REBUILDING THE BRAIN: ENGINEERING NEUROMORPHIC PROCESSING

Transcription

1 Conference Section B12 Paper #170 Disclaimer This paper partially fulfills a writing requirement for first year (freshman) engineering students at the University of Pittsburgh Swanson School of Engineering. This paper is a student, not a professional, paper. This paper is based on publicly available information and may not provide complete analyses of all relevant data. If this paper is used for any purpose other than these authors partial fulfillment of a writing requirement for first year (freshman) engineering students at the University of Pittsburgh Swanson School of Engineering, the user does so at his or her own risk. REBUILDING THE BRAIN: ENGINEERING NEUROMORPHIC PROCESSING Jonathan Coles, jmc307@pitt.edu, Sanchez 5:00,, drv7@pitt.edu, Mahboobin 10:00 Abstract We will discuss the societal impact of neuromorphic technology, a programming method that mimics human thought with silicon microchips. With radically higher processing speeds, these microchips, called neuromorphic chips, greatly overpower digital computers. It is likely that if they are perfected, brain-mimicking microchips will be found in nearly every computational device in the near future. The basis of neuromorphic chip technology is mimicking neural function using silicon. The chips are comprised of millions of neurons that emulate the human brain s processing capability. The chips basic composition is that of today s microprocessors with additional silicon neurons, which work in an analogous manner to process information. The newest front of neuromorphic engineering is IBM s chip, TrueNorth. This new chip has approximately a million silicon neurons, similar to the brain of a bee. Because of this, it is capable of executing upwards of 40 million operations a second for every watt of energy it consumes, a vast improvement over today s digital computers. The processing power comes from the neural networks created by silicon circuits within the chip, which is about the size of a postage stamp. TrueNorth integrates pattern recognition and memory, the core of artificial intelligence. Our paper will elaborate on how neuromorphic engineering will have a strong global impact on technology and society. We will use examples of current research, detailed technological description, and rational assessment of ethics to ensure the determination of neuromorphic engineering s possibility and potential. Key words artificial intelligence, biomimicry, computer science, electrical engineering, neuromorphic engineering, Google, processing, Qualcomm. MODERN PROCESSING AND NEURAL BIOMIMICRY Within the last century, computers have become an integral part of our everyday lives. 2 billion personal computers received frequent use in 2015, with higher amounts of computers being used in 2017 [1]. The functionality of computational processing that is programmed into our laptops and PCs has allowed for our species to take great leaps into historically untouched depths of science and engineering. Our computers have evolved along with our scientific knowledge; despite this we have still been using iterations of the same computer for decades. Although innovations have been made in computer processing by tech giants such as IBM, Intel, and Qualcomm, industry-standard computers are still designed to function using analog number-crunching that reads binary data; information that consists of billions of ones and zeroes. The speed and power of a computer using this form of computation are limited by battery power and random access memory (RAM) storage; the true innovations to digital computation come from innovations to RAM and batteries. Digital processors in modern computers have gotten faster at reading large amounts of numbers, but doing so still takes far too much energy and hardware space. Neuromorphic technology cuts that energy use, and increases speed while reducing processor size. Neuromorphic engineering is a field focused around biomimicry of the mammalian brain. Scientists in this field practice using tiny silicon nodes that mimic the neurons that exist within the brain of an animal. Information can be processed by these neurons organically. Much like our own brain, a chip using neuromorphic technology can process images very quickly, learn and study patterns of information, make informed inductions, and execute millions of operations every second while using very little energy. This technology is reinventing how computers work as a whole, and is doing so by designing solutions to all of modern computational processing s biggest problems. We will describe and illustrate the functionality and design of current neuromorphic research by detailing the technology and construction of neuromorphic chips. We shall simultaneously exemplify that neuromorphic engineering is vital in improving information processing and artificial intelligence. If done correctly, neuromorphic engineering could have massive global implementations in every field of engineering and computer science, and will find its way into the modern computer. WHAT IS NEUROMORPHIC ENGINEERING? To fully comprehend all aspects of neuromorphic engineering, one must have an advanced understanding of not only the development of electronics, but also a deep University of Pittsburgh Swanson School of Engineering Submission Date

