Can AI be psychoanalyzed? What is the basic principle of artificial intelligence operation?
AI cannot be psychoanalyzed, AI works on the principle that a computer, using sensors (or human input), will gather facts about a scene. The computer will compare this information with information already stored to determine what it means.
Based on the information it has collected, the computer will calculate various possible actions and then predict which action is best. A computer can only solve problems that its program allows it to solve, and it does not have analytical capabilities in a general sense. The effectiveness of psychoanalysis as a method of psychotherapy is controversial.
There is also a lot of controversy about artificial intelligence. Heap these on yourself and you may find negative negativity. You can first try to find a Lacanian psychoanalytic AI. At the same time, artificial intelligence can also be analyzed as human beings by people who don't know the truth.
You can consider the AI to have passed the Turing test if it convinces a large percentage of people that it is human in the process. You can check out this "Google Engineer Claims AI Already Conscious" story, which was buzzing in 2022 for a while.
The evolution of artificial intelligence has a certain logical relationship. It starts from cognition, through logical training and deep learning, and evolves into a self-learning process of neural networks. This process is very long, from the birth of the PC to the current mobile Internet.
With half a century, with modeling algorithms, based on the development of communication technology, we have entered the fourth-generation communication and semiconductor GPU era, and this phenomenon has developed at a high speed. We know that Baidu has released version 1. In 2019, zero unmanned driving will be achieved.
Artificial intelligence is the study of the laws of human intelligent actions, the construction of artificial systems with a certain degree of intelligence, and the study of how to make computers complete tasks that previously required human intelligence, that is, the basic theory and method of how to use computer hardware and software to simulate human intelligent behavior. And technology.
Artificial intelligence is the process of simulating human thinking and judgment with machines. Generally, artificial intelligence consists of two parts: algorithms and training data. Through algorithms and training data, a set of thinking and judgment methods can be obtained, which can be applied to the judgment of on-site data. This is General content of artificial intelligence.
Google AI Writing Project: Neural Network Pseudo-Original
What is the principle of artificial intelligence
The principle of artificial intelligence, a simple description is: artificial intelligence = mathematical calculation writing cat . The intelligence of a machine depends on the "algorithm". Initially, it was discovered that 1 and 0 could be represented by the on and off of the circuit.
Then many circuits are organized together, and different arrangements can represent many things, such as colors, shapes, and letters. Coupled with logic elements (transistors), the mode of "input (press the switch button) - calculation (current through the line) - output (light is on)" is formed.
Imagine a two-way switch in your home. In order to achieve more complex calculations, it eventually became, "Large-Scale Integrated Circuits" - chips. After the circuit logic is nested layer by layer and packaged layer by layer, our method of changing the current state becomes a "programming language". That's what programmers do.
It will execute whatever the programmer tells the computer to do, and the entire process is fixed by the program. So, in order for a computer to perform a task, the programmer must first fully understand the flow of the task. Take the joint control elevator as an example: don't underestimate this elevator, it is quite "smart".
Consider what judgments it needs to make: up and down directions, whether it is full, peak hours, whether the stop time is sufficient, single and double floors, etc., you need to think about all the possibilities in advance, otherwise there will be bugs. In a way, programmers control the world.
But it is always like this, the programmers are too tired, you can see that their eyes are red from working overtime. So I thought: Can the computer learn by itself and solve problems by itself? And we just need to tell it a set of learning methods.
Everyone still remembers that in 1997, IBM used a specially designed computer to win the chess championship.
In fact, its method is very stupid - brute force calculation, the term is "exhaustive" (in fact, in order to save computing power, IBM manually trimmed a lot of unnecessary calculations for it, such as those obvious stupid moves, and for the card Sparov's style is optimized).
The computer calculates all the moves of each move, and then compares the game records of humans to find the optimal solution. In a word: work hard to create miracles! But when it comes to Go, it is no longer possible to exhaustively. No matter how powerful you are, there is always a limit.
The possible moves of Go far exceed the sum of all atoms in the universe (known), even if the most powerful supercomputer is used at present, it will take tens of thousands of years. Until quantum computers mature, electronic computers are next to impossible.
Therefore, the programmer added an extra layer of algorithm to AlphaGo: A. Calculate first: where to calculate and where to ignore. B. Then, calculate in a targeted manner. ——In essence, it is still calculation. How can there be any "perception"! In step A, how does it judge "where needs to be calculated"?
This is the core issue of "artificial intelligence": the process of "learning". Think about it, how do humans learn? All human cognition comes from summarizing observed phenomena, and predicting the future based on the summarized laws.
When you see a four-legged, short-haired, medium-sized, long-mouthed, barking animal and call it a dog, you will classify all similar objects you see in the future as dogs. However, the learning method of machines is qualitatively different from that of humans: humans can infer most unknowns by observing a few features.
Take one corner and turn three corners. The machine has to observe many, many dogs before it can know whether the one that is running is a dog or not. Can such a stupid machine be counted on to rule mankind? It just relies on computing power to do it recklessly! Work hard. Specifically, the algorithm it "learns" is called a "neural network" (more bluffing).
(The feature extractor summarizes the characteristics of the object, and then puts the features into a pool for integration, and the fully connected neural network outputs the final conclusion) It requires two prerequisites: 1. Eat a lot of data for trial and error, and gradually adjust your own Accuracy; 2. The more layers of the neural network, the more accurate the calculation (limited), and the greater the computing power required.
