FEATURED

IBM has adapted a convolutional neural network to work on a neuromorphic chip

According to IBM, TrueNorth accuracy corresponds to the best modern systems of image and voice recognition, but the system consumes less energy and works faster. The team of researchers of the company is sure that the combination of convolutional networks with neurromorphic microcircuits will allow to create more sophisticated smart cars and smartphones that correctly recognize a person's voice command even if he speaks with his mouth full. Let's try to figure out what the advantages and disadvantages of TrueNorth are, and where it found application.

The human brain contains about 86 billion neurons – cells that connect with thousands of other neurons with the help of processes – synapses. The neuron receives signals from many others, and when the stimulation reaches a certain threshold, it "activates", sending its own signal to the surrounding neurons. The brain learns, in particular, by adjusting strong ties. When the sequence of actions is repeated, for example, through practice, the concomitant synapses become stronger, and the resulting lesson or skill "fits" into the network.

In the 1940s, scientists began to model neurons mathematically, and in the 50s – to create Network of neurons and computers. Artificial neurons and synapses are much simpler than in the brain, but act on the same principles. The set of simple units – neurons – connects to others through "synapses" with their numerical values ​​depending on the values ​​of units.

Convolutional neural networks (CNN) is a special type of network that has gained popularity in recent some years. CNN extract important functions from incentives – usually photographs. Take, for example, a photo of a dog. It can be represented as a layer of neurons, where the activation of one neuron represents one pixel in the image. On the next layer, each neuron will receive input from the set of the first layer and activate if it detects a certain pattern in this set, acting as a sort of filter.

In the next layers, neurons will look for patterns in patterns, and so on. Within a single layer, filters can be sensitive to certain structures. First to the boundaries of the figures, then to the paws, then to the dogs, until the network can determine the difference between the dog and the toaster.

Now such calculations are expensive. Given that there are billions of neurons and trillions of synapses in the human brain, the imitation of each is not yet possible. Even simulating a small part of the brain will require millions of calculations for each input element, which requires enormous computing power. The largest modern CNN can be millions of neurons and billions of synapses, but no more.

The classical computing architecture of central processors, designed to process one instruction at a time, is not suitable for performing the tasks that CNN imposes. Therefore, scientists turned to parallel calculations, which can handle several at once. Modern neural networks use graphics processors, which usually calculate the graphics of video games and CAD. Because of the architecture and the similarity of mathematical calculations, video cards are suitable for in-depth training.

But still the hardware does not do as much with deep training as a brain that can drive a car and simultaneously talk about the future of autonomous vehicles , While consuming less energy than a lightbulb.

In the 1980s, engineer Carver Mead coined the term "neurromorphic processors" to describe computer chips that work in a sense in terms of upgrade equ brain. His work laid the foundation for this field. Although the term "neuromorphic" is now applied to a wide range of solutions, they all try to replicate the mechanism of operation of neural networks at the hardware level, avoiding the bottlenecks faced by traditional processors.

Having seen the need for rapid and efficient machine learning, The Office of Advanced Projects of the US Department of Defense (better known by the acronym DARPA) has been actively funding the IBM HRL Laboratories corporate laboratory since 2008 to develop neurromorphic machines that can be easily scaled Vat.

TrueNorth

In 2014, IBM introduced its TrueNorth chip on the cover of the journal Science. Since then, the company has been developing systems based on TrueNorth with the financial support of the US Department of Energy, the Air Force and the Army. One such chip contains a million "neurons", each of which is represented by a group of digital transistors, and 256 million "synapses" – wired connections between the chips.

Neuromorphic architectures become more efficient in comparison with conventional chips due to two functions. First, such a chip, like the brain, communicates through "flashes" – one-dimensional packets of information sent from one neuron to the descending neurons. The signals are simple (there is a flash or not) and are transmitted only occasionally when the neuron transmits the packet. Secondly, as well as in the brain, processing and memory are located in neurons and synapses. On a traditional computer, the data processing unit continually extracts information from individual memory areas, performs operations, and then returns new information to the memory. This leads to a lot of slow and energy-consuming activities.