2 comprehension of how our brains work. Tom Simonite wrote on this topic in an article for the MIT Technology Review Journal. Brains compute in parallel as the electrically active cells inside them, called neurons, operate simultaneously and unceasingly. Bound into intricate networks by threadlike appendages, neurons influence one another s electrical pulses via connections called synapses. When information flows through a brain, it processes data as a fusillade of spikes that spread through its neurons and synapses [2]. The brain contains approximately two hundred billion neurons. There are approximately one quadrillion synapses in the human brain. Each synapse functions similarly to a microprocessor by passing nervous system signals from one neuron to another. Micro-processing power is based on the total number of petaflops contained within the system. One petaflop is equal to one quadrillion calculations per second. The most powerful supercomputer can process about 16 petaflops of data, while human brains can process over 35 [3]. Neurological processing outweighs the ability of any modern computer. As you read this sentence, billions of neurons are firing in your brain, allowing you to process the information written on this page. Your brain is also simultaneously processing motor control, breathing, organ function, and auditory and visual cognition. There is almost no conscious thought put into any of these processes. Biological and Silicon Systems There are only two effective computing devices in the physical world, digital computers and brains. Both are physical machines; one is built from silicon and metal while the other is made from hydrocarbons and aqueous solutions. They both are able to cipher information. Brains, however, can actually interpret and react to sensory and motor interfaces within their environment whilst preforming cognitive-level tasks [4]. It is almost unbelievable how efficiently the brain is able to compute data. If this efficiency were integrated into normal computers, there would be a huge technological shift. Information processing would be utterly revolutionized. Uncovering the innermost functions of the human brain can lead to these technological advancements. This is the basis of neuromorphic engineering: understanding the computational principles used by the brain to develop technology that mimics biological processes. Neuromorphic engineering, a relatively new subfield of electrical engineering, uses electrical analog circuits to mimic neuro-biological structures present in the nervous system [4]. It is a comprehensive integration of brain function and technology: the frontier of artificial intelligence. Neuromorphic engineering is an integral part of electrical engineering; whose principal undertaking is to propel technology forward. The amalgamation of the body and technology has propelled our civilization forward within the past century. The future of technology lies in the hands of neuromorphic engineers. The most influential neuromorphic engineering professional groups is the Qualcomm Zeroth team. They are dedicated to creating neuromorphic technology. Their goals include biologically inspired learning, enabling devices to see and perceive the world as humans do, and creating and defining a neural processing unit. Instead of the conventional coding technique, preprogramming behaviors and their possible outcomes, the engineers at Qualcomm have created software that adapts to its environment and learns through possible consequences as it processes information. This is a very important feature in their technology because it truly mimics neurological methods of computation. Another pivotal point in the Zeroth technology is that it replicates biological senses and simulates synapses in brain communication. The ability to program these senses comes from mathematical data collected by neuroscientists that depicts neuron behavior. This data is able to precisely characterize when certain neurons send and receive information and then analyze how and where it is processed within the brain. Neurons can be timed through the electrical impulses in each cell s membrane. This innovation in neuroscience has helped inspire innovations of neuromorphic technology. The Qualcomm engineers are striving to create a definite form of neural processing that can be directly implemented into society [5]. The high caliber of work that these engineers are doing creates a precedent for future data processing. This research is an important stepping stone for all future neuromorphic engineering endeavors. A relatively recent invention, the neuromorphic chip, integrates all of this information and implements true neurobiological processing into conventional computing. NEUROMORPHIC CHIPS: THE FUTURE OF INFORMATION PROCESSING Neuromorphic engineers have integrated biological language with computing in order to effectively imitate subconscious brain function. The basis of neuromorphic chip technology is mimicking neural function using silicon. The chips are comprised of millions of neurons that emulate the human brain s processing capability. The chips basic composition is that of today s microprocessors with additional silicon neurons, which work in an analogous manner to processes information. Engineers are trying to imitate brain functionality by building mechanical versions of neurobiological systems. These chips are designed to processes auditory and sensory data and then respond based on qualitative measures. These responses have not been preprogrammed but are actually learned through observation and careful analysis of environmental stimuli. [6]. Generally, the chips are made by using collections of transistors to imitate the electrical spiking behavior of a neuron and then creating silicon synapses between them by 2