Therefore, the method of neural network, although it existed many years ago (it was still called "perceptron" at that time). However, limited by the amount of data and computing power, it has not been developed. The neural network sounds better than the perceptron. I don't know where the high-end is!
This once again tells us how important it is to have a nice name for research (zhuang) research (bi)! Now, both of these conditions are in place - big data and cloud computing. Whoever owns the data can do AI.
At present, the common application fields of AI: image recognition (security recognition, fingerprint, beautification, image search, medical image diagnosis), using "convolutional neural network (CNN)", which mainly extracts the characteristics of spatial dimensions to identify images.
Natural language processing (human-computer dialogue, translation) uses "recurrent neural network (RNN)", which mainly extracts the characteristics of the time dimension. Because speech is sequential, the time at which words appear determines the semantics. The design level of the neural network algorithm determines its ability to describe reality.
Top-notch Daniel Wu Enda once designed up to 100 multi-layer convolutional layers (too many layers are prone to fitting problems). When we deeply understand the meaning of calculation: there are clear mathematical laws. Then, this world has quantum (random) characteristics, which determines the theoretical limitations of computers.
— In fact, computers can't even generate truly random numbers. — Machines are still dumb. If you want to know more about Shenyou's in-depth artificial intelligence knowledge, you can send a private message to ask.
artificial neural network, what does artificial neural network mean
1. The concept of artificial neural network Artificial Neural Network (ANN), referred to as neural network (NN), is based on the basic principles of neural networks in biology. After understanding and abstracting the structure of the human brain and the response mechanism to external stimuli, Based on network topology knowledge, it is a mathematical model that simulates the processing mechanism of complex information by the nervous system of the human brain.
The model is characterized by parallel distributed processing capabilities, high fault tolerance, intelligence, and self-learning capabilities. It combines information processing and storage, with its unique knowledge representation and intelligent adaptive learning capabilities. attention in various subject areas.
It is actually a complex network with a large number of simple components connected to each other, with a high degree of nonlinearity, a system capable of complex logic operations and nonlinear relationships. A neural network is an operational model consisting of a large number of nodes (or neurons) connected to each other.
Each node represents a specific output function, called an activation function.
The connection between each two nodes represents a weighted value for the signal passing through the connection, called weight, and the neural network simulates human memory in this way. The output of the network depends on the structure of the network, the way the network is connected, the weights and the activation function.
The network itself is usually an approximation to a certain algorithm or function in nature, or it may be an expression of a logical strategy. The concept of neural network construction is inspired by the operation of biological neural networks.
The artificial neural network combines the understanding of the biological neural network with the mathematical statistical model, and realizes it with the help of mathematical statistical tools.
On the other hand, in the field of artificial perception of artificial intelligence, we use the method of mathematical statistics to enable the neural network to have human-like decision-making ability and simple judgment ability. This method is a further extension of traditional logic calculus.
In artificial neural networks, neuron processing units can represent different objects, such as features, letters, concepts, or some meaningful abstract patterns. The types of processing units in a network fall into three categories: input units, output units, and hidden units.
The input unit accepts signals and data from the outside world; the output unit realizes the output of system processing results; the hidden unit is a unit between the input and output units and cannot be observed from the outside of the system.
The connection weight between neurons reflects the connection strength between units, and the representation and processing of information is reflected in the connection relationship of network processing units.
Artificial neural network is a non-programmable, adaptive, brain-style information processing. Its essence is to obtain a parallel and distributed information processing function through the transformation and dynamic behavior of the network, and to imitate human beings in different degrees and levels. The information processing function of the brain nervous system.
Neural network is a mathematical model that uses information processing similar to the synaptic connection structure of the brain. It is simulated on the basis of human understanding of the combination of brain organization and thinking mechanism. It is rooted in neuroscience. , mathematics, thinking science, artificial intelligence, statistics, physics, computer science and a technology of engineering science.
Second, the development of artificial neural network The development of neural network has a long history. Its development process can be roughly summarized as the following four stages.
1. The first stage - the enlightenment period (1), MP neural network model: In the 1940s, people began to study neural networks.
In 1943, American psychologist McCulloch and mathematician Pitts proposed the MP model, which is relatively simple but of great significance.
In the model, the algorithm is realized by treating the neuron as a functional logic device, and the theoretical research of the neural network model has been initiated since then.
(2) Hebb's rule: In 1949, the psychologist Hebb published "The Organization of Behavior", in which he put forward the hypothesis that the strength of synaptic connections is variable.
This hypothesis holds that the learning process ultimately occurs at the synaptic site between neurons, and the strength of the synaptic connection changes with the activity of neurons before and after the synapse. This assumption developed into the very famous Hebb rule in neural networks.
This law tells people that the connection strength of synapses between neurons is variable, and this variability is the basis of learning and memory. Hebb's law lays the foundation for constructing a neural network model with a learning function.
(3) Perceptron model: In 1957, Rosenblatt proposed the Perceptron model based on the MP model.
The perceptron model has the basic principles of modern neural networks, and its structure is very consistent with neurophysiology.
This is a MP neural network model with continuously adjustable weight vectors. After training, it can achieve the purpose of classifying and identifying certain input vector patterns. Although it is relatively simple, it is the first real neural network.
