“Perception systems can be defined as a machine or Edge device, which has embedded advanced intelligence, which can perceive its surroundings, taking meaningful abstractions out of it and allow itself to take some decisions in real-time,” said Pradeep Sukumaran, VP, AI&Cloud at Ignitarium, at the Machine Learning Developers’ Summit (MLDS) in his talk titled “Hardware Accelerators in the World of Perception AI”.
The key components of perception system AI include sensing systems like camera, Lidar, Radar, and microphone.
Pradeep says, “Looking at the cost and power parameters, and now with the advent of Deep Learning, which is a subset of ML, and availability of some very interesting hardware options, I think this has opened up the use of Deep Learning. In some cases, completely replacing the traditional signal processing algorithms going way beyond what was done earlier, in terms of the amount of data it can process and also in some cases, there is a combination of traditional signal processing with Deep Learning.”
Perception AI: Use cases
Automotive and Robotics
Sensors guide a truck from source to destination on dedicated lanes in the trucking industry. There are also lower end-use cases like robotics which can be used for services or delivery where the robots use sensors to understand their surroundings to find their way around.
Companies use vibration sensors attached to motors to understand specific signatures. And these are typically done using ESP pattern recognition, but now they are being replaced by ML and Deep Learning which can be used by low power hardware.
Surveillance is done with a combination of Deep Learning and specialises hardware in the multimodal use cases. Now, there are multiple sensors with audio and video combined, trying to get information from the surroundings. The 2D cameras with 3D LiDARS can be used in traffic junctions to monitor vehicles and pedestrian movement. Sometimes, the 2D cameras miss out on many images due to excessive light or rain or environmental conditions obstructing standard cameras. 3D LiDAR can actually detect objects in such conditions and use a combination of these two to get the traffic pattern for a more intelligent traffic management system.
Medical equipment
The medical field is also using Deep Learning and FPGAs specifically for smart surgery, smart surgical equipment etc.
Edge AI
General Purpose Hardware like the CPU, DSP , GPU is strapped to DNN Engine.
Deep learning models require specific hardware to run it efficiently. These are called DNN engines. So they are strapping these to the CPUs and DSPs and the GPUs, basically allowing the CPUs to offload some of the work to these engines that are tightly coupled to the same chip. The general-purpose hardware is now getting variance and is tuned for AI.
FPGAs are programmable devices, and the companies providing FPGAs want to enable AI in their key applications across industries. They want to get high performance with low power where you can write the code, burn it in the FPGA, and design it on the field. The trade off is the lack of software developer friendliness. The developers have to use hardware to implement neural nets. However, companies are building tools and SDKs that make it easier, but still it’s a long way to go.
ASICs are basically application specific integrated circuits specifically designed for AI workloads.