3 wiring them together. The chips also contain silicon microcolumns, repeating clumps of neurons that perform certain functions [2]. Programming all of these aspects in a harmonious way is an extremely delicate process. The largest working surface neuromorphic engineers use is smaller than a pinhead and contains hundreds of neurons and thousands of synapses that need to be individually programmed. Inside the Neuromorphic Chip Neuromorphic chips transmit and respond to information sent in spikes of energy. This differs from regular super computers, which operate with continuously varying voltages. These spikes charge only a small fraction of the silicon neuron as they are processed. Conventional processing techniques have to keep every single transmission line at a certain voltage, constantly and continuously [7]. This creates a dramatic difference in the amount of power used for conventional processing. The neuromorphic chips are not endlessly running and changing voltage; they are programmed with the ability to pinpoint and spike certain neural processes at the time they occur. This is wildly more efficient. This difference of energy consumption is one of the most important factors when comparing conventional processing to neuromorphic processing. The newest design for these neuromorphic chips is based on two different technologies: lateral spin valves and memristors. A spin valve is a device that consists of conducting magnetic materials. The electrical resistance between the materials can change depending on the relative alignment of the magnetization between the layers. The lateral spin valves within the chips are tiny magnets connected by metal wires that can switch orientation depending on the spin of the electrons passing through them [7]. They act like a brain s neural network. The way these values are connected throughout the chip creates a framework that information can easily pass through. Then energy is sent throughout this framework in spikes as mentioned above. The spin valves operate at terminal voltages which are measured in millivolts; this is significant again because it is much less energy than conventional microchips. The memristors are two terminal electrical components that link electric charge and magnetic flux. Fundamentally, they act like resistors with memory storage. They are able to regulate the way the lateral spin valves store information. The structure that these two forms of technology create is the ideal environment for the exact kind of processing biological systems do well: analog-data-sensing, cognitive-computing, and associative memory [7]. Memristors and lateral spin valves are still both relatively new in the field of neuromorphic engineering; their development only further pushes the boundaries between biological and synthetic systems. This technology is not the only way that neuromorphic chips are being made. Another method of constructing neuromorphic chips is through programmable neural silicon. Carver Mead, the creator of the first silicon retina, started mimicking ion-flow across a neuron s membrane with electron flow through a transistor s channel. The electrons within the transistor act very similarly to the chemical ion s we have in our neural membrane. This is integrated into the chip s hardware by the creation of versatile ion-channel analog and reconfigurable synaptic connections [8]. This means that the analog emulates a specific range of behaviors based on the arrangement of particular displays or transmissions. The neural silicon is softwired within this analog and assigned unique addresses so that every time a spike occurs the chip actually outputs that specific neuron s address; this process is something that can be found in a human brain as well. These silicon neurons are modeled after biological neurons using transistors and circuits. A circuit is created and converted into an array of silicon neurons. The neurons are linked with transistors throughout the circuit. The purpose of the neuron is to integrate synaptic inputs leading to depolarization, which is how signals can actually travel through the circuit, and to initiate action potential that propagates to the neurons terminals within the transistors. Theses terminals effect the channel current through specific voltages [9]. Because of the capability of the silicon neurons, the terminal voltages are spikes rather than continuous current. This leads to much less energy consumption. The components of the circuit mimic biological processing. FIGURE 1 [11] Example of a silicon neuron within engineered circuit TrueNorth, Pointing towards the Future The newest frontier of neuromorphic engineering is IBM s chip, TrueNorth. The chip is made up of hundreds of circuits working in parallel. The chip contains 5.4 billion transistors but only uses 70 milliwatts of power. By comparison, modern computers have approximately 1.4 billion transistors but use watts of power. The efficiency of this chip is astronomical. It can compute over 46 billion operations a 3

