Prof. Steve Farber from the University of Manchester, the developer of the first British personal computer BBC as well as the ARM architecture found in most mobile phones today, wants to build computers with artificial intelligence and learning ability. For this purpose, he built in his laboratory a computer based on a neural network that imitates the activity of the human brain
Prof. Steve Farber from the University of Manchester, the developer of the first British personal computer BBC as well as the ARM architecture found in most mobile phones today, wants to build computers with artificial intelligence and learning ability. To this end, he built a neural network-based computer in his laboratory, Farber was one of the recipients of the "Global Industry Leader" award at the ChipEx2016 conference.
"Neural networks have revolutionized machine learning and artificial intelligence, but they are still ineffective. For example, to recognize a voice on the phone, the task is transferred to the cloud and not done on the phone itself. The reason - today's computing architectures still do not work well with neural networks. We have a lot of work ahead of us before we know how to answer these problems optimally."
"Dr. Alan Turing's house, in the last six years of his life, is about 2 km from my house. In Britain, Turing is known to the general public for his work during World War II on deciphering the Enigma Nazi cipher. To the rest of the world he is known for his contribution to computer science. He laid the foundations for modern computing - and for artificial intelligence and came up with ideas that are now integrated into every computer we use."
"He came to Manchester because there was an active computer there that he didn't build but based on his ideas. He was interested in MORPHOGENESYS - how the symmetry of the cells is broken so that the cells that divide and become an embryo are identical, the symmetry is broken and they become specialized cells. He spent a lot of time developing the mathematics of the symmetry breaking process. But his most significant contribution was in the article Computer Machinery and Intelligence."
In 1950, when the first computers were two years old and included 126 bits of memory, he predicted that by the end of the 20th century machines would have processing capacities of gigabytes and memories of hundreds of megabytes. The article includes several predictions including the prediction that all a computer will need to be as smart as humans is more memory and he said that by the end of the century machines will have gigabytes, when a computer had 126 bits of memory. Indeed these were the capabilities of a typical personal computer in the year 2000.
""He is also known as the one who invented the Turing test for artificial intelligence, according to which if it is not possible to tell whether the interlocutor on the other side is a person or a computer, it is a sign that he has passed the test that reflects his artificial intelligence capabilities. In my opinion, Turing was surprised that even today, no computer passes his test. There are many reasons for this, but in my opinion the basic reason for the difficulty in building artificial intelligence is that we don't know much about my natural intelligence. For the past twenty years I have been researching the use of computers to speed up the understanding of the brain's working processes."
"In the press there are a lot of articles about artificial intelligence. We have all heard of deep learning - a new trend in machine learning. Google and Facebook are spending a lot of money to research the issue. Until ten years ago neural networks were always the second most important problem for solving computing problems. Today, thanks to the deep study, the method has become the main method for solving complex problems.
A few weeks ago we were informed about the breakthrough of the Alpha-Go system built by a British company that was acquired by Google. When Garry Kasparov was defeated by IBM's DEEP BLUE computer 20 years ago, he said that chess was an easy game to compute, even Turing wrote chess software for a computer that didn't exist yet. He (Kasparov) recreated this software and beat it easily. Then the DEEP BLUE computer managed to beat it but no deep learning was required. Go is a much more complicated game, and in less than 20 years we have software that beats the best human player.
"Alpha Go contains many deep learning techniques. It is actually a neural network and is improved by playing with copies of it millions of times and that's how it learned to defeat the best human player. Many people see this as a sign that we are approaching a new artificial intelligence. And some even warned of the consequences of this including important people in industry and science such as Stephen Hawking, Elon Musk and Bill Gates who warned that artificial intelligence is an existential threat to humanity.
"They support the singularity theory according to which if we build artificial intelligence it will improve exponentially all the time. In my opinion these talks are not justified. The more I examine human reason, the more I see that it has many layers and there is no specific factor that can be accelerated and have the ability to calculate like the brain. It is sophisticated and multidimensional, I don't see the idea of the singularity happening."
"As for the risk to humanity - artificial intelligence is not at the highest level of risk, people are more dangerous. The Pentagon will invest 12-15 billion dollars in weapons research based on artificial intelligence. Perhaps an international agreement should be reached to ban autonomous weapons. They are not ethically capable of making a decision, so there should be human involvement in the loop at least for the foreseeable future."
