How Machine Learning Works
First, here’s a (very) basic look at how this works. What we’re talking about here is machine learning, a subcategory of AI. You can read my previous article on Machine Learning here.
The building blocks of machine learning are called perceptrons and neurons. We will use perceptrons and neurons synonymously in this article. However, you can think of neurons as complex perceptrons. In layman’s terms, a perceptron algorithm is used to create a line that is used for classification of entities. A single-layer perceptron accepts a number of inputs, applies an assigned weight, and then creates (fires) an output — either a zero or a one, which can be thought of its entity A classified as “insert classification here” “yes” or “no” respectively.
To understand the concepts better, let’s use a basic example: Given a 10×10 grayscale png picture of a single handwritten digit, how can we detect what figure has been written?
Obviously, the picture of the digit could contain any number from 0-9. For a human, all it takes is a quick glance at the image and within milliseconds, your brain’s neural networks magic and recognizes the digit (as long as it’s not a doctor’s handwriting).
This is a bit more of a complex problem for a computer algorithm. One method to achieve your desired results would to first break the image down to the pixel level, each pixel representing a perceptron in the network. Thus the first layer, called the input layer, would be represented by 100 perceptrons. The input would be zero for black, and one is white. Each perceptron would be given a weight. For example, the pixel in the top left corner of the image would most likely have a much smaller weight than a pixel near the center of the image. The higher the weight, the more important the neuron.
Each perceptron would also be assigned a threshold, which allows it to fire. For example, a neuron with a threshold of 10 will be more difficult to fire off than a neuron with a threshold of 5. Once your neural network is set up, you will need to feed it training data. For this example, there is a plethora of training data available.
The training data is then plugged into the system, and its hidden layer (the layer between the input and output layers that does all the heavy lifting) is adjusted by modifying the weights and thresholds to provide a maximum success rate.
This involves some intense math (for me at least) including gradient descent and other methods.
The output layer will consist of 10 neurons, a neuron for zero, a neuron for 1, all the way up to 9. And thus your neural network is complete.
That is a very simplified explanation, but gives you some idea of how machines begin this work. The amazing part is that as the system identifies and processes data, it will adjust itself and make adjustments as it goes, improving the overall success rate.
Think of that like “learning as you go,” just as a human does. Some of these systems can be trained to achieve close to 100 percent accuracy; incorrect items are typically due to unreadable handwriting in our example above.
One stunning example of this is an AI system using deep learning defeated a human in a game called Defense of the Ancients 2 (DOTA 2). While machines have surpassed people in other games such as chess and Go, those have a limited number of variables. This game contains an infinite number of possibilities.
Plus, it’s not a turn-based game. In real-time, players manage dozens of characters, all with different abilities, compete to collect items and control territory. The system, built by the Elon Musk-backed nonprofit, Open AI, demonstrates it’s possible to build systems that can be trained to outperform humans at infinitely complex tasks. I implore you to read the details on this project.
The Cost Factor Machine learning is a new, exciting field, and we’re just tapping into the potential. Still, SMBs and even some enterprises may wonder where they fit into all this. After all, leveraging AI uses the latest technology.
While standard computer functions operate on CPUs (central processing units), this type of computing requires graphics processing units or GPUs. These were initially designed by gaming companies such as Nvidia and ATI to process the functions and graphics in video games.
While you’re paying a premium for that type of power and the people who can create it, the return on investment makes the cost irrelevant. Many companies will eventually supplement AI in exchange for some human labor, especially teams that work in forecasts and predictions. If you paid a person to sit down and study 30 years of health data, he or she might take months or even years to do that work. A human must sleep and may make mistakes.
Meanwhile, machines will find hidden patterns we can’t see and won’t overlook anything. On top of that, they’re always learning and improving. The person who played DOTA2 against AI spent years training on the game. The machine trained for one month.
Imagine if we set AI systems to train and study the world’s problems for years. They are going to discover things humans never would.
The Data Factor Another challenge for businesses: having enough data. To apply AI, you need massive data sets, with hundreds of thousands (or millions) of points. The larger the data set, the better it is for the machine so it can learn.
You might find some data within your organization, information about your customers or clients, for example. Analyzing that may benefit a company in many ways, from sales and marketing to improving your product or service.
Companies can also rely on public data sets. Governments, non-governmental organizations (NGOs), nonprofits, and others publish plenty of data about our world and its people. This includes crime data for your neighborhood, Census data, financial information, even the post office. A law firm might analyze local crime data, for example.
Employee data is also useful. Consider these ideas:
● Determine people’s most productive times so you can schedule meetings that don’t interfere with people’s best working hours.
● In manufacturing, tracking production per hour and by day to determine high/low productivity and using that to predict when problems may occur.
● In construction, using the probability of sick days and inclement weather to predict a more accurate project timeline.
AI can make companies more efficient and reduce costs and improve customer service. Staying on top of this trend will make your business more competitive in an increasingly tech-driven world.
Each company’s need or use for studying data is different, which is why we and others aren’t offering a specific product. There is no one-size-fits-all option, but contact us about custom solutions that can improve your business.
Remember, 20 years ago, the internet wasn’t mainstream. Where will your company be 20 years from now? AI is the future; embrace it.
Got A New Project?
Book a meeting with one of our team member or get ball park estimation on your project.