Neural Networks for Dummies and Humanities: When Robots Replace People
Posted: Wed Jan 29, 2025 6:41 am
Artificial neural networks surround us everywhere: Alice will tell you the weather for the day, the navigator will build a quick route to work, and the smart feed will show a selection of news based on your interests. Thanks to neural networks, anyone can feel like a great artist or writer, even if they can’t draw or express their thoughts beautifully. However, for many, they still remain a mystery. As well as the phrase Big Data, which we have already talked about .
Try to guess where the neural network worked and where a person did! We have come up with a short test in which we invite you to compare the results and test your intuition. At the end, you will find some tips on how to distinguish an author's work from a machine. For the test, we used the Balabob and MidJourney services , for which we are immensely grateful to their developers.
Take the test
In his book "How a Machine Learns: The Deep Learning Revolution in uganda whatsapp list Neural Networks" Yann LeCun explains how neural networks work and where they are used. The author is a laureate of the Turing Award, the equivalent of the Nobel Prize in computing. He is called the godfather of neural networks. The review will be useful for those who use the achievements of neural networks and want to learn more about them without delving into complex technical details.
Cover of the book 'How a Machine Learns'
How neural networks learn
The first learning machine was created in 1957 by American psychologist Frank Rosenblatt at the Cornell University Aviation Laboratory in Buffalo, USA. The scientist was inspired by the work of neurons in the human brain and, by analogy, created an artificial neural network, which he called a perceptron.
The neural network in the human brain consists of 86 billion nerve cells, or neurons, that are connected to each other. Artificial neural networks, in turn, consist of artificial neurons, mathematical functions, which are also connected to each other.
Like a human, a neural network learns by changing the connections between neurons. The easiest way to track this process is with the aplysia mollusk. It has a very simple nervous system that controls external gills. If you touch the gills, the mollusk will first pull them in, and then release them after a while. If you touch the gills again and again, the mollusk will gradually start releasing them faster, and then stop pulling them in altogether. This is how neural connections adapt to external stimuli, that is, they learn.
There are many ways to train neural networks. Most of them consist of two stages: finding the main rule and debugging. In the first stage, neural networks are shown billions of pictures and told what they depict. The machine finds distinctive features of different objects and develops its own algorithm for distinguishing them. In the second stage , they check whether the neural network can correctly name pictures that it has not yet seen. If the machine makes a mistake, the operator informs it about it. Then the neural network reconfigures its internal connections to give the correct answer next time.
For example, to teach a machine to distinguish ships from planes, you first need to collect thousands of photos of both and upload them to a neural network. Then show it an image of a ship. If the machine gives the right answer, nothing needs to be changed. If the machine gives the wrong answer, you need to adjust the system parameters so that its answer is closer to the correct one.
Similarly, to build a car that can drive itself, you first need to collect data from an experienced driver. To do this, every split second, you need to record the position of the car on the road and how the driver turns the steering wheel so that the car stays within the lane. As a result, in an hour of observation, scientists receive 36,000 car positions and steering wheel angles. The neural network then learns from this information.
The “magic” of learning is that a trained machine can go beyond what it was shown. It can correctly identify what is in a picture it sees for the first time. Or make the right decision on the road, even when it encounters a new obstacle.
Try to guess where the neural network worked and where a person did! We have come up with a short test in which we invite you to compare the results and test your intuition. At the end, you will find some tips on how to distinguish an author's work from a machine. For the test, we used the Balabob and MidJourney services , for which we are immensely grateful to their developers.
Take the test
In his book "How a Machine Learns: The Deep Learning Revolution in uganda whatsapp list Neural Networks" Yann LeCun explains how neural networks work and where they are used. The author is a laureate of the Turing Award, the equivalent of the Nobel Prize in computing. He is called the godfather of neural networks. The review will be useful for those who use the achievements of neural networks and want to learn more about them without delving into complex technical details.
Cover of the book 'How a Machine Learns'
How neural networks learn
The first learning machine was created in 1957 by American psychologist Frank Rosenblatt at the Cornell University Aviation Laboratory in Buffalo, USA. The scientist was inspired by the work of neurons in the human brain and, by analogy, created an artificial neural network, which he called a perceptron.
The neural network in the human brain consists of 86 billion nerve cells, or neurons, that are connected to each other. Artificial neural networks, in turn, consist of artificial neurons, mathematical functions, which are also connected to each other.
Like a human, a neural network learns by changing the connections between neurons. The easiest way to track this process is with the aplysia mollusk. It has a very simple nervous system that controls external gills. If you touch the gills, the mollusk will first pull them in, and then release them after a while. If you touch the gills again and again, the mollusk will gradually start releasing them faster, and then stop pulling them in altogether. This is how neural connections adapt to external stimuli, that is, they learn.
There are many ways to train neural networks. Most of them consist of two stages: finding the main rule and debugging. In the first stage, neural networks are shown billions of pictures and told what they depict. The machine finds distinctive features of different objects and develops its own algorithm for distinguishing them. In the second stage , they check whether the neural network can correctly name pictures that it has not yet seen. If the machine makes a mistake, the operator informs it about it. Then the neural network reconfigures its internal connections to give the correct answer next time.
For example, to teach a machine to distinguish ships from planes, you first need to collect thousands of photos of both and upload them to a neural network. Then show it an image of a ship. If the machine gives the right answer, nothing needs to be changed. If the machine gives the wrong answer, you need to adjust the system parameters so that its answer is closer to the correct one.
Similarly, to build a car that can drive itself, you first need to collect data from an experienced driver. To do this, every split second, you need to record the position of the car on the road and how the driver turns the steering wheel so that the car stays within the lane. As a result, in an hour of observation, scientists receive 36,000 car positions and steering wheel angles. The neural network then learns from this information.
The “magic” of learning is that a trained machine can go beyond what it was shown. It can correctly identify what is in a picture it sees for the first time. Or make the right decision on the road, even when it encounters a new obstacle.