Rosenblatt proved that two-layer perceptrons can classify inputs, and he also proposed an important research direction of three-layer perceptrons with hidden layer processing elements.
Rosenblatt's neural network model contains some of the basic principles of modern neural computers, thus forming a major breakthrough in neural network methods and technologies.
(4) ADALINE network model: In 1959, famous American engineers B.Widrow and M.Hoff proposed the adaptive linear element (Adaline for short) and Widrow-Hoff The neural network training method of learning rules (also known as the minimum mean square error algorithm or δ rule) and applying it to practical engineering has become the first artificial neural network used to solve practical problems, which has promoted the research and application of neural networks and develop.
The ADALINE network model is an adaptive linear neuron network model with continuous values, which can be used in adaptive systems.
2. The second stage--the low tide period, Minsky and Papert, one of the founders of artificial intelligence, conducted in-depth research on the functions and limitations of the network system represented by perceptrons mathematically, and published a sensational paper in 1969 The book "Perceptrons" pointed out that the function of a simple linear perceptron is limited, and it cannot solve the classification problem of two types of samples that are linearly inseparable. For example, it is impossible for a simple linear perceptron to realize the logical relationship of "exclusive or".
This assertion brought a heavy blow to the research of artificial neural network at that time. The 10-year low tide period in the history of neural network development began.
(1) Self-organizing neural network SOM model: In 1972, Professor KohonenT. of Finland proposed the self-organizing neural network SOM (Self-Organizing feature map).
Later neural networks were mainly implemented based on the work of KohonenT. SOM network is a class of unsupervised learning network, which is mainly used for pattern recognition, speech recognition and classification problems.
It adopts a "winner takes king" competitive learning algorithm, which is very different from the previously proposed perceptron. At the same time, its learning and training method is unguided training, which is a self-organizing network.
This learning and training method is often used as a training for extracting classification information when it is not known which classification types exist.
(2) Adaptive Resonance Theory ART: In 1976, Professor Grossberg of the United States proposed the famous Adaptive Resonance Theory ART (Adaptive Resonance Theory), whose learning process has the characteristics of self-organization and self-stabilization.
3. The third stage - the renaissance period (1), Hopfield model: In 1982, the American physicist Hopfield (Hopfield) proposed a discrete neural network, namely the discrete Hopfield network, which effectively promoted the neural network. Network research.
In the network, it introduced the Lyapunov function into it for the first time, and later researchers also called the Lyapunov function an energy function. Proves the stability of the network.
In 1984, Hopfield proposed a continuous neural network, changing the activation function of neurons in the network from discrete to continuous.
In 1985, Hopfield and Tank used the Hopfield neural network to solve the famous Traveling Salesman Problem. Hopfield neural network is a set of nonlinear differential equations.
Hopfield's model not only summarizes the nonlinear mathematical information storage and extraction functions of artificial neural networks, proposes dynamic equations and learning equations, but also provides important formulas and parameters for network algorithms, so that the construction and learning of artificial neural networks have a theory Guidance, under the influence of the Hopfield model, a large number of scholars have aroused their enthusiasm for the study of neural networks and actively devoted themselves to this academic field.
Because the Hopfield neural network has great potential in many aspects, people attach great importance to the research of neural network, and more people start to study neural network, which greatly promotes the development of neural network.
(2) Boltzmann machine model: In 1983, Kirkpatrick et al. realized that the simulated annealing algorithm can be used to solve NP complete combinatorial optimization problems. This method of simulating the annealing process of high-temperature objects to find the global optimal solution was first developed by Metropli et al. in 1953 proposed in the year.
In 1984, Hinton and young scholars such as Sejnowski proposed a large-scale parallel network learning machine, and clearly proposed the concept of hidden units. This learning machine was later called the Boltzmann machine.
Using the concepts and methods of statistical physics, Hinton and Sejnowsky first proposed a multi-layer network learning algorithm, called the Boltzmann machine model.
(3), BP neural network model: In 1986, on the basis of the multi-layer neural network model, melhart et al. proposed a backpropagation learning algorithm for multi-layer neural network weight correction---- BP algorithm (Error Back-Propagation) solves the learning problem of multi-layer forward neural network, and proves that multi-layer neural network has strong learning ability, it can complete many learning tasks and solve many practical problems.
(4) Parallel distributed processing theory: In 1986, "Parallel Distributed Processing: Exploration in the Microstructures of Cognition" edited by Rumelhart and McCkekkand, in this book, they established a parallel distributed processing theory, mainly dedicated to the microcosmic of cognition At the same time, the error backpropagation algorithm of the multi-layer feedforward network with nonlinear continuous transfer function is analyzed in detail, that is, the BP algorithm, which solves the problem that there is no effective algorithm for weight adjustment for a long time.
It can solve the problems that the perceptron cannot solve, answered the questions about the limitations of the neural network in the book "Perceptrons", and proved that the artificial neural network has a strong computing power in practice.
(5) Cellular neural network model: In 1988, Chua and Yang proposed the cellular neural network (CNN) model, which is a large-scale nonlinear computer simulation system with the characteristics of cellular automata.
Kosko built the bidirectional associative memory model (BAM), which has unsupervised learning capabilities. (6) Darwinism model: The Darwinism model proposed by Edelman had a great influence in the early 1990s. He established a neural network system theory.
(7) In 1988, Linsker proposed a new self-organization theory for the perceptron network, and formed the maximum mutual information theory based on Shanon's information theory, thus igniting the light of NN-based information application theory.
