A Fierce reader asked: How do analog sensors work at the edge with AI, since AI functions are digital?
Glad you asked!
So many edge applications are being invented today to take advantage of AI data for better decision making and they often work alongside sensors for inputs. And yes, many sensors are analog. A common analog output in an industrial workplace might be 4-20mA current and 0-10V voltage. Analog sensors are still widely used because they are simple and cost effective.
Often, Al functions at the edge with an analog-to-digital converter (ADC), which translates the sensor’s continuous analog signal (say, temperature or pressure or light) into a digital binary format of 1’s and 0’s that can be easily read by a computer or a microcontroller, including those with AI capabilities like AI PCs that process AI models, often at low-power. Major companies that make ADCs include Analog Devices, Microchip Technology, Texas Instruments, Renesas and STMicroelectronics. The list also includes Cirrus Logic, Ashai Kasei Microdevices and Teledyne e2v.
And yes, there are so-called digital sensors that contain an integrated ADC, which converts the continuous analog signal into a digital value. Digital signals are less susceptible to noise during a transmission and may be more precise or reliable in an application.
In edge processing, digital data is processed in an AI inference model that has already been trained on a large dataset and runs right on the local chip in an edge device in an edge application. (The main advantage of edge operations is mainly to avoid the need to consult a big data center far away.)
You may already understand inference, but in relation to sensing, it might be as simple as detecting an anomaly. For example, an industrial machine that is supposed to run with a steady humming sound might develop an anomaly such as a sudden loud, repetitive noise due to wear and tear. An AI model will have been trained to know the normal operating sound of a machine or similar machines, and will report any variance to what’s normal. Based on that analysis reported out, the edge device could do many things, such as triggering an alarm (for human interaction) or adjusting a setting (turning down the speed of a machine or even turning it off).
The value of edge computing has been covered by Fierce many times and is beginning to be well accepted by developers. Edge computing offers multiple advantages. One, it is valuable to use edge to reduce latency in decision making, especially with robots and safety systems. Waiting for data to travel to the cloud and back with a decision may take too long. Two, processing locally saves energy costs, and is valuable for battery-powered and other smaller devices. Three, local systems can function even if a network connection to the cloud is lost. Four, sensitive data and private data can be processed locally rather than transmitted over a network to improve security.
So, what's new?
New technologies are emerging that include analog AI chips. They perform AI calculations, such as matrix multiplications, directly in the analog domain. That approach means they are even more power-efficient and faster in some workloads. With neuromorphic computing, some analog AI devices mimic the structure and function of the human brain to process information efficiently in areas such as pattern recognition and sensory data.
IEEE Spectrum in June featured a new neuromorphic chip from Dutch firm Innatera that is optimized for AI at the extreme edge of networks. Called Pulsar, it can lower latency to as little as one-one-hundredth of conventional processes and consume just one-five-hundredth the power in AI applications, according to the company.
The Pulsar chip uses a hybrid analog-digital approach with 12 digital cores for spiking neural networks and four analog cores. Developers can pick which cores they want to load into their models based on their needs. Pulsar is designed for ultralow-power AI sensors applications in consumer, industrial and IoT. Think smart doorbells with super long battery life. Innatera is working with Socionext to develop a radar-based sensor that can detect people accurately, even when they are standing perfectly still, based on their body movements as they breathe.
It turns out that the Pulsar chip uses an ADC and a specialized encoder to process analog sensor data. The company says the chip can handle input from a wide variety of traditional sensors. The design also includes a traditional digital signal path and a power-efficient analog signal path. An encoder is used to convert sensor inputs, including from the ADC into the spikes needed for the chip’s core spiking neural network (SNN) engines, according to ElectronicDesign.
One of the biggest concerns with using neuromorphic computing is that developers will face a deep learning curve as they try to run models on such devices. But, hey, that’s the kind of challenge most engineers want, right? And Innatera has a Talamo software development kit to help developers build spiking models in PyTorch.