Neural networks can generate content, improve product features, help scientists conduct experiments, and perform many other tasks. They work much faster than humans and effortlessly process huge amounts of data. They can operate 24/7 and don’t require any social benefits. But it's not all so clear-cut.
Will neural networks fully replace people? Let’s find out!
Neural network: what is it?
If you watch Netflix, you probably love its recommendations. The streaming service knows what you enjoy watching and helps you find top-notch content in your favorite genre. Have you ever wondered how this system works? It seems to rely on complex algorithms. Well, neural networks work to enhance them. You probably aren’t even aware of this technology. But, you regularly come into contact with it.
A neural network operates as a digital analog of the human brain. Our brain consists of cells, called neurons. Thanks to them, we can memorize facts, think, and make decisions. The small elements that a neural network consists of are called nodes. A node is a place where computation happens. A node to a neural network is what a neuron is to a brain.
A neural network is a subset of artificial intelligence, but the two are not the same thing. AI is the research area of computation applied to thought. Supervised learning and neural networks specifically are a particularly successful subset of it.
Today, people already use AI in some sense. It is capable of doing some pretty advanced things, in the right circumstances. There is a chance that, over time, AI might be able to develop a consciousness like the one that humans possess (although, scientists are split on that). If AI achieves this goal, it will turn into AGI — artificial general intelligence, like the character Data from Star Trek. Some people believe that large language models already serve as examples of AGI today. Yet, in fact, they only mimic AGI’s manner of communication with humans.
Neural networks, meanwhile, lack such ambitions. They were invented to recognize patterns in data and help people make decisions, based on these patterns. Before neural networks came to life, people used simple linear regression and other supervised learning techniques to make predictions based on previous observations. Compared to these techniques, neural networks can take a lot more parameters into account and thus are considered a breakthrough.
Once a neural network is trained to achieve a task, it can do it much more reliably and rapidly than a human. For example, it can write a sonnet in a second or label millions of pictures based on which ones have cats and which have dogs.
A neural network consists of at least three layers:
- Input. At this stage, it gets a set of data to process — for example, the medical records of all the children who were born in the area where an environmental disaster broke out.
- Hidden. The network processes this information — for instance, singles out the children who had issues with their thyroid glands. Actually, the hidden stage can consist of any arbitrary number of layers, performing various transformations and computations. These layers define the overall architecture of the network.
- Output. The network transforms this information into a comprehensive format for humans, such as spreadsheets and graphs.
The researchers will be able to compare the statistics from that particular time and area with the national or global average. It will help them analyze the impact of the disaster on the babies’ thyroid glands. Such an approach is much more precise and efficient than checking every medical record and organizing the data manually.
How can people use neural networks?
The scope of the capabilities of neural networks keeps expanding. Here are a few examples of tasks they can perform — and that you most likely know about:
The human brain is versatile. It can handle a wide variety of tasks: recognize objects, compose texts, tell the truth from a lie, calculate, compare objects, prioritize things, and so on. Neural networks specialize in their respective niches. For example, Amazon Go’s neural network scans the content of the basket where you put your purchases and automatically charges funds when you leave the store. AlphaFhold by DeepMind, meanwhile, predicts the protein structure for biologists. Amazon Go and AlphaFhold’s neural networks can’t substitute for each other.
How to train neural networks
The method used to teach neural networks is called machine learning. It operates at the intersection of computer science and statistics. Within this method, there is a specific topic, called deep learning.
If education involves intense human intervention, it’s called machine learning. People hand-pick the relevant pieces of information, label them, explain to the network which features are important, and so on.
Let’s imagine that you want to teach a traditional statistical model to identify bicycles in pictures through machine learning:
- Take thousands of pictures of different vehicles and label them with words (“bicycle”, “truck”, “motorcycle”, and so on).
- Single out the features that the network should pay attention to in order to understand that it’s looking at a bicycle (such as two wheels).
- Check the results of the model and improve them. For example, it might fail to understand that a pink plastic toy bicycle is also a fully-fledged bicycle — and you need to tell the model that it is.