(8) In 1988, Broomhead and Lowe used radial basis function (RBF) to propose a design method for hierarchical networks, thus linking the design of NN with numerical analysis and linear adaptive filtering.
(9) In 1991, Haken introduced synergy into the neural network. In his theoretical framework, he believed that the cognitive process is spontaneous, and asserted that the pattern recognition process is the pattern formation process.
(10) In 1994, Liao Xiaoxin proposed the mathematical theory and foundation of cellular neural networks, which brought new progress in this field.
By expanding the activation function class of neural network, more general models of time-delayed cellular neural network (DCNN), Hopfield neural network (HNN), and bidirectional associative memory network (BAM) are given.
(11), in the early 1990s, Vapnik et al. proposed the concept of support vector machines (Supportvector machines, SVM) and VC (Vapnik-Chervonenkis) dimension.
After years of development, hundreds of neural network models have been proposed.
What is an Artificial Neural Network?
How does artificial intelligence work?
AI works like this: A computer gathers facts about a situation through sensors (or through human input). The computer compares this information with information already stored to determine what it means.
Based on the information gathered, the computer calculates possible actions and predicts which action will work best. Computers can only solve problems that the program allows to solve, and do not have analytical capabilities in a general sense.
Introduction: Artificial Intelligence (AI), abbreviated as AI in English, is a new technical science that studies and develops theories, methods, technologies and application systems for simulating, extending and expanding human intelligence.
Artificial intelligence is a branch of computer science that attempts to understand the essence of intelligence and produce a new intelligent machine that can respond in a manner similar to human intelligence. Research in this field includes robotics, language recognition, image recognition, natural language processing and expert systems, etc.
Artificial intelligence is a new technical science that studies and develops theories, methods, technologies and application systems for simulating, extending and expanding human intelligence. Since the birth of artificial intelligence, the theory and technology have become increasingly mature, and the application fields have also continued to expand, but there is no unified definition.
Artificial intelligence is the simulation of the information process of human consciousness and thinking. Artificial intelligence is not human intelligence, but it can think like human beings, and it may surpass human intelligence. But this kind of advanced artificial intelligence that can think for itself still needs breakthroughs in scientific theory and engineering.
Scientific introduction: 1. Practical application of machine vision: machine vision, fingerprint recognition, face recognition, retina recognition, iris recognition, palmprint recognition, expert system, automatic planning, intelligent search, theorem proving, game, automatic programming, intelligent control , robotics, language and image understanding, genetic programming, etc.
2. Disciplinary scope Artificial intelligence is a frontier discipline, belonging to the intersection of natural science and social science. 3. Involving disciplines such as philosophy and cognitive science, mathematics, neurophysiology, psychology, computer science, information theory, cybernetics, and uncertainty theory.
4. Research areas Natural language processing, knowledge representation, intelligent search, reasoning, planning, machine learning, knowledge acquisition, combinatorial scheduling problems, perception problems, pattern recognition, logic programming soft computing, imprecise and uncertain management, artificial life , Neural Networks, Complex Systems, Genetic Algorithms.
5. Consciousness and artificial intelligence Artificial intelligence, in its essence, is a simulation of the information process of human thinking.
The Differences and Connections Between Artificial Intelligence, Machine Learning and Deep Learning
Some people say that artificial intelligence (AI) is the future, artificial intelligence is science fiction, and artificial intelligence is also a part of our daily life. These evaluations can be said to be correct, depending on which kind of artificial intelligence you are referring to.
Earlier this year, Google DeepMind's AlphaGo defeated the South Korean Go master Lee Sedol 9-dan.
When the media described the success of DeepMind, artificial intelligence (AI), machine learning (machine learning) and deep learning (deep learning) were all used.
All three played a role in AlphaGo's victory over Lee Sedol, but they don't mean the same thing. Today we use the simplest method - concentric circles, to visually show the relationship and application of the three of them.
As shown in the figure above, artificial intelligence is the first to appear, and it is also the largest and outermost concentric circle; the second is machine learning, a little later; the innermost is deep learning, the core driver of today's artificial intelligence explosion. In the 1950s, artificial intelligence was once extremely optimistic.
Afterwards, smaller subsets of artificial intelligence developed. First came machine learning, then deep learning. Deep learning is again a subset of machine learning. Deep learning has had an unprecedented impact.
From conception to prosperity In 1956, several computer scientists gathered at the Dartmouth Conference (Dartmouth Conferences) and proposed the concept of "artificial intelligence".
Since then, artificial intelligence has been lingering in people's minds and slowly incubating in scientific research laboratories. In the following decades, artificial intelligence has been reversing the poles, or it has been called the prediction of the dazzling future of human civilization;
Frankly speaking, until 2012, these two voices still existed at the same time. In the past few years, especially since 2015, artificial intelligence has exploded. A large part of this is due to the widespread use of GPUs, making parallel computing faster, cheaper, and more efficient.
Of course, the combination of infinitely expanding storage capacity and the sudden burst of data flood (big data) has also resulted in a massive explosion of image data, text data, transaction data, and mapping data.
Let's take a look at how computer scientists have developed artificial intelligence from the earliest semblance to applications that are used by hundreds of millions of users every day.
| Artificial Intelligence (Artificial Intelligence)—Giving intelligence to machines As early as the meeting in the summer of 1956, the pioneers of artificial intelligence dreamed of using computers that had just appeared at that time to construct complex machines with the same essential characteristics as human intelligence. machine.