"What is artificial intelligence - the controversial human brain project funded by the European Union's ICT program to support neuroscience research. Today, the output is so great that no scientist can read all the articles, so they use neuron-based computing NEUROFORMATS to analyze and help read and understand studies that are enormous."
"Therefore, at the University of Manchester we built the SPINNAKER MACHINE with 500 ARM cores that occupies six cabinets at a cost of half a million dollars. It simulates the activity of half a billion neurons, much less than one percent of the capacity of the human brain, but nevertheless our research budget is limited."
16 תגובות
Finally someone tells the truth - the king is naked...
As of today we are stuck with an impressive toaster that can estimate a function based on statistical inference = lack of artificial intelligence, even if the toaster does it really, really, really fast it is still not consciousness or even something close
Singularity followers will probably have to wait for the next millennium...
I did not understand how the text in the title is reflected in the body of the article.
Correction - "but are already managing to perform ground-breaking tasks that until recently were considered the exclusive ability of the human brain."
I don't understand how it is possible to see all the amazing breakthroughs that are happening in front of our eyes in recent years in the field of artificial intelligence and in particular in the field of neural networks and continue to make claims like:
"It won't happen in the near future... we are far from understanding how the brain works... we don't know what to calculate... the field of artificial intelligence has been stuck for 50 years and is not progressing anywhere..."
Based on the knowledge we have already accumulated about the brain and based on principles of operation we have already learned about, we have been able to build neural networks that are not even a thousandth of our brain in size and complexity, but are already successful in performing groundbreaking tasks and were only recently considered an exclusive ability of the human brain (or a developed biological brain). Today, neural networks recognize a face better than a human, distinguish between a dog and a cat, learn to play by themselves without any guidance or direction in dozens of computer games, including XNUMXD games, and reach the level of a professional player in many of them...!
If someone doesn't understand what a huge breakthrough the neural network that beat the world champion in the Alpha-Go game is, I suggest reading here:
http://www.haaretz.co.il/magazine/.premium-1.2898680
From the article: "According to accepted calculations, chess has 10 solutions to the 50th power, and Alpha Go - 2 times 10 to the 170th power. This difference makes Go not only bigger and more complex than chess: it is so much bigger than chess that it is already something else altogether.
…. The breadth and depth of the game options in Go make the computational component - the attempt to calculate the game ahead impractical... to win it, the best Go players in the world use a cognitive component known as intuition... Computers know how to do millions of calculations per second without making a single mistake - but they have no intuition. And that's exactly the reason why before the start of the competition, I was convinced of his victory."
Miracles,
"I think you are ignoring something fundamental - a system like IBM's is likened to a dead person's brain. The mind is not hardware on which software ran. The brain is dynamic and constantly changing. The dynamics of the brain builds up over time, and a critical component is input from all our body's senses.'
I don't understand how you can make a claim that is so far from reality. IBM's neural networks are likened to a dead brain??
These neuron networks learn, new memories are created in them (which allow them, for example, to recognize faces, or distinguish between a dog and a cat), these networks are dynamic and the connections in the content are constantly changing according to the inputs they receive, for example pixels from a computer screen on which an Atari game is displayed, an input corresponding to the input of an eye Human watching the game.
miracles are you serious? A dead brain has such abilities?? Can a dead brain learn anything?
The neuron networks today are still not aware of themselves, but we must not forget that they do not contain even a thousandth of the number of connections and neurons that exist in a human brain. What will happen in 20-35 years when we already have cognitive computer chips that contain the amount of connections and nerve cells that exist in the brain? And they will be placed inside a robotic body that moves in space and receives inputs from cameras and sensors?
Is there any valid reason to assume that this chip will not at some point be at least as intelligent as a human if not more so? Is there any reason to assume that all the phenomena that exist in the human brain, such as thoughts and self-awareness, will not develop in him?
You can argue and say that it won't happen, but the way I see it, all the amazing developments we've seen in the field in recent years definitely point to a clear trend.
When the simultaneous calculation power increases 10 times by many powers (thanks to hardware synapses and qubits) and the method of calculation changes. It turns from a deterministic algorithm to machine learning, or to a superposition of solution modes, which corrects and improves its algorithms, I would not rule out the possibility that the computer will start to become self-aware. I can't be sure of either option. A computer that lets him make nuclear decisions and track down terrorists.
Look at the synapse prototype demos on YouTube - of the DARPA project. Think what would happen if the amount of his parallel calculations increased by a factor of ten and if he thought in neurons and qubits instead of bits. I see very powerful computing power happening much faster than even the ARM supercomputer, and making decisions in all kinds of areas of life, which do not fall short of a person, but even exceed him in certain ways. Does it have consciousness? Don't know in advance you are right. I wouldn't rule out the possibility that he will start to understand himself. The system will learn problems, analyze them and even understand how to optimize its own algorithms.
Joseph
It doesn't matter how fast you calculate - you need to know what to calculate 🙂 and we are very far from understanding that.
A quantum computer works up to 10 to the minus 14th power of a second and at the same time on a large number of calculations.
Joseph
I am one of those who study the subject at university...
The last wording is condescending. The intention is not condescending. Because of this, the neurons synapse in the hardware, today logic gates work at ten to the minus 10th power of a second, and quantum computers at less than that. The problems faced by the mathematics of machine learning, pattern recognition, optimal control, and game theory are dynamic.
If you remember in the movie Terminator in the trilogy it all started with a military computer.
A neuron is a dynamic entity. The neurons of the IBM processor are much "stronger" than those of ARM. These are hardware neurons. Using object-oriented software, for example, you can turn on neurons (create), and turn them off (annihilate). The computing power of IBM's planned supercomputer, based on SYNAPSE, will exceed that of ARM by tens of meters. You are directed to YOUTUBE videos that illustrate how it will be connected to the traffic cameras and the data bases and will perform actions like a human. But as far as I remember, he will also be responsible for the American missile program. Obviously not alone, but a main decision supporter. A second IBM supercomputer will be built in the next decade as a quantum computer. Here it is enough to apply 100 qubits in order to obtain a calculation power that exceeds any super-orthodox computer. People speculate that the human mind is very, very far from implementation. not so far As I explained - the mathematics of sensing and perception are already known. The mathematics of some decision-making processes is also known. This mathematics - it is for the reason that you read material about it before deciding whether it is qualified or not qualified to deal with the challenges you mentioned. I believe that the generation currently studying in universities is less skeptical than you.
Joseph
I think you are ignoring something essential - a system like IBM's is likened to the brain of a dead person. The mind is not hardware on which software ran. The brain is dynamic and constantly changing. The dynamics of the brain builds over time, and a critical component is input from all of our body's senses (and there are more than five senses).
It is one thing to build a simulation of a brain with 100 billion neurons and 100 trillion (and possibly many more) connections - but each such neuron contains dynamic information received from other neurons. We are very, very far from knowing how to sample this information in a short time, and the time must be extremely short because everything changes...
The mathematical theory of perception (cognition) and sensing through the senses is known to us. Basically it's machine learning and pattern recognition. Mathematics is different from all mathematics: not deterministic but probabilistic. But not probabilistic in the conventional sense. Includes elements of optimal control (mathematical theory) of game theory (the mind plays against nature) and differential geometry. Why geometry: a problem with many variables - arrange them all in a row and separate them with commas - a vector. Therefore the possible tool is geometry. In studying functions, most of us remember that we subtracted twice to find a minimum. Here the minimum is a minimum of action over time according to a defined constraint. Therefore differential = because they determine. Using such methods, decisions are made with a minimum of information, faces are recognized, complicated quantum problems are solved, and much more. It is possible that such structures of decision-making exist in us unconsciously.
What is not completely clear to us is probably consciousness. How consciousness is formed. I think means I exist. It is possible that a critical mass of nerve cells is created, and there may be another theory that we are not familiar with that requires exposure.
For God's sake, why isn't the article proofread before publication? Lots of errors.
IBM is developing a computer with human capabilities and with exposure to road cameras and NSA databases. The computer will be based on IBM synapse processors which are processors with hardware neurons (active filters) in the amount of billions of gates in all processors. Despite all the skeptics, IBM has a tradition of implementing such things. The second supercomputer will consist of superconducting devices cooled by liquid nitrogen. 100 quantum bits produce the computing power of a supercomputer.
That's why Farber's opinion is in one direction and IBM's in another. To all the skeptics, I would address my concern about the Terminator movie coming to fruition. What happens when a military computer that controls missile decision making processes on the one hand and people elimination decisions on the other hand comes to the conclusion that it doesn't need us. Will we have time to reach the switch and turn it off? I don't think the option will come up at all.