Artificial intelligence is full of confusing terminology. Narrow intelligence, general intelligence, even long short-term memory. In this blog, we look at what Narrow AI is and give some examples.
Over the ages, many people have sought to define exactly what is meant by intelligence. My favorite definition comes from the Oxford English Dictionary:
Intelligence (noun). The ability to acquire and apply knowledge and skills.
This definition makes it clear that there are two distinct aspects of intelligence. Firstly, intelligence is about learning (acquiring) knowledge and skills. Secondly, it is about applying these to solve a problem.
General vs narrow intelligence
Human intelligence is described as general intelligence. That is, we are able to apply our existing knowledge to solve completely new or abstract problems. This is so innate that we often aren’t conscious of doing it. For centuries, we assumed this form of intelligence is what sets us apart from animals. But over recent decades, many studies have shown that animals are capable of some degree of general intelligence.
By contrast, narrow intelligence is only able to solve very specific problems. Where the problem fits with its learning, a narrow intelligence will solve it extremely well. But any problem that it hasn’t been explicitly trained to do is beyond its capabilities. This is the sort of intelligence demonstrated by dogs when they learn new skills.
AI or artificial intelligence refers to any software that demonstrates some degree of intelligence. Remember the definition above? According to that, software becomes intelligent if it is able to acquire knowledge and skills and apply these to a problem. In other words, learning lies at the heart of AI. Or to be more exact, machine learning lies at the heart of AI.
What is machine learning?
In machine learning, the computer is taught to recognize certain patterns in data. It then applies this learning to spot the pattern in new data. There are 3 forms of ML.
Supervised learning. Here you train the model with known, labeled data. E.g. you might show a computer thousands of labeled photos of animals and teach it to identify the ones that are cats.
Unsupervised learning. This time, you give the computer a set of data and it simply tries to identify any interesting patterns. Typically, you can use this to identify clusters of similar data.
Reinforcement learning. This approach is similar to humans learning from our mistakes. You give the computer unlabelled data, but each time it identifies something correctly, it is “rewarded”.
These 3 forms of machine learning are the basis of Narrow AI. Indeed, you will find one of these if you look into almost any system that is described as artificially intelligent.
Applications of narrow AI
In general, narrow AI can quickly learn to outperform a human at a given task. However, they can’t apply this learning to other tasks. Below we explore three common applications of narrow AI
Computer vision requires three distinct tasks. First, you need to segment the image. That means you need to work out which bits of the image relate to each other. For this, you can use a form of unsupervised machine learning, such as K-means clustering.
Next, you need to classify each region of the picture. One of the classic approaches is using a convolutional neural network (CNN). This is designed to replicate the way a human brain learns to recognize patterns. The final step is semantic segmentation. This involves working out how the bits of the image relate to one another and trying to actually understand what the image shows.
Natural language processing
Natural language processing (NLP) involves teaching a computer to recognize human (natural) language. This is a central requirement for a computer to pass Alan Turing’s famous Imitation Game (or Turing Test), making it one of the oldest disciplines in computer science. Nowadays, virtual assistants like Siri and Alexa have made NLP seem almost commonplace. However, it is only recently that computers have become so good at NLP.
NLP requires the computer to break each sentence into its constituent grammatical parts. It then uses a set of rules and grammar to understand how these parts are related. This allows it to extract the actual meaning in the sentence. However, getting this right is pretty hard since language relies on context and often uses idiom. Just consider these three sentences. “Time to get up!” “Rise and shine.” “Get your butt out of bed!”. These all mean it’s time to get out of bed. But it isn’t easy for a computer to understand that.
Machine learning is very good at recognizing patterns in data. Indeed, that is its main skill. As a result, it can be used for a whole range of forecasting applications. These vary from predicting when heavy machinery will fail or understanding future demand for electricity. Let’s look at a simple example to understand how forecasting works.
Say you manage a chain of clothing stores. You only have limited space to store stock, but you also don’t want to run out of anything in your stores (since empty shelves look bad). This means you need to forecast demand for products. This requires two things. Firstly, understanding past demand and how external factors affected this. Secondly, a model of future demand based on predictions and current data. Your model can take inputs from fashion magazines to predict fashion trends. It can use long-term weather patterns for seasonal effects and medium-range weather forecasts to check the weather. The model will output an estimate of demand for each type of product.
Creating good narrow AI models is a demanding job. Models can take months to create and verify. Even then, each model is only able to perform one task. SAIBRE changes that. It is a complete AI ecosystem that is designed to help us deliver end-to-end AI applications. With SAIBRE, it's extremely easy to reuse data sets between models. Models can be easily chained together to achieve more complex tasks. Once we've built your model it can be deployed in our infrastructure or in your own backend. And finally, our Smart Monitoring systems makes MLOps easy for anyone.
Practical applications of SAIBRE
SAIBRE's approach to AI can be applied to solve a huge range of business problems. These include:
Demand forecasting. For instance, forecasting electricity demand over the coming 8 hours, allowing utility companies to plan resources better.
Stock forecasting. SAIBRE can create models that predict how much stock you will need to hold and ensure you keep costs and waste to a minimum.
Elastic pricing. Dynamic price adjustments help you sell more products while minimizing your costs. This is the secret that made Amazon such a successful retailer.