4 second per watt of energy [6]. Its low power consumption could have a serious impact on present technology. Energy consumption could be tremendously decreased if IBM s technology were integrated into conventional computers. The transistors in the chip work similarly to that of the neural networks within the brain. TrueNorth has one million silicon neurons, which is about as complex as the brain of a bee. The neurons are able to signal when certain information passes a specific threshold [6]. They organize this data based on pre-programmed patterns. The difference between this processing technique and conventional computation is that the chip is able to learn from these patterns and integrate new solutions based on its environment. This all comes from the bases of neuromorphic engineering. The chip can hold all of this information, much like a memory, and use it for future tasks. TrueNorth actually integrates pattern recognition with memory storage, which allows the chip to function like many tiny microcolumns, each performing a specific task. This is the core of artificial intelligence. Engineers have created a mind that can process and store information based on the past and the present. TrueNorth, IBM s brain-like microprocessor, has been found to be exceptionally proficient at inference work for deep neural networks. In particular, the chip has demonstrated it s especially good at image recognition, being able accurately classify such data much more efficiently, from an energy perspective, than traditional processor architectures, suggesting new applications in mobile computing, IoT, robotics, autonomous cars, and HPC. [7]. TrueNorth is one of the first truly energy-efficient neuromorphic chips. The neural networks within the chip are much faster at processing information than the computers of today. This technology can propel many aspects of mainstream machinery forward. Processing speed will increase and power consumption will decrease through the implementation of this technology. Implementing neuromorphic concepts will lessen the gap between humans and machines. This is a slightly frightening idea. Do we really want to bridge this divide? Will creating artificial intelligence hinder human progress? These are ethical questions neuromorphic engineers must face when completing their research. There is still a long way to go before completely replicating the human brain, but neuromorphic engineering is a step towards the future. FIGURE 2 [10] Example of a TrueNorth chip core FUTURE IMPROVEMENTS ON CURRENT COMPUTING STANDARDS Modern computers function using digital processing, a method that reads binary data in ones and zeroes. These ones and zeroes signal on or off to the computer, informing the processor of the details of the data (such as pixel arrangement or numerical value.) This is the language of modern digital computers. A byte of information consists of eight digits of binary code, for example: Spelling out the word porridge in binary code is done with the following string of ones and zeroes: Most images on our computer consist of several megabytes of binary code; one megabyte of binary code is approximately eight-million digits. Programs and applications can have several gigabytes of information; upwards of eight-billion ones and zeroes. Some hard-drives can store a terabyte of information, or eight-trillion binary digits. If an image has three megabytes of information, a computer has to process twenty-four-million ones and zeroes to be able to display the image. Modern computers are capable of reading this amount of information in a few milliseconds, but doing so still takes a lot more energy and time than our brains would take to analyze the same image. Neuromorphic engineering rethinks the modern approach to information processing. In modelling computer chips from the mammalian brain, neuromorphic engineers are designing new chips that will conquer the current standard of large data and high energy usage [10]. These chips do this by mimicking organic neural synapses with silicon nodes. Electrical current is transferred very quickly between the fabricated neural net of the neuromorphic chips, producing an emulation of mammalian thought and reaction [11]. This is a much faster process than crunching billions of numbers to process an image. Computers with neuromorphic chips can be programmed to be capable of human-level vision and speech functions that use microphones and cameras. These features 4