This is what we now call "General AI". This omnipotent machine, which has all our senses (even more than humans), all our rationality, can think like us.
People have always seen such machines in movies: friendly, like C-3PO in Star Wars; evil, like the Terminator. Strong artificial intelligence exists only in movies and science fiction for now, and it's easy to understand why we can't achieve them, at least not yet.
What we can achieve at present is generally called "Narrow AI". Narrow AI is technology that can perform specific tasks as well as, or even better than, humans. For example, image classification on Pinterest; or face recognition on Facebook.
These are examples of Narrow AI in practice. These technologies realize some specific parts of human intelligence. But how are they achieved? Where does this intelligence come from? This brings us to the inner layer of the concentric circle, machine learning.
| Machine Learning - An Approach to Artificial Intelligence Machine learning, at its most basic, is the use of algorithms to parse data, learn from it, and then make decisions and predictions about real-world events.
Unlike traditional software programs that are hard-coded to solve specific tasks, machine learning is "trained" with large amounts of data, and various algorithms learn from the data how to complete tasks. Machine learning comes directly from the early field of artificial intelligence.
Traditional algorithms include decision tree learning, inferential logical programming, clustering, reinforcement learning, and Bayesian networks, among others. As we all know, we have not achieved strong artificial intelligence. Early machine learning methods couldn't even achieve weak AI.
The most successful application area of machine learning is computer vision, although it still requires a lot of hand coding to get the job done.
People need to manually write classifiers, edge detection filters so that the program can recognize where the object starts and where it ends; write a shape detection program to determine whether the detected object has eight sides; write a classifier to recognize the letters "ST-OP ".
Using these hand-written classifiers, one can finally develop algorithms to perceive an image and determine whether it is a stop sign or not. That's not bad, but not the kind of success that lifts people up.
Especially in cloudy and foggy days, when the signs are not so clearly visible, or partly blocked by trees, the algorithm will hardly succeed. This is why, some time ago, the performance of computer vision has been unable to approach human capabilities. It is too rigid and too easily disturbed by environmental conditions.
Over time, the development of learning algorithms changed everything.
| Deep learning - a technology to achieve machine learning Artificial Neural Networks (Artificial Neural Networks) is an important algorithm in early machine learning, after decades of ups and downs.
The principle of neural networks is inspired by the physical structure of our brains - neurons that are interconnected with each other. But unlike a neuron in the brain that can connect to any neuron within a certain distance, artificial neural networks have discrete layers, connections, and directions in which data travels.
For example, we can divide an image into image patches and input them to the first layer of a neural network. Every neuron in the first layer passes data to the second layer. The neurons in the second layer do a similar job, passing the data to the third layer, and so on, until the last layer, and then generating the result.
Each neuron assigns weights to its inputs, and the correctness of this weight is directly related to the task it performs. The final output is determined by summing these weights. Let's still take the example of the Stop sign.
All the elements of a stop sign image are smashed and "examined" with neurons: octagonal shape, firetruck-like red color, bold lettering, typical dimensions of a traffic sign, and still motion features and more.
The task of the neural network is to come up with a conclusion whether it is a stop sign or not. Based on all the weights, the neural network will come up with a well-thought-out guess — a "probability vector."
In this example, the system may give the following results: 86% may be a stop sign; 7% may be a speed limit sign; 5% may be a kite hanging from a tree and so on. The network structure then tells the neural network whether its conclusions are correct.
Even this example is relatively advanced. Until recently, neural networks were largely forgotten by the AI community. In fact, in the early days of artificial intelligence, neural networks already existed, but the contribution of neural networks to "intelligence" was minimal.
The main problem is that even the most basic neural networks are computationally intensive. The computing requirements of neural network algorithms are difficult to be met.
However, there are still some devout research teams, represented by Geoffrey Hinton of the University of Toronto, who persist in research and realize the operation and proof of concept of parallel algorithms targeting supercomputing. But those efforts didn't bear fruit until GPUs became widely available.
Let's go back to this stop sign recognition example. Neural networks are modulated, trained, and error-prone from time to time. What it needs most is training.
Hundreds of thousands or even millions of images are required for training until the input weights of the neurons are modulated very precisely, no matter whether it is foggy, sunny or rainy, the correct result can be obtained every time.
Only then can we say that the neural network has successfully learned to look like a stop sign; or in the Facebook application, the neural network has learned your mother's face; or in 2012, when Professor Andrew Ng Google implemented a neural network to learn the appearance of a cat and so on.
Professor Wu's breakthrough lies in significantly increasing these neural networks from the foundation. There are many layers and many neurons, and then a large amount of data is input to the system to train the network. In Professor Wu's case, the data are images from 10 million YouTube videos.
Professor Wu added "deep" to deep learning. The "depth" here refers to the many layers in the neural network.
Now, image recognition trained with deep learning can even do better than humans in some scenarios: from identifying cats, to identifying early components of cancer in blood, to identifying tumors in MRIs.
Google's AlphaGo first learned how to play Go and then trained it playing Go against itself. The way it trains its neural network is to keep playing chess with itself, repeatedly, without stopping.
| Deep learning, giving artificial intelligence a bright future Deep learning enables machine learning to achieve many applications and expands the scope of artificial intelligence. Deep learning achieves a variety of tasks so destructively that it seems possible that all machine-assisted functions are possible.
