DIPOL Weekly Review – TV and SAT TV, CCTV, WLAN

Training artificial intelligence models on edge devices using a small amount of memory.

Microcontrollers, or miniature computers that can execute simple commands, are the basis of billions of devices: from the Internet of Things (IoT) to sensors used in cars. However, low-cost, low-power microcontrollers have very limited memory and no operating system. All of this makes it challenging to train artificial intelligence models on "edge devices" that operate independently of central computing resources. Training a machine learning model on a smart edge device enables it to adapt to new data and make better predictions. For example, training a model on a smart keyboard can enable it to continuously learn based on how the user types. However, the training process requires so much memory that it is usually carried out using powerful computers in the data center before the model is deployed on the device. This is more costly and raises privacy issues due to the need to transfer data to a central server.

To solve this problem, researchers at MIT and the MIT-IBM Watson AI Lab have developed a new technique that allows training on a device using less than a quarter of a megabyte of memory. Other training solutions for networked devices can use more than 500 megabytes of memory, far exceeding the 256-kilobyte capacity of most microcontrollers (there are 1024 kilobytes in one megabyte). The framework and smart algorithms that the researchers have developed reduce the amount of computation needed to train the model, making the process faster and more memory-efficient. This technique can be used to train a machine learning model on a microcontroller in minutes. It ensures privacy by storing data on the device, which can be particularly beneficial for sensitive data, such as in medical applications. It can also allow the model to be customized to the users' needs.

A popular type of machine learning model is the neural network. These models, loosely based on the human brain, contain layers of interconnected nodes or neurons that process data to perform a task such as recognizing people in photos. The model must first be trained, which means showing it millions of examples so it can choose the correct solutions on its own. As the model learns, it increases or decreases the strength of the connections between neurons, which are called weights. The model can go through hundreds of updates as it learns, and intermediate activations must be stored during each round. Since there can be millions of weights and activations, training a model requires much more memory than running a pre-trained model.

The researchers tested their framework by training a computer vision model to detect people in pictures. After just 10 minutes of training, it learned to successfully perform the task. The optimization required only 157 kilobytes of memory to train the machine learning model on a microcontroller, while other techniques designed for lightweight training would still need 300 to 600 megabytes. With this method, the model could be trained more than 20 times faster. Now the researchers want to apply the technique to language models and different types of data, such as time series data. At the same time, they want to use what they have learned to reduce the size of larger models without losing accuracy.

Operation of CCTV cameras by the monitor of Hikvision IP video intercom system.

A camera built into the video intercom is usually unable to be aimed at the desired point, and its field of vision may be obscured by the caller. A solution to this may be the use of cameras from the existing surveillance system and adding them to the video intercom monitor. The integration is possible by adding:

Up to 16 IP cameras can be added to the monitor. It can decode images from cameras with a maximum resolution of 720p with H.264 compression set. It does not mean, however, that the monitor cannot be equipped with cameras of higher resolution. For example, if an 8 MP camera is added where the main stream is set to the maximum resolution and H.265 compression, and the auxiliary stream is set to 640 x 480 resolution with H.264 compression, the monitor will automatically decode the auxiliary stream. If the main stream was downgraded to 720p resolution and H.264 compression was enabled, the monitor would decode the main stream. Adding cameras is possible directly at the monitor level , or using the iVMS-4200 client application. To do this, go to the remote settings and then: Remote settings->Intercom->IP camera details.

SMATV system receiving TV signals from two DVB-T2 transmitters.

The main task of the installer was to perform a SMATV system to receive 5 DVB-T/DVB-T2 multiplexes from two transmitters: 4 multiplexes from a local 100 kW transmitter (a distance of 12 km), 1 multiplex from a 50 kW transmitter 50 km away, and then distribution of the signal in a TV installation of about 70 receivers. For this purpose, an antenna set consisting of two DIPOL SMART antennas was used: Horizon A2230, City A2010 and Terra PA420T R82516 channel amplifier.

Due to the short distance of the reception site from the powerful transmitter, a DIPOL SMART CITY A2010 antenna operating in passive mode was used to receive channels from the no. 1 transmitter. The DIPOL SMART Horizon A2230 antenna operating in active mode was directed to the distant transmitter no. 2. The PA420T R82516 channel amplifier was connected to the antenna set configured in that way.

You can read more about it here

Why use the outdoor twisted pair cable.

Outdoor telecom cables usually have a sheath made of polyethylene (PE). This material is tougher and has greater mechanical strength than the polyvinyl chloride (PVC) used in indoor cables. What is also important, this material is much more resistant to UV rays, so the lifespan of outdoor cables exposed to sunlight is much longer compared to indoor cables operating under the same conditions.

The video below shows typical problem of using PVC-coated cable to connect an outdoor antenna with a router. The cable laid along the facade, exposed to high and low temperatures, as well as to sunlight and rain, cracks after several years at the latest. Afterwards, water gets inside, and after a short time, due to gravitational force, it "flows" down to the router. This may of course damage of the router itself and, in case of large accumulation of water in the area of the router, damage of other elements, such as furniture.

DIPOL offers the following types of outdoor twisted pair cables:

10 Gbps via one optical fiber.

Internet providers who build their networks over a larger area or partially rely on the infrastructure of larger operators are interested in optimal use of optical fibers. Here, the SFP modules, using one instead of two fibers for transmission, are of help.

The speed of 10 Gbps has become a standard in backbone links even for smaller operators. The SFP+ Ultimode L1430 and L1432 modules make it possible to set up such a link using one single-mode fiber. The modules operate at wavelengths of 1270 nm and 1330 nm, i.e. the second transmission window, the safest in terms of any installation issues (poor splices, fiber bends, etc.).

How to power a camera and 12 V microphone using PoE? Numerous systems are faced with the problem of powering an IP camera via PoE from one UTP cable and another device requiring 12 VDC supply (e.g. an external microphone or an infrared illuminator). This can be done with any PoE (802.3af or PASSIVE) switch and using one output and the PoE Atte ASDC-12-124-HS M18953 stabilizer...>>>more