Qualcomm Looking to Remove Barriers to Visual AI
16-02-2021 | By Robin Mitchell
Recently, Qualcomm announced that it is looking to remove the barriers that prevent visual AI systems from being implemented in everyday solutions. What challenges does AI face, what is Qualcomm looking to do, and how does this mimic biology's neural system?
What challenges does visual AI face?
The development and implementation of AI has been a long journey that started several decades ago. The first AI concepts have dated back to the first computers, but it was the advent of neural networks that saw the acceleration of AI development. Furthermore, data gathering and the increasing processing power of computers allowed AI systems to become more advanced and eventually become practical.
However, AI still suffers from a range of problems that prevent its integration into products or make its unpopular with customers. To start, AI is an energy-intensive task and typically requires a powerful processor. This makes integrating AI into small microcontroller-based systems extremely challenging.
AI also suffers from privacy concerns, especially when considering systems that can record potentially sensitive information such as visual and audio information. Since AI is generally too demanding to run on the device itself, this data can be sent to a remote system for processing, but this requires sensitive information to leave the device and be left exposed on a remote system that could be hacked.
How Qualcomm Wants to Change visual AI Systems
Qualcomm is a leader in the tech field and highly specialises in the development of mobile devices. Since mobile devices typically struggle with AI applications, it makes sense that Qualcomm recently announced that it intends to explore and develop AI systems that will solve the problems it faces.
If Qualcomm can integrate specialised AI hardware into their products while simultaneously improving their capabilities, developers of low-powered devices requiring AI capabilities would most likely choose Qualcomm as the underlying technology. According to Qualcomm, they are specifically looking at reducing the power requirements for visual AI systems by minimising data that needs processing and ignoring redundant data.
One option that can help with this is to reduce the total number of processed frames from a camera. Instead of examining every single frame from a camera, an AI system can instead use alternative frames. Since video generally includes little change over individual frames, many of these frames can be skipped.
The second method that Qualcomm is looking into is only processing change in data as opposed to absolute data. Instead of examining every skipped frame, the difference between the two frames is taken, and an AI system then processes this data. If the difference between two skipped frames falls under a threshold, then no data is sent to an AI processor, reducing energy consumption. Such a system is similar to what researchers are trying to achieve with light sensors that only react to a