5 would make computer image and sound recognition function at the same level that human sight and listening occur [12]. The main innovation made by neuromorphic engineers comes from the shift in thought of the processing ideology. Instead of encoding data into machine-dictated strings of countless numbers, the neuromorphic chips are designed to read data as a human or animal would. These chips are similar to an animal s brain in more ways than one; neuromorphic processing is more energy-efficient than digital processing. This increase in efficiency reduces the need for large batteries, as less energy is needed in order for the computer to operate. Smaller batteries and greater energy efficiency will improve the environmental sustainability of our computers, as less electricity, battery acids and metals will be used to fabricate and operate the billions of computers in use today. In addition, neuromorphic chips are also much faster and more powerful than digital processors, just like our brains [2]. This increase in speed and power would diminish the need for high amounts of RAM and memory storage, as neuromorphic chips would process data with greater speed and energy efficiency; this would also greatly decrease the size of the hardware and alleviate the environmental strain of manufacturing large computers. Societal Impact If modern computers were to be built using neuromorphic chip processors, members of every engineering field, as well as all computer-using consumers, would experience a strong upgrade in available technology. New computers built with neuromorphic chips would feature battery life nearly twice as long as the current standard, smaller devices, more ergonomic hardware design, 3000% faster processing times, and similarly higher processing power [12]. With all of these innovations, new software and operating systems would be able to feature historically unprecedented functionality such as speech and sight [12], because of the staggeringly high processing speed and power of the neuromorphic chips. A computer that uses neuromorphic processing would function faster and stronger than a digital computer while using a fraction of the energy. This would cut the massive amounts of energy used by the worlds 2 billion computers in half. The largest, most expensive digital computers used by engineers could be replaced by the smallest of laptops that use neuromorphic processing, as the two would be equal in speed and power. Software developers would be able to create programs with functionalities that are currently only dreamt of, as they would not be nearly as restricted by RAM and processing speed. The possibility of software innovations occurring as a result of neuromorphic engineering is highlighted in Jeffrey Burt s article on the future of memristor research for Next Platform magazine: Much of the talk around artificial intelligence these days focuses on software efforts various algorithms and neural networks and such hardware devices as custom ASICs for those neural networks and chips like GPUs and FPGAs that can help the development of reprogrammable systems. [13] With faster development of programs that can function on less energy and power restriction, the average quality and capability of software would greatly increase. Glitches and drops in frame rate would not occur. Programs would run on less energy, making our computers more sustainable. Applications could be developed that would predict and evaluate real world situations, and make the engineers job much easier. Engineers could get a lot more done with less stress, and do better work [12]. Here is an example: when using Google SketchUp, a popular 3D modelling program used by many architects and civil engineers, my 2015 MacBook Pro was freezing every couple seconds. The more details I add to my model, the more frequently I experienced glitches. This is because my computer s processor was struggling to analyze the billions of zeroes and ones that it was being fed by the program. With a neuromorphic processor, the program could function much more simply. My computer would be able to comprehend the needs of a program the same way that our brains would. The program would communicate with the silicon neurons of the neuromorphic chip, and interpret my interactions with the keyboard and mouse. It would be able to organically understand what I was trying to do; how I was trying to edit my model, and it would do just that for me. The neuromorphic chips can function the same way for other programs. This would make everyday computer use much simpler and faster once completed neuromorphic processors are implemented into commercial computers. Market and Industry Sustainability With nearly a decade of research on neuromorphic technology, developers are nearing completion of marketready processors. The efficiency of modern computers has been rising slowly for many years, and the technology industry is in need of a big leap. The switch to neuromorphic computing from the current standard of digital computing would be very worthwhile, as upwards of 2 billion computers would use less energy, and function up to 3000 times faster [12]. The digital processing computers that would become obsolete would create a lot of waste after being replaced by neuromorphic computers; this would put strain on the environment. Unfortunately for the earth, they would be replaced anyway in 2-4 years, which is the common lifecycle of a computer [14]. Many computers would have already gone through their lifecycles by this time; implementing neuromorphic chips would simply replace them with much better technology. When neuromorphic chips are developed to be industryready, computer companies such as Apple, Microsoft, and Dell must be prepared to make the switch from digital to neuromorphic. And likely, they will be, as implementing this new technology will greatly increase the performance of their products. Most Dell and Apple computers already use digital processors developed by the same technology companies that 5