With deep learning, the process would look like this:
- People feed thousands of images labeled by vehicle type to the neural network.
- The network analyzes these pictures, singling out shared features.
- People check the results of the neural network’s work and improve them.
Sometimes, human professionals don’t know which patterns to look for, or there are just too many patterns for them to keep track of. This is when deep learning algorithms come in handy. For example, ornithologists can feed the network video footage of birds that live in remote areas. A neural network can detect that these birds tend to wake up at a specific time and keep singing their songs for around 42 minutes in sunny weather. This data facilitates and accelerates the process of studying the birds’ behavior.
Pros and cons of neural networks
Neural networks are revolutionizing the labor market thanks to the following benefits:
- High efficiency (relative to a human) and ability to operate 24/7.
- Multitasking: people can’t compare 1,000 objects simultaneously while AI can
- Extensive range of applications: from composing music playlists to suggesting treatments to patients.
- Continuous learning, flexibility, and adaptivity.
- Convenience in creating backups and retrieving data.
- User-friendly interfaces. Once it is packaged up in an application, it’s just as easy to operate a neural network as it is to use a smartphone.
However, they have at least five big disadvantages:
- Dependency on costly hardware and software: for example, if the owner of a healthcare center wants to start applying AI in radiology, they should be ready to invest at least $100,000 in it. Meanwhile, the process of training a large language model such as GPT-3 can cost over $4 million.
- Humans need to spend months collecting and labeling the massive amount of data neural networks require to be effectively trained.
- Black box nature: people can’t check what happens behind the scenes. They have to trust the results that the neural network delivers — but they can’t ask it to explain how it achieved these results.
- Dependency on data: if the input data contains errors or is incomplete, the result at the output stage can be inaccurate.
- No guarantee of 100% accuracy even if the input data is sound. When you ask a neural network to draw a hand, there can be a random number of fingers on it.
You can’t fully trust neural networks yet. Even ChatGPT, one of the most powerful solutions that can imitate human communication, tends to provide wrong answers — and is unlikely to overcome this problem in the next few years.
Neural networks don’t understand what they say. In this aspect, they resemble parrots who can speak a human language, but can’t interpret the meaning of words and phrases.
Impact of neural networks on the labor market
According to McKinsey, by 2030, automation can displace up to 800 million professionals around the world. PwC predicts that by the early 2030s, automation can put nearly 40% of US jobs at risk. At the same time, AI will create new workplaces for people who will create and operate neural networks.
For example, neural networks can analyze X-ray images. Every year, medical professionals from all over the world carry out 3.6 billion medical examinations of this type. Doctors and patients can benefit from neural networks in the following ways:
- Neural networks diagnose some diseases better than human professionals — such as tuberculosis and lung cancer.
- The workload of human professionals decreases. They can take care of more patients or devote more time to research.
On the flip side, doctors who work with neural networks face many challenges:
- The accuracy of AI work depends on the amount of data it has analyzed. Due to data privacy laws in some territories, it might get access only to limited pools of information.
- The regulations for applying neural networks differ from one territory to another.
- The result of the network’s analysis might depend on the X-ray machines and lenses used, as well as the settings for acquiring the images.
95% of radiologists wouldn’t trust AI algorithms to run autonomously.
No matter which sector you take, you’ll see that neural networks can’t fully replace human specialists. Instead, these two will be working hand in hand.
To remain relevant in the labor market, people should be ready to switch occupations and upgrade their skills. Humans need to understand neural networks, manually correct their answers, feed the correct answers to them, and provide the networks with even larger sets of data to analyze. The future opens impressive opportunities to those who love to learn, earn money, and bring value to society!
If you’re excited about neural networks, consider becoming a specialist in data science or data analytics. It’s not as difficult as some might think! At Practicum, you can learn data science and data analytics in just 9 and 10 months respectively. Both programs are fully remote and ideal for beginners who start from scratch. It’s part-time education, so you won’t need to quit your current job. After graduation, you’ll become a pro in machine learning and deep learning and possess the skills to create and supervise neural networks!