layers, the network combines these details into more complex shapes, ultimately recognizing the "2" as a whole.

An essential part of CNNs is their pooling layers, which reduce the image’s resolution while preserving key features. This optimization helps the network focus on the most crucial information, making it faster and more efficient. Finally, fully connected layers at the end compile all these learned features and assign probabilities to each possible outcome—in our example, numbers from 0 to 9.

Training a CNN involves a process similar to teaching a person to recognize numbers, incorporating a concept called backpropagation. Initially, the network’s connections are random, and when it makes a mistake, backpropagation adjusts these connections.

Random numbers are used to initialize the connections in neural networks to ensure that the model doesn’t start with any inherent bias. By beginning with random values, each neuron in the network learns independently, rather than starting with identical values that could cause the neurons to learn the same features. This randomness allows the network to explore diverse patterns during training, which leads to a more balanced and robust learning process as the model gradually adjusts and fine-tunes these initial values through backpropagation.

This process calculates how wrong the network's guess was, then fine-tunes the internal filters by strengthening connections that led to correct identifications and weakening those that didn’t. Over time, as the network learns from its mistakes, it refines its ability to recognize patterns, developing an “intuition” to accurately predict a "2" when it sees one in the future.