Driverless cars, preventive healthcare, and even better movie recommendations are all here, or on the verge of being there. Artificial intelligence is now and tomorrow. With deep learning, artificial intelligence can even reach the level of science fiction we imagined.
I took your C-3PO, as long as you have your Terminator.
What is the relationship between artificial intelligence, machine learning and deep learning?
Some people say that artificial intelligence (AI) is the future, artificial intelligence is science fiction, and artificial intelligence is also a part of our daily life. These evaluations can be said to be correct, depending on which kind of artificial intelligence you are referring to.
Earlier this year, Google DeepMind's AlphaGo defeated the South Korean Go master Lee Sedol 9-dan.
When the media described the success of DeepMind, artificial intelligence (AI), machine learning (machine learning) and deep learning (deep learning) were all used.
All three played a role in AlphaGo's victory over Lee Sedol, but they don't mean the same thing. Today we use the simplest method - concentric circles, to visually show the relationship and application of the three of them.
Turn left|Turn right As shown in the picture above, artificial intelligence is the first to appear, and it is also the largest and outermost concentric circle; the second is machine learning, a little later; the innermost is deep learning, the core driver of today's artificial intelligence explosion . In the 1950s, artificial intelligence was once extremely optimistic.
Afterwards, smaller subsets of artificial intelligence developed. First came machine learning, then deep learning. Deep learning is again a subset of machine learning. Deep learning has had an unprecedented impact.
From conception to prosperity In 1956, several computer scientists gathered at the Dartmouth Conference (Dartmouth Conferences) and proposed the concept of "artificial intelligence".
Since then, artificial intelligence has been lingering in people's minds and slowly incubating in scientific research laboratories. In the following decades, artificial intelligence has been reversing the poles, or it has been called the prediction of the dazzling future of human civilization;
Frankly speaking, until 2012, these two voices still existed at the same time. In the past few years, especially since 2015, artificial intelligence has exploded. A large part of this is due to the widespread use of GPUs, making parallel computing faster, cheaper, and more efficient.
Of course, the combination of infinitely expanding storage capacity and the sudden burst of data flood (big data) has also resulted in a massive explosion of image data, text data, transaction data, and mapping data.
Let's take a look at how computer scientists have developed artificial intelligence from the earliest semblance to applications that are used by hundreds of millions of users every day.
| Artificial Intelligence (Artificial Intelligence) - to give machines the intelligence to turn left | A machine with the same essential properties as human intelligence.
This is what we now call "General AI". This omnipotent machine, which has all our senses (even more than humans), all our rationality, can think like us.
People have always seen such machines in movies: friendly, like C-3PO in Star Wars; evil, like the Terminator. Strong artificial intelligence exists only in movies and science fiction for now, and it's easy to understand why we can't achieve them, at least not yet.
What we can achieve at present is generally called "Narrow AI". Narrow AI is technology that can perform specific tasks as well as, or even better than, humans. For example, image classification on Pinterest; or face recognition on Facebook.
These are examples of Narrow AI in practice. These technologies realize some specific parts of human intelligence. But how are they achieved? Where does this intelligence come from? This brings us to the inner layer of the concentric circle, machine learning.
| Machine Learning - An Approach to Artificial Intelligence Turn Left|Turn Right Machine learning, at its most basic, is the use of algorithms to parse data, learn from it, and then make decisions and predictions about real-world events.
Unlike traditional software programs that are hard-coded to solve specific tasks, machine learning is "trained" with large amounts of data, and various algorithms learn from the data how to complete tasks. Machine learning comes directly from the early field of artificial intelligence.
Traditional algorithms include decision tree learning, inferential logical programming, clustering, reinforcement learning, and Bayesian networks, among others. As we all know, we have not achieved strong artificial intelligence. Early machine learning methods couldn't even achieve weak AI.
The most successful application area of machine learning is computer vision, although it still requires a lot of hand coding to get the job done.
People need to manually write classifiers, edge detection filters so that the program can recognize where the object starts and where it ends; write a shape detection program to determine whether the detected object has eight sides; write a classifier to recognize the letters "ST-OP ".
Using these hand-written classifiers, one can finally develop algorithms to perceive an image and determine whether it is a stop sign or not. That's not bad, but not the kind of success that lifts people up.
Especially in cloudy and foggy days, when the signs are not so clearly visible, or partly blocked by trees, the algorithm will hardly succeed. This is why, some time ago, the performance of computer vision has been unable to approach human capabilities. It is too rigid and too easily disturbed by environmental conditions.
Over time, the development of learning algorithms changed everything.
| Deep learning - a technology to achieve machine learning Turn left | Turn right Artificial Neural Networks (Artificial Neural Networks) is an important algorithm in early machine learning, after decades of ups and downs.
The principle of neural networks is inspired by the physical structure of our brains - neurons that are interconnected with each other. But unlike a neuron in the brain that can connect to any neuron within a certain distance, artificial neural networks have discrete layers, connections, and directions in which data travels.
For example, we can divide an image into image patches and input them to the first layer of a neural network. Every neuron in the first layer passes data to the second layer. The neurons in the second layer do a similar job, passing the data to the third layer, and so on, until the last layer, and then generating the result.
