2014 New Technology: Neuromorphic Chip

A little robot named Pioneer slowly walked over to a carpet with "Captain America." Outside, a group of people around it looked at it. Here is the scene of Qualcomm's experimental test simulation. Simulating a child's bedroom. Pioneer paused, as if measuring the surrounding environment, then rolled up the carpet like a snowball, turned around, and prepared to put it in a toy box. Qualcomm's senior engineer, Ilwoo Chang, swings his hands toward the Pioneer Robot and signals that it should not be placed here. Pioneer looked at this gesture with its camera and then accepted the order obediently. Then it will return to its position, put down the carpet and find another Spider-Man doll. This time, the Pioneer robot rushed directly to the doll, disregarding the surrounding chess, and put the toy in the box without anyone's guidance.

This demonstration took place at Qualcomm's headquarters in San Diego. Although it was a demonstration, it could see the shadow of future calculations. The robots needed powerful computing power to handle their missions. In the past, these calculations were performed on large-scale equipment that consumed more electricity. The Pioneer robot uses only a smart phone chip and runs specially-made software. It can recognize objects that have not been seen before, classify them according to the similarity of related objects, and place different items in the correct position of the room. Tiring programming, you only need to show it the object and location. This small robot can do this because it simulates the state of human brain work.

Later this year, Qualcomm will reveal how science and technology are embedded in everyday electronic device chips. These "neuromorphic" chips (so named because they mimic the brain) are designed to process sensor data (images, sounds, etc.) and respond to changes in unprogrammed data. These chips have fulfilled the tasks that the field of artificial intelligence can take decades to accomplish, allowing machines to understand the world and interact with the world as humans do. Medical sensors and devices can track the patient's vital signs, take medical measures according to time, learn to adjust the dose, and even find the condition as soon as possible. The smartphone can learn the text you will enter next, and then set reminders in the background. Google’s self-driving car may no longer require the driver’s help. Sweeping robots will never get stuck under the sofa. Qualcomm Technical Director Matthew Grob said: "We are blurring the gap between the chip and the biological system."

Qualcomm's chips will not be available until 2015. This year, Qualcomm will issue chips to researchers for testing. As long as a shipment of Qualcomm, "zero plan" will be the first commercial platform for large-scale neural model computing. Qualcomm will also implement these capabilities before many universities and research institutes, such as IBM Labs and HRL Labs. The two companies have spent 100 million U.S. dollars to develop neuromorphic chips for the U.S. Department of Defense Advanced Research Projects Agency. In addition, researchers of the European Brain Project joint University of Heidelberg and the University of Manchester also spent 100 million euros to study the neuromorphic project. Another German group recently reported that the use of neuromorphic chips and software to simulate the odor processing system of insects can be used to judge plant species by smelling flowers.

Today's computers use the so-called "von Neumann architecture", which means that data is calculated back and forth between the processor and memory. This method is very suitable for digital operations and execution of well-written programs, but it is not suitable for processing images, sounds, and other perceivable content. In 2012, Google’s artificial intelligence software successfully identified cats with up to 16,000 processors without telling them what a “cat” was.

In order to increase the performance of these processors, manufacturers have added more transistors, buffers, and data channels to the chips. However, the temperature at which the chips operate is limited by the operating speed of the chips. In the mobile field, the power supply is required. More stringent. This limits the device's ability to effectively process images, sounds, and other sensory data, and it cannot be used to perform facial recognition or robot and vehicle navigation.

In addition to Qualcomm, no one can better solve the physical challenges faced by these chips. The demand for performance of mobile devices is increasing. But today's personal assistants, such as Apple's Siri, Google's Google Now, and so on, have limited functionality. Because the device's performance is not enough, data must be sent to the cloud for processing. Jeff Gehlhaar, vice president of Qualcomm Technologies, who led Zero Plan, said: "We just want to hard hit."

The human brain has billions of neurons and hundreds of billions of synapses that can simultaneously process visual and audio signals. The neuromorphic chip simulates the ability of the human brain to process multiple data simultaneously in the chip. Depending on changes in images, sounds, or other signals, neurons can change the connections with other neurons. This process is called learning. These neuromorphic chips mimic the neural network of the human brain and can achieve some of the functions of the human brain. This is why Qualcomm's robot can put it in accurate position before it knows who Spiderman is.

Qualcomm can add "neural processing units" to smart phone chips to help process sensor data, such as image recognition.

Even if the capabilities of a neuromorphic chip are far from the human brain, its ability to process sensory data and learn data changes is faster than any computer. Jeff Hawkins, founder of Numenta, a brain-inspired software company, said that if a traditional computer is used to run a special computer to simulate a human brain like Google, the efficiency of such a smart device is too low. “Artificial intelligence cannot be achieved by software. Need to use chips to complete."

Nerve channel

The popularity of smart phones also made Qualcomm feel emboldened, and Qualcomm’s current market value has surpassed Intel’s. The reason for this may be related to several hundred wireless communications patents that Qualcomm has mastered. Now Qualcomm is ready for the next round of breakthroughs. The first step was to collaborate with Brain Corp. and Qualcomm invested in a neuroscience startup with outstanding achievements in simulating human brain computing. The "zero project" at the beginning of the project is to serve the robot so that the robot can interact with the real world, and then it can be used on smart phones or other products. "Zero Project" from Isaac? Asimov’s zeroth robot rule: Robots must not harm the human whole, nor should they cause harm to humanity as a whole without taking action.

The concept of neuromorphic chips dates back decades. In 1990, Carver Mead, Emeritus Professor of California Institute of Technology, gave its definition in a paper. “Analog chips are different from digital chips that only have binary results (on/off) and can be different from the real world. As a result, the brain's neurons and synapse's electronic activity can be simulated." But he couldn't finish the task of designing this analog chip. Only one company, Audience, has developed a controversial neuromorphic chip that can suppress noise and has sold hundreds of millions of tablets. These chips are designed based on the human cochlea and can isolate noise like humans. They only listen to human voices and have been used on mobile phones such as Apple and Samsung.

As a commercial company, Qualcomm hopes that product design is more practical than performance. This means that Qualcomm’s neuromorphic chips are still being developed on digital chips. This is easier than researching and developing analog chips and production is easier. The simulation chip is to complete the simulation of the brain, and Qualcomm's chip simulates the behavior of the brain. For example, neuromorphic chips program and transmit data in a manner that mimics the electronic pulses that the brain processes when processing sensory data. M.Anthony Lewis, a “zero project” engineer, said: “Even if we do it digitally, we can already replicate many of the behaviors of the brain.”

This batch of chips is also very consistent with Qualcomm's existing business, although Qualcomm has now occupied the position of the boss in the smart phone field, but income growth began to slow down. Qualcomm Snapdragon mobile processors can join "neural processing units" to handle sensory data, image recognition, and robot navigation tasks. Because Qualcomm generates a large portion of its revenue from licensing patents to other companies, it can also sell its use of neuromorphic chip algorithms to these companies. This kind of chip can be used in other applications such as vision and motion control.