The TrueNorth system is flexible enough because it can be programmed to implement networks of different sizes and shapes and to scale by "splitting" several chips. In its scientific work, the IBM team used a neurromorphic chip to identify people, bicycles and cars in a video shot on the street. A comparative experiment showed that TrueNorth software running on a traditional microprocessor used 176,000 times more energy for this task.

A key part of the IBM project was the creation of not only a chip, but also software. The company has created a simulator, a new programming language and a library of algorithms and applications. The company then provided these tools to more than 160 researchers in academic, national and corporate laboratories. The TrueNorth design was completed in 2011, and the revolution of convolutional neural networks occurred in 2012 as part of the ImageNet Challenge. Therefore, some people began to wonder if TrueNorth chips could handle these networks.

CNN uses the method of back propagation of the error. Every time the network is mistaken, the difference between its assumption and the correct answer is calculated. The error back propagation algorithm considers each neuron in the final layer and calculates how much the change in the output of this neuron will reduce the overall error. Then he returns to the previous neurons and calculates how much the change in the strength of each incoming synapse will reduce the overall error.

It is necessary to find out whether to increase or decrease the synaptic force, so the algorithm slightly adjusts each weight in the right direction. Subsequently, the algorithm calculates a new error using the new weights and repeats the entire process. After many such steps, the error decreases in a process called a gradient descent.

Initially, TrueNorth was considered incompatible with the error back propagation algorithm, because gradient descent requires making tiny adjustments to the weights and seeing tiny improvements. TrueNorth maximizes its effectiveness using only three different weight values: -1, 0 and 1, and the output from the neuron is 0 or 1. No gradients, only discrete steps.

One of the key achievements of the team was a series of methods for Execution of an algorithm for back propagation of an error with impulse neural networks. Researchers solved this problem by teaching the program model of the chip programmed to use hardware approximation that is compatible with the gradient descent.

Another key development was the comparison of CNN with a lot of connections with neurons on the chip, which involves only 256 connections per neuron. This was achieved by assigning certain pairs of neurons that worked simultaneously, which were combined into one neuron through the inputs and outputs.

Despite the rather high performance of TrueNorth, it was created without taking into account the features of deep neural networks and CNN, Other systems he has flaws. For example, to make a network of 30,000 neurons work, you need 8 chips (8 million neurons). In addition, TrueNorth is a fully digital chip, when as some have analog components. Their work is more unpredictable, but still more effective. And although each TrueNorth chip is divided into 4096 "cores" that work in parallel, 256 neurons in each core are updated only sequentially and one at a time.

Serial processing of neurons in the TrueNorth core can create a bottleneck, but it also Provides regularity. And this means that the behavior of the chip can be simulated with high accuracy on desktop computers. At the same time, the chip is universal – it can support many different types of networks, and the current goal of the chip makers team, led by IBM's chief brain scientist Dharmendra Modha, is to deploy several different networks working together to achieve compositionality.

Plans for the future

In addition to achieving compositionality, the Modha team seeks to explore various teaching methods. Also, scientists note that the methods described in their work can be applied to neuromorphic chips, other than TrueNorth. Moreover, in addition to new methods of teaching, the team is also thinking about more radical achievements. According to the US Department of Energy's 2015 report on neuromorphic calculations, currently about 5-15% of the world's energy is consumed in some form of data processing and transmission. Along with this, the Department wants to increase the speed, efficiency and resiliency of networks. This report prompted the IBM team to think about developing materials with new physical properties.

The global goal is to replace giant data centers with chips in smartphones, homes and cars that can "think" for themselves: conduct conversations, do scientific and medical Discover, manage cars, robots or limb prostheses. Ideally, these chips can achieve even greater success, for example, solve the problem of hunger in the world.

Several research laboratories are already actively using TrueNorth. In August 2016, Samsung demonstrated a system that uses a video stream to create a 3D environmental map in three dimensions at a rate of 2000 frames per second, consuming a third of the watts. The company used this system to control TV gestures.

Lawrence Livermore National Laboratory has a payment of 16 TrueNorth chips, which is used to enhance cybersecurity and ensure the security of US nuclear weapons. The Air Force Research Laboratory, which uses TrueNorth to provide autonomous navigation for unmanned aerial vehicles, recently announced plans to test an array of 64 chips.

About the author

admin

Add Comment

Click here to post a comment

Your email address will not be published. Required fields are marked *