Each neuron assigns weights to its inputs, and the correctness of this weight is directly related to the task it performs. The final output is determined by summing these weights. Let's still take the example of the Stop sign.
All the elements of a stop sign image are smashed and "examined" with neurons: octagonal shape, firetruck-like red color, bold lettering, typical dimensions of a traffic sign, and still motion features and more.
The task of the neural network is to come up with a conclusion whether it is a stop sign or not. Based on all the weights, the neural network will come up with a well-thought-out guess — a "probability vector."
In this example, the system may give the following results: 86% may be a stop sign; 7% may be a speed limit sign; 5% may be a kite hanging from a tree and so on. The network structure then tells the neural network whether its conclusions are correct.
Even this example is relatively advanced. Until recently, neural networks were largely forgotten by the AI community. In fact, in the early days of artificial intelligence, neural networks already existed, but the contribution of neural networks to "intelligence" was minimal.
The main problem is that even the most basic neural networks are computationally intensive. The computing requirements of neural network algorithms are difficult to be met.
However, there are still some devout research teams, represented by Geoffrey Hinton of the University of Toronto, who persist in research and realize the operation and proof of concept of parallel algorithms targeting supercomputing. But those efforts didn't bear fruit until GPUs became widely available.
Let's go back to this stop sign recognition example. Neural networks are modulated, trained, and error-prone from time to time. What it needs most is training.
Hundreds of thousands or even millions of images are required for training until the input weights of the neurons are modulated very precisely, no matter whether it is foggy, sunny or rainy, the correct result can be obtained every time.
Only then can we say that the neural network has successfully learned to look like a stop sign; or in the Facebook application, the neural network has learned your mother's face; or in 2012, when Professor Andrew Ng Google implemented a neural network to learn the appearance of a cat and so on.
Professor Wu's breakthrough lies in significantly increasing these neural networks from the foundation. There are many layers and many neurons, and then a large amount of data is input to the system to train the network. In Professor Wu's case, the data are images from 10 million YouTube videos.
Professor Wu added "deep" to deep learning. The "depth" here refers to the many layers in the neural network.
Now, image recognition trained with deep learning can even do better than humans in some scenarios: from identifying cats, to identifying early components of cancer in blood, to identifying tumors in MRIs.
Google's AlphaGo first learned how to play Go and then trained it playing Go against itself. The way it trains its neural network is to keep playing chess with itself, repeatedly, without stopping.
| Deep learning, giving artificial intelligence a bright future Deep learning enables machine learning to achieve many applications and expands the scope of artificial intelligence. Deep learning achieves a variety of tasks so destructively that it seems possible that all machine-assisted functions are possible.
Driverless cars, preventive healthcare, and even better movie recommendations are all here, or on the verge of being there. Artificial intelligence is now and tomorrow. With deep learning, artificial intelligence can even reach the level of science fiction we imagined.
I took your C-3PO, as long as you have your Terminator.
Artificial intelligence, machine learning, deep learning, what is the difference
Some people say that artificial intelligence (AI) is the future, artificial intelligence is science fiction, and artificial intelligence is also a part of our daily life. These evaluations can be said to be correct, depending on which kind of artificial intelligence you are referring to.
Earlier this year, Google DeepMind's AlphaGo defeated the South Korean Go master Lee Sedol 9-dan.
When the media described the success of DeepMind, artificial intelligence (AI), machine learning (machine learning) and deep learning (deep learning) were all used.
All three played a role in AlphaGo's victory over Lee Sedol, but they don't mean the same thing. Today we use the simplest method - concentric circles, to visually show the relationship and application of the three of them.
As shown in the figure above, artificial intelligence is the first to appear, and it is also the largest and outermost concentric circle; the second is machine learning, a little later; the innermost is deep learning, the core driver of today's artificial intelligence explosion. In the 1950s, artificial intelligence was once extremely optimistic.
Afterwards, smaller subsets of artificial intelligence developed. First came machine learning, then deep learning. Deep learning is again a subset of machine learning. Deep learning has had an unprecedented impact.
From conception to prosperity In 1956, several computer scientists gathered at the Dartmouth Conference (Dartmouth Conferences) and proposed the concept of "artificial intelligence".
Since then, artificial intelligence has been lingering in people's minds and slowly incubating in scientific research laboratories. In the following decades, artificial intelligence has been reversing the poles, or it has been called the prediction of the dazzling future of human civilization;
Frankly speaking, until 2012, these two voices still existed at the same time. In the past few years, especially since 2015, artificial intelligence has exploded. A large part of this is due to the widespread use of GPUs, making parallel computing faster, cheaper, and more efficient.
Of course, the combination of infinitely expanding storage capacity and the sudden burst of data flood (big data) has also resulted in a massive explosion of image data, text data, transaction data, and mapping data.
Let's take a look at how computer scientists have developed artificial intelligence from the earliest semblance to applications that are used by hundreds of millions of users every day.
| Artificial Intelligence (Artificial Intelligence)—Giving intelligence to machines As early as the meeting in the summer of 1956, the pioneers of artificial intelligence dreamed of using computers that had just appeared at that time to construct complex machines with the same essential characteristics as human intelligence. machine.
This is what we now call "General AI". This omnipotent machine, which has all our senses (even more than humans), all our rationality, can think like us.
People have always seen such machines in movies: friendly, like C-3PO in Star Wars; evil, like the Terminator. Strong artificial intelligence exists only in movies and science fiction for now, and it's easy to understand why we can't achieve them, at least not yet.