6 are leaders in neuromorphic engineering, such as Qualcomm and Intel. The computer companies can simply replace the digital processors in their designs with neuromorphic processors made by the same company, making the switch very simple. The future market implementation will be dictated by how the companies that produce the neuromorphic processors market their products to the companies producing the computers. Research and development is set to be concluded on the first round of market ready chips by 2026; the global market for neuromorphic chips is predicted to grow to 10 billion dollars by then [15]. This means that by 2026, 10 billion dollars must be spent in order for these chips to have a prominent presence in the market. This seems expensive until compared to the 2.9-trillion-dollar size of the current global technology market. A 10-billion-dollar investment is very small in the grand scheme of the technology industry. This investment would lead to massive improvements to the overall functionality and environmental sustainability of our modern technology, diminishing the energy use of 2 billion computers by almost 50%. The market implementation of neuromorphic chips would not only revolutionize computer processing, but would also be a fantastic victory for global environmentalists. CONCLUSION The possibilities of neuromorphic technology elaborate the prospect of the next great revelation in computer and electrical engineering. Neuromorphic processing presents a rise in computer capability that is historically unrivalled, and could lead way to revolutionary changes to our everyday lives. The concept of organic interaction with our computers gives fuel to the hope of creating a truly capable form of artificial intelligence, and further underlines the true possibilities that could arise from neuromorphic technology. Silicon synapse technology used in neuromorphic processors is the greatest opportunity the human race has had thus far to move past the centuries-old method of binary computing, and head into a new age of technological advancement. With robust improvements to energy usage, and potential for massive increases on the functionality of modern software, neuromorphic technology is in a position to make the world s 2 billion computers run faster, and be more environmentallyfriendly. It is set to take over the insides of our beloved laptops and cellphones within a decade. SOURCES [1] How Many Computers Are in The World? Worldometers.info. Accessed [2] Simonite, T. Thinking in Silicon. MIT Technology Review Accessed [3] Webster, S. Earth s supercomputing power surpasses human brain three times over. Rawstory Accessed [4] Giacomo,Indiveri. Horiuchi, Timothy K. Frontiers in Neuromorphic Engineering. Frontiers in Neuroscience. Date of Publication 03/10/2011. Accessed /full [5] Introducing Qualcomm Zeroth Processors: Brain Inspired Computing. Qualcomm. Date of Publication 03/10/2013. Date Accessed ng-qualcomm-zeroth-processors-brain-inspired-computing [6] Hof, R. Neuromorphic Chips. MIT Technology Review Accessed chips/ [7] Intel Reveals Neuromorphic Chip Design. MIT Technology Review Date Accessed [8] Programmable Neural Silicon. Brains in Silicon, Stanford University Date Accessed [9] Designing and Testing Neuromorphic Chips. Brains in Silicon, Stanford University Date Accessed 3/23/ [10] Markoff, John. IBM Develops a New Chip That Functions Like a Brain. The New York Times. Date of Publication Date Accessed [11] Feldman, M. IBM Finds Killer App for TrueNorth Neuromorphic Chip. Top 50 List Date Accessed [12] Esser, Steven. Convolutional Networks for Fast, Energy Efficient Neuromorphic Computing. IBM Research Almaden Date Accessed [13] Burt, Jeffrey. Memristor Research Highlights Neuromorphic Device Future. Next Platform Date Accessed [14] Dubash, Manek. The Desktop Lifecycle: How Long is it Anyway? The Register Accessed [15] Global Neuromorphic Chip Market to surpass US$ 10 Billion in Revenues by 2026; Smart Machine Integration and AI Systems Driving Market Growth. Future Market Integration Date Accessed

7 ACKNOWLEDGEMENTS Firstly, we would like to thank Beth Newborg, who did a fantastic job providing feedback as our writing instructor. Also, Alyssa Srock and Mr. Wunderley for guiding us through the writing process as our conference chairs. Of course, we would like to thank our parents, Radisav Vidic, Natasa Vidic, Nicholas Coles, and Jennifer Matesa, all four of whom work for the University of Pittsburgh and have greatly influenced our education. We would not be where we are today if it was not for their amazing guidance and trust. Lastly, we would like to thank our friends. Without them our days would have been much harder and filled with a lot less laughter. 7

8 8