Prof. Shahar Kotinsky from the Technion and his colleagues from Tower developed hardware inspired by the brain and implementing artificial intelligence algorithms

In a study now published in Nature Electronics Prof. Kotinsky demonstrates the effectiveness of the said technology in neural networks. Dr. Wei Wang and Dr. Louis Daniel, who were members of Prof. Kotinsky's research group, and researchers from Tower Semiconductor are partners in the development

Computers broke into the world about eighty years ago, and their involvement in our lives is constantly deepening. From decade to decade, computers become smaller and faster, but their basic structure has not changed: processor and memory. This dual structure has many disadvantages, including the resources dedicated to communication between them - the processor performs calculations and sends them to memory storage, and the memory transmits information to the processor as needed. This communication requires valuable resources of time and energy, a problem that is particularly prominent in terminal devices such as a mobile phone.

Prof. Shahar Kotinsky from the Viterbi Faculty of Electrical and Computer Engineering Offers - and presents - a promising alternative: Memory that performs calculations on its own. It is a pure hardware device, that is, a chip in which the software is already embedded and is not external. In a study now published in Nature ElectronicsProf. Kotinsky demonstrates the effectiveness of the said technology in neural networks. Dr. Wei Wang and Dr. Louis Daniel, who were members of Prof. Kotinsky's research group, and researchers from Tower Semiconductor are partners in the development. "We usually refer to the computer as a 'brain'," says Prof. Kutinsky, "but in our real brain there is no separation between memory and processor or between thinking areas and storage areas. The hardware we developed is neuromorphic hardware - it draws inspiration from the brain, and like the brain, there is no separation between memory and processing. Its structure is made of neurons and synapses built using electrical components. This way we eliminate the bottleneck created by the need for communication between the two parts of the traditional computer."

Even in the world of classical, dual computing, major successes have been recorded in recent decades. The media enthusiastically reported on the victory of the "Deep Blue" software over the chess player Garry Kasparov and later on the victory of AlphaGo over the "Go" champion Lee Sedol. However, Prof. Kotinsky notes, "The AlphaGo software works on 1,500 processors, and in every single game it consumes electricity that costs $3,000. At the same time, a human 'Go' player can make do with an average sandwich, but no less important - the human player is also able to speak, drive and perform many other tasks that AlphaGo cannot perform. It is therefore clear today that the achievements in classical computing are very limited, and for a leap forward a paradigm leap is required."

In a previous article, Prof. Kotinsky demonstrated the effectiveness of a commercial memory component from Tower Semiconductor as a building block in artificial intelligence calculations. Now he has developed a multi-capability chip based on these memory components. This chip stores the information and processes it in the same unit, similar to the human brain. The greatness of the chip is its ability to learn from examples, and the task in the current study was handwriting reading. The chip, which was "trained" on a huge database of samples, provided high accuracy in handwriting recognition, while consuming only little energy - in contrast to existing software systems dealing with handwriting recognition.

Conceptually, artificial neural networks are similar to our brains: they receive examples of the subject being studied - handwriting, in this case - and deduce by themselves the differences between the letters and how to correctly identify letters. When the neural network is embedded in the hardware, the learning process strengthens the connectivity between components on the chip, similar to strengthening the connections between the synapses in our brain during learning. The new chip has countless potential uses. For example, it can be integrated into camera sensors of phones and other devices, which will save the need to convert between analog and digital signals and allow analysis of the image itself without the need to translate it into a digital format.

"Commercial companies are in a constant race to improve their products," says Prof. Kotinsky. "They cannot go back to the drawing board again and again and design completely new products, because the investment is huge and the risk is clear. This is where the advantage we have in the academy comes in - our ability to develop new technological concepts that will significantly improve existing products on the market."

The research, under the guidance of Prof. Kotinsky, was led by Dr. Wei Wang, who did his post-doctorate under the guidance of Prof. Kotinsky and currently heads his own research group in Shenzhen, China; and Dr. Louis Daniel, who completed his doctorate under the guidance of Prof. Kotinsky and currently works at Mobileye. Dr. Wang developed the theoretical concept of the chip and conducted the experiments and Dr. Daniel designed the chip and led the steps that led to its production.

The research was supported by the European Union under the ERC and FETOPEN programs.

for the scientific article inNature Electronics click here

More of the topic in Hayadan:

• The next manned spacecraft - mid-October 2005 * Li Wei will not take off on it
• Technion researchers present a breakthrough in the use of computer memory to perform computational operations
• Researchers at the Technion and the Tower Jazz company have developed a technology that adapts commercial transistors to the age of artificial intelligence
• The solution to the energy problem - movement of animals (and humans)
• Technion researchers have developed a platform that accelerates the learning process of artificial intelligence systems 1,000 times