Says Dr. Eli David from Bar Ilan University and a serial entrepreneur in the field of artificial intelligence.
"The deep neural networks will be in the 21st century what electricity was in the 20th century." Says Dr. Eli David from Bar Ilan University and a serial entrepreneur of start-up companies in the field of artificial intelligence. Dr. David will be one of the lecturers at the Silicon Club meeting that will take place on 18/9 in the Nofim Hall at the Eretz Israel Museum in Ramat Aviv, starting at 18:30 p.m.
Dr. Eli David is a leading expert in the field of artificial intelligence specializing in deep learning (neural networks) and evolutionary computation. He has published more than 40 articles in leading magazines in the field of artificial intelligence and conferences, and focuses mainly on applications of deep learning and genetic algorithms in various fields in the real world. He is the founder of Falcon, an advanced chess game software based on genetic algorithms and deep learning. The program came in second place in the Computer Speed Chess World Championship. Dr. David is the founder of Dist Instinct, which applies deep learning to the cyber field, and an AI consultant to large companies, venture capital funds and more.
"In the last two or three years we have seen a rapid development of the field of deep artificial neural networks - or deep learning. Until a few years ago, this field was considered completely negligible and many of the experts looked at it as far-fetched. The field of deep learning managed to reach overnight a situation where it affects all fields: computer vision, text processing, cyber security, games and more, and gave a knock out to all other methods of artificial intelligence."
"Deep learning is itself a subfield within machine learning. There are many methods to teach the machine. One of them is artificial neural networks, also known as deep learning. The significant difference is that in all classical methods of machine learning we require two steps. Let's take for example the face recognition problem. With all the classic methods, we cannot take the raw data in the image and put it directly into a machine learning module, whether it's pixels in the image, letters in the text, and more. Rather, we were required to carry out a step called feature extraction in which we use a human expert who builds us a set of features that can help identify a face - the distance between the pupils, or the distance between the nose and the mouth, and so on - dozens of such features."
"In the second step, these characteristics are entered into a deep learning module, which compares the dozens of characteristics and thus can identify the image. The feature extraction step is a bad step. By taking the rich raw data -- an image consisting of millions of pixels and turning it into a list of a few tens or hundreds of features, however good they may be, we've still lost most of the raw data."
The big problem has more to do with us as humans. Even the experts have a hard time looking at complicated problems and translating them into an orderly list of features. example. Each of us, if we see a picture of a dog or a cat, will tell within milliseconds whether it is a dog or a cat one hundred percent. Try to explain what exactly is the difference between a dog and a cat? I do this in an exercise with my students at the beginning of each course.
"The roots of the field of deep neural networks have their roots in the seventies and eighties. Already then there were neural networks when scientists tried to imitate the brain but achieved few and unimpressive successes. Improving the algorithm and improving the hardware, mainly thanks to Nvidia's graphics processors (GPU) we were able to go from flat networks of 4-2 layers to turn them into deep networks of 20 and thirty layers. And instead of having only a few thousand synapses (a synapse connects two neurons, just like in the brain), today there are billions of synapses in the large neuronal networks.
This is what allowed us to skip the feature extraction step for the first time. If we take facial recognition as an example, we put the pixels into a deep neural network without adding any human knowledge. The deep web learns by itself. In all areas where deep neural networks were used the improvement was unimaginable. In some cases, such as computer vision, we have even surpassed the capabilities of humans.
In the coming years I foresee a continuation of the accelerated improvement we have seen in recent years. We see no sign of slowing down in the pace of improvement in this area. Every day we see dramatic results and in many areas are approaching or even surpassing the capabilities of humans. My expectation for the coming decades is that we will see the continued exponential improvement in results and I am guessing that in most areas the artificial modules will be no less good than humans."
Will computers surpass humans in everything?
This comparison is currently out of place because the largest deep learning module we can train on the most powerful computer, which simulates billions of synapses, is 100 times smaller than the computational capacity of the brain. For comparison, just in the area of the outer cortex of the brain, which is the area of interest from the cognitive aspect, there are 150 trillion synapses. On the other hand, today's computers are already able to run neural networks a million times stronger than 30 years ago. If in the next thirty years we see a million-fold improvement and overcome the hardware problem, we will have artificial networks that will surpass the biological networks, and it will be possible to say that computers will be smarter than humans. There will even be a stage where the cognitive gap between computers and humans will be like the gap between humans and chimpanzees. What will happen then? This is a subject in which many and disturbing discussions are held. At one extreme there is the fear that the computers will take over us and even destroy us, expressed by people like Bill Gates, Elon Musk and Stephen Hawking, peace be upon him. At the other extreme, there are those who claim that it doesn't matter how smart the computers are, they will always serve us."
Will computers replace humans in specific tasks?
"In the coming years we will see a wave of replacing humans with machines. In ten or at most twenty years, a profession like taxi driver will be irrelevant and will no longer exist. The profession of a radiologist who interprets X-ray images, MRI, etc., will not exist in the way we know it. Already today, machine learning surpasses humans in many aspects of interpreting radiological images."
However, Dr. David is optimistic about the job opportunities that artificial intelligence will create. "At the beginning of the twentieth century, there were many workers in every city in the Western world, whose job it was to turn on the street lights one by one every evening. Once electricity came in there was a wave of unemployment, all these were unnecessary. But let's look at the huge amount of jobs created thanks to electricity. I strongly believe in this, that artificial intelligence will do for the 21st century what electricity did for the 20th century. Against the wave of unemployment there will be a wave of professions where a person will still have an advantage. The total replacement is not yet close.”
Will we see an interface between artificial intelligence and the human mind?
"The rate of development of artificial neural networks is very dizzying and the developments accordingly. On the other hand, the pace of improving our understanding of the human brain is still progressing at a very slow pace. We haven't seen a significant and big breakthrough that helped to better understand our brain. Still the most advanced and expensive machines in the world are very far from reaching the state of measuring the action of individual neurons and synapses in the brain during the thinking process. I think, again this is my personal opinion, that much faster we will reach a situation where we have an artificial intelligence as good as us and superior to us than a situation where we can connect to our brain and insert chips into it and expand the memory."
More of the topic in Hayadan: