WILL LEARNING MACHINES TAKE OVER THE WORLD?

Learning is an integral part of our lives, and the lives of all animals, but do you realize how wondrous our learning ability is? When we try to build machines that can learn, we are faced with deep questions about the nature and functioning of intelligence. In this article, I will tell you about special artificial networks, called neural networks, that mimic the brain to produce intelligent behavior. Neural networks are an integral part of artificial intelligence, widely used in many daily applications, from face recognition to autonomous driving. I am certain that neural networks will play a key role in our future lives, making them more comfortable and improving our understanding of big riddles such as what intelligence is and how our brains work.


ARE MACHINES SMARTER THAN HUMANS?
Are you intelligent?I assume that all of you would say "yes".Now, what about your dog or cat?And how about your smartphone?Though intelligence is intuitive for us, it is actually a complex phenomenon [ ] and there is no clear definition of intelligence and no absolute measure of how intelligent a system is.What we often do in artificial intelligence (AI) is try to reproduce, or surpass, the

ARTIFICIAL INTELLIGENCE (AI)
The ability to reproduce, and even surpass, with machines the capabilities of humans and animals.
capabilities of humans and animals, particularly intellectual abilities like understanding language and reasoning.
AI is a moving target.Whenever some problem is solved so that machines outperform humans in a specific task, it is no longer AI.Many things which used to be part of AI are not any more, such as using GPS to navigate the best route or performing complex arithmetical calculations.Through advancing AI we also learned that humans are relatively good, or intelligent, at certain things but pretty bad at other things.For example, until recently, people thought that we were unbeatable at board games like Chess and Go.But today we know that advanced AI systems can beat the best human players-see this documentary to learn about the first AI system (AlphaGo) to beat a human world champion, in .Today, there are other tasks which AI systems can perform better than humans, such as instantaneous translation and image recognition.But even the most intelligent systems are highly specialized: they are good only at a very narrow range of tasks [ ].
Truly intelligent machines must be able to do what all animals do-learn [ ]. Learning means adapting, or using knowledge gained in experiences to perform well in a task within new scenarios.Humans and animals learn naturally, but how do machines learn?That is the research of machine learning (ML).In ML, we develop systems

MACHINE LEARNING (ML)
Teaching machines how to learn, or improve their performance, as result of previous experience, with as little human intervention as possible.
and algorithms that take in data and use it to continuously update ALGORITHM A set of instructions that tells machines how to operate.their behavior and improve their performance.ML is the basis of many applications, like automatic translation of webpages to di erent languages, voice control and face recognition in smartphones, driving autonomous cars and analyzing medical data.But the best learning machines still can not learn like humans.For example, it takes humans only about -h to learn how to drive properly, but currently no machine can drive autonomously as well as a human, even after thousands of hours.A big ML breakthrough was when we figured out how to imitate the way the brain learns.it with other neurons (Figure A; to learn more about neurons and synapses, see this paper).When we learn something, the strengths and locations of synapses in our brain change-think of learning as a network of units that receive inputs from other units, perform a computation, and transfer and output via junctions whose strengths of connections change over time.Using these principles, we can create a network of artificial neurons that can learn (Figure B).

ARTIFICIAL NEURONS
A representation of a neurons inside a computer.Artificial neurons receive digital inputs, perform a calculation, and transmit a digital output to other artificial neurons.

NEURAL NETWORKS-BIOLOGICAL AND ARTIFICIAL
A great leap in ML occurred with the birth of neural networks (NNs)

NEURAL NETWORK (NN)
A network of artificial computing elements that simulates the operation of neurons in the brain.
[ ].A NN is a network of connected artificial neurons, simulated within a computer and organized in layers (to learn more about NNs, read this paper on artificial neural networks or this paper on the similarity between nerve cells and AI).The first layer, which is the input layer (Figure ), receives the data that we feed it, for example, an image, video, or audio file.The middle layers, called the hidden layers, process the data, and the output layer gives the result of the computation (for example, identification of objects in an image or the transcription of voice to text  Now comes an interesting part, where our network must learn. Learning means updating the weights in the DNN (like the changing synapses in the brain), to perform the task better next time.Think of every weight in the DNN as a knob which can be tuned to change the overall result-usually tens of thousands of them.To tune these knobs, the DNN first compares the outcome it got (the activation of its neurons in the output layer, each representing a specific outcome) with the right outcome it should have gotten (expected activation of each neuron in the output layer).Then, it updates the weights in all layers, from the last to the first hidden layer, to minimize the di erence between the desired outcome and actual outcome.