What we can achieve at present is generally called "Narrow AI". Narrow AI is technology that can perform specific tasks as well as, or even better than, humans. For example, image classification on Pinterest; or face recognition on Facebook.
These are examples of Narrow AI in practice. These technologies realize some specific parts of human intelligence. But how are they achieved? Where does this intelligence come from? This brings us to the inner layer of the concentric circle, machine learning.
| Machine Learning - An Approach to Artificial Intelligence Machine learning, at its most basic, is the use of algorithms to parse data, learn from it, and then make decisions and predictions about real-world events.
Unlike traditional software programs that are hard-coded to solve specific tasks, machine learning is "trained" with large amounts of data, and various algorithms learn from the data how to complete tasks. Machine learning comes directly from the early field of artificial intelligence.
Traditional algorithms include decision tree learning, inferential logical programming, clustering, reinforcement learning, and Bayesian networks, among others. As we all know, we have not achieved strong artificial intelligence. Early machine learning methods couldn't even achieve weak AI.
The most successful application area of machine learning is computer vision, although it still requires a lot of hand coding to get the job done.
People need to manually write classifiers, edge detection filters so that the program can recognize where the object starts and where it ends; write a shape detection program to determine whether the detected object has eight sides; write a classifier to recognize the letters "ST-OP ".
Using these hand-written classifiers, one can finally develop algorithms to perceive an image and determine whether it is a stop sign or not. That's not bad, but not the kind of success that lifts people up.
Especially in cloudy and foggy days, when the signs are not so clearly visible, or partly blocked by trees, the algorithm will hardly succeed. This is why, some time ago, the performance of computer vision has been unable to approach human capabilities. It is too rigid and too easily disturbed by environmental conditions.
Over time, the development of learning algorithms changed everything.
| Deep learning - a technology to achieve machine learning Artificial Neural Networks (Artificial Neural Networks) is an important algorithm in early machine learning, after decades of ups and downs.
The principle of neural networks is inspired by the physical structure of our brains - neurons that are interconnected with each other. But unlike a neuron in the brain that can connect to any neuron within a certain distance, artificial neural networks have discrete layers, connections, and directions in which data travels.
For example, we can divide an image into image patches and input them to the first layer of a neural network. Every neuron in the first layer passes data to the second layer. The neurons in the second layer do a similar job, passing the data to the third layer, and so on, until the last layer, and then generating the result.
Each neuron assigns weights to its inputs, and the correctness of this weight is directly related to the task it performs. The final output is determined by summing these weights. Let's still take the example of the Stop sign.
All the elements of a stop sign image are smashed and "examined" with neurons: octagonal shape, firetruck-like red color, bold lettering, typical dimensions of a traffic sign, and still motion features and more.
The task of the neural network is to come up with a conclusion whether it is a stop sign or not. Based on all the weights, the neural network will come up with a well-thought-out guess — a "probability vector."
In this example, the system may give the following results: 86% may be a stop sign; 7% may be a speed limit sign; 5% may be a kite hanging from a tree and so on. The network structure then tells the neural network whether its conclusions are correct.
Even this example is relatively advanced. Until recently, neural networks were largely forgotten by the AI community. In fact, in the early days of artificial intelligence, neural networks already existed, but the contribution of neural networks to "intelligence" was minimal.
The main problem is that even the most basic neural networks are computationally intensive. The computing requirements of neural network algorithms are difficult to be met.
However, there are still some devout research teams, represented by Geoffrey Hinton of the University of Toronto, who persist in research and realize the operation and proof of concept of parallel algorithms targeting supercomputing. But those efforts didn't bear fruit until GPUs became widely available.
Let's go back to this stop sign recognition example. Neural networks are modulated, trained, and error-prone from time to time. What it needs most is training.
Hundreds of thousands or even millions of images are required for training until the input weights of the neurons are modulated very precisely, no matter whether it is foggy, sunny or rainy, the correct result can be obtained every time.
Only then can we say that the neural network has successfully learned to look like a stop sign; or in the Facebook application, the neural network has learned your mother's face; or in 2012, when Professor Andrew Ng Google implemented a neural network to learn the appearance of a cat and so on.
Professor Wu's breakthrough lies in significantly increasing these neural networks from the foundation. There are many layers and many neurons, and then a large amount of data is input to the system to train the network. In Professor Wu's case, the data are images from 10 million YouTube videos.
Professor Wu added "deep" to deep learning. The "depth" here refers to the many layers in the neural network.
Now, image recognition trained with deep learning can even do better than humans in some scenarios: from identifying cats, to identifying early components of cancer in blood, to identifying tumors in MRIs.
Google's AlphaGo first learned how to play Go and then trained it playing Go against itself. The way it trains its neural network is to keep playing chess with itself, repeatedly, without stopping.
| Deep learning, giving artificial intelligence a bright future Deep learning enables machine learning to achieve many applications and expands the scope of artificial intelligence. Deep learning achieves a variety of tasks so destructively that it seems possible that all machine-assisted functions are possible.
Driverless cars, preventive healthcare, and even better movie recommendations are all here, or on the verge of being there. Artificial intelligence is now and tomorrow. With deep learning, artificial intelligence can even reach the level of science fiction we imagined.
I took your C-3PO, as long as you have your Terminator.