Intel Loihi processors contain electronic devices that function very much like neurons. Despite their names, neural networks have little to do with what you find in your brain. While their organization and how data is moved across processing layers may be very similar to real neural networks, the data and computations that are performed on them look very familiar to a standard CPU.
On Thursday, Intel released the latest version of its symmetric neural hardware version called Loihi. The new version includes what you'd expect from Intel: a better processor and some basic computing improvements. But it also has some fundamental hardware changes that allow it to run entirely new classes of algorithms. And while Loihi is currently a research-focused product, Intel is also releasing a compiler that it hopes will be widely accepted. Start by taking a look at a little neurobiology and then build from there.
From Neurons to Computation
The basis of the nervous system is a type of cell called a neuron. All neurons have many functional characteristics in common. At one end of the cell is a structure called dendrites, which you can think of as receptors. This is where neurons receive input from other cells. Neurons also have axons that act as transmitters and communicate with other cells to transmit signals. The cell membranes of spinal neurons descend from the axons until they reach the meeting point of other cells (synapses), at which time they are converted into a chemical signal that is transmitted to nearby dendrites. This chemical signal opens channels that allow ions to flow into the cell, causing a new mutation to occur in the recipient cell. Show that he should be calm, how active he has been in the past, etc. - and use it to determine his state of activity. When a threshold is exceeded, they rapidly increase their axons and activate other cells.
Normally, when neurons are not receiving, activities are sporadic and random. High Input However, when it starts receiving the signal, it activates and fires a group of groups one by one. The axon (a long extension at the bottom right) is visible. "src=" https://safirsoft.com/picsbody/2109/10669-1.jpg "alt =" https://safirsoft.com Understanding Neural Computing and Why Intel is so excited about it "srcset =" https://cdn.arstechnica.net/wp-content/uploads /2021/09/20130314-neuron-640x640.gif 2x "> Zoom in/A neuron, with dendrites (raised protrusions on top) and part of the axon (long extension at lower right) visible NIH
How does this process encode information? This is an interesting and important question, and we have just answered it.
The answer we came up with is the so-called theoretical neurobiology (or computational neurobiology), which involved trying to create mathematical models that reflect the behavior of the nervous system and neurons , hopefully that will lead to our ability to identify some basic principles: Neural networks that focus on the organizational principles of the nervous system have been one of the efforts that have come out of this field: augmenting neural networks, which attempt to measure the behavior of individual neurons to be created, one d It is another thing.
Augmentation of neural networks in software can be implemented on traditional processors It is also possible to implement it through hardware, as Intel is doing with Loihi. The result is a processor completely different from any you may be familiar with. Each of these nuclei contains a large number of "neurons" or executive units. Each of these neurons can receive inputs in groups from any other neuron - a neighbor in one nucleus, a unit in a different nucleus in one chip, or entirely from another chip. Neurons integrate the bumps they receive over time and use them to decide when to send their clusters to each associated neuron, based on behavior programmed with them.
All spike signals occur asynchronously at specified intervals, with x86 cores embedded in a synchronous chip. At that time, the neuron changes the weight of its various connections again - essentially, the amount of attention to pay attention to all the individual neurons sending the signal.
As a real neuron, part of the execution unit on the chip acts as dendrites, partially processing the signals received from the communication network based on the weight of the previous behavior. Then a mathematical formula was used to determine when an activity exceeds a critical threshold and to generate its own mutations when used. Then Axon executes the executive unit, which communicates with it and sends a hill to each of them. A neuron is recorded only when it receives a neuron.
Unlike a normal processor, there is no external RAM. Instead, each neuron has a small memory to use. This includes the weights assigned to the different neuron inputs, the storage of recent activity, and a list of other neurons to which the spike is sent.
Another big difference between neural-shaped chips and conventional processors is their energy efficiency, as neural-shaped chips are appearing in the future. IBM, which introduced its TrueNorth chip in 2014, was able to do useful work from it, even though its timing was at 1 kHz, and it used less than .0001% of the power needed to simulate an acute neural network. In traditional processors, Mike Davis, director of Intel's Neural Computing Lab, said Loihi processors process 2,000 times more in certain workloads. "We typically find 100 times less power for SLAM and other robotic loads," he added.
Understand Neural Computing and why Intel is excited about it
to explore faraway worlds, now's the time. The team behind the constell...
Welcome to version 4.21 of the Missile Report! There's a lot of news this wee...