LeCun
There are two common ways to train DNNs.The first is supervised training, which is more accurate but less e cient, where systems scan many data samples already labeled by people (e.g., images of cats and dogs correctly categorized).The second way is self-supervised training, where the network tries to reconstruct information it already has (its input) after representing the information within its layers-this does not require human intervention.
The common algorithm used for updating DNN weights is called backpropagation [ , ]. Backpropagation allows the DNN to improve its performance after every trial.The DNNs we currently train keep improving after every trial, but their rate of improvement decreases, so though at some point we stop investing in training, nonetheless the network keeps improving.After a network learns a training set of data, it then receives a test set to check its performance.When its performance is good enough for the specific task, it is ready to be used on new unseen data.

UNDERSTANDING THE WORLD USING CONVOLUTIONAL NEURAL NETWORKS
In the visual cortex-our brain's visual information processor-neurons respond to di erent visual features, called motifs, such as the orientation of lines [ ], or edges.Collections of neurons, each detecting a di erent motif, are repeated everywhere inside the visual cortex, detecting the same motifs on di erent parts of the image we see.Inspired by this architecture of the brain, researchers applied the same basic structure on artificial neural nets [ ].I further developed these neural nets, which were trained using backpropagation [ ] and that eventually led to the development of convolutional neural networks (CNNs).
The assumption behind CNNs is that signals we perceive from the world are composited of simple elements that build on top of each other to create complex objects.Simple motifs (e.g., lines and edges) are combined into parts of objects (e.g., square surfaces and legs), which are then combined into more complex objects (e.g., a table) and even categories of objects in a hierarchical manner (Figure ).
In a CNN, we take a set of neurons and connect them to a small patch in our input data, an image for example.Then, we take the same set of neurons and copy them all over the image, so that they look at every di erent part or patch.The neurons' output determines the presence or absence of a feature within each patch of the image.As the information propagates throughout the CNN layers, these features become more composite, or complex (to see an intuitive demonstration of how CNNs work, see this video).Amazingly, we do not have to tell the CNNs which features to look for-we just train them end-to-end using backpropagation, and they-quite

CAN NEURAL NETWORKS HELP US UNDERSTAND THE BRAIN?
As we have seen, the development of artificial neural networks was inspired by the brain's operations.So can NNs be useful in understanding the brain itself?My answer is they are necessary to understand the brain.The brain is a very complex organ and there are probably some underlying principles governing its abilities that we have not uncovered yet.Though neuroscientists have collected a huge amount of experimental data, there is no solid theory about how the brain works.To show that a theory of brain functioning is correct, we need to recreate it within a computer and see that it works in ways that are somewhat similar to that of the brain.If we manage to build a computerized brain that acts similarly to the biological brain, this indicates we captured common operating principles, despite the di erences between the systems.LeCun Today, many brain scientists are already using deep learning and NNs as models for explaining brain activity [ ], particularly in visual cortex, but it is also relevant for explaining how we process speech and text.NNs allow us to inspect information processes by recording the activity of all the neurons and using that to understand how they represent data.But to understand the collective operation of millions or billions of these elements, at some point we must describe the operation of the whole network at an abstract level.Having NNs that operate in similar ways to the brain would provide a great leap in our overall understanding of the brain, and of intelligence.

WILL MACHINES TAKE OVER THE WORLD?
I think the issue of controlling AI has become a modern "bogeyman", with horror forecasts of machines becoming smarter than us and dominating us (Figure ).However, in humans and thus in machines, intelligence does not equal a will to dominate the other.Another concern is aligning behavior of intelligent machines with humanity's values.Although it is hard to "educate" machines to behave properly, we can manage it-the same way we educate our kids to behave in society, and with the same regulations to guide social functioning.We can define intrinsic objectives that machines will pursue (think of like "core values") that they cannot violate or modify, ensuring that the machines' behavior stays aligned with our values and goals.

ZI-AN, AGE:
Hi, I am Zi-An, coming from a family of teachers, I have inherited a love for knowledge and learning.But my biggest joy?That is definitely my little brother.I absolutely love goofing, around and making him laugh!I am fascinated about science.I want to research the secrets of everlasting life, so that people I love will never grow old or die.

FigureFigure
Figure LeCun magically-figure out themselves which features should be used to identify the object.

Figure
Figure

Figure
Figure Image recognition using convolutional neural networks.Convolutional neural networks (CNNs) are used for various applications, including image recognition.The network represents an image in the input layer as a list of values (indicating the color of di erent pixels in the image).It then propagates the information into more complex objects (such as ears, eyes, whole face), until eventually the network output determines what object is present in the image (cat or dog in this example).

Figure
Figure

Figure
FigureShould we be afraid of machines?I believe the worry about machine dominance is unfounded as we can ensure that machines will be friends not foes.