Fanless Edge system for neural networks
The built-in eBox560-900 system is Axiomtek’s response to the growing popularity of graphics processing units (GPUs), which have a parallel structure and can handle a large amount of simple computational operations at the same time.
Develop applications with ease
Axiomtek takes precedence in applications where the number of simultaneous computational operations is too large and at the same time they are too monotonic for the central processing unit (CPU).
In addition to the gaming world, rendering and non-negligible mining operations, the need for parallel computing is present in artificial intelligence (AI) systems, especially in machine deep learning. The latter areas of artificial intelligence deal with neural networks. You can read more about neural networks at the bottom of the page.
Use a robust embedded system
The compact and robust design of the eBOX560-900, compliant with IP40 protection, operates in the extended temperature range from -10 °C to 50 °C and is resistant to vibration up to 3G, and complies with CE and FCC Class A certification. DIN rail or wall mounting.
The computer has USB, two Ethernet and HDMI output that supports 4 K resolution. There is room for a PCIE Mini expansion card, for WLAN and Bluetooth or a 3G/4G/LTE/GPRS module, which provides wireless connectivity via the option of connecting four external antennas.
Exploit amazing development tools
In terms of software, the computer is supported by the Linux 16.04 operating system. The Jetson Development Pack (JetPack) SDK is available to developers , which contains all the necessary tools for application development. JetPack includes machine learning tools, machine vision, GPU calculations, and multimedia.
JetPack tools:
- Deep Learning: TensorRT, cuDNN, NVIDIA DIGITS™ Workflow
- Computer Vision: NVIDIA VisionWorks, OpenCV
- GPU Compute: NVIDIA CUDA, CUDA Libraries
- Multimedia: ISP Support, Camera imaging, Video CODEC
Teach your machines
It is common for people to learn how machines work and how to operate them. If the pieces from the production line do not meet the quality standard, the operator corrects the machine and tool parameters according to experience. Or if the car body part is not well painted, additional human work is required, which again, according to experience, corrects the defective work of the robot.
With Industry 4.0, where processes are digitized, AI does the machine management. As the concept of machine learning suggests, machines themselves understand processes and adjust their parameters according to their surroundings.
Implement neural networks
Neural network is a computational model for parallel information processing, where the basic building blocks of the network are artificial neurons or threshold functions. Networks have multiple inputs and one or more outputs. The inputs are weighted differently and are connected to the output via one or more levels – one level is shown in the sketch.
The mathematical model of a neuron sums the signals (w n ) of differently weighted inputs (x n ). If the sum exceeds a certain threshold, the neuron sends a signal according to the activation function. If the network is multilevel, and usually is, the transmitted signal is received by another neuron in the next layer, which also processes it.
Meet the elements of a neural network
- Synaptic weights: Wn values are weights associated with a single neuron that determines the strength of the input vector Xn,
- The neuron threshold is the value at which the output is activated,
- Activation or a transfer function, which can be of various forms of functions (sigmoid in the figure), performs a mathematical operation on the output signal,
- The grid output is the weighted sum of the inputs.
Teach the network how to react to a pattern
The weights of neurons and the threshold at which a neuron emits a signal are formed by learning. In learning mode, therefore, we teach the network how to react to a particular pattern of inputs. We use a large database of inputs for learning and determine the desired output. To get the desired result, the learning network searches for optimal weight values and thresholds.
Operations with neural networks are performed in parallel, as each neuron operates relatively independently of the others, so neural networks are able to adapt to a complex environment, with many input parameters, in real time and with minimal delays.
Want to know more?
Fill out the contact form and we will get back to you shortly.
Don’t miss any news.
Subscribe to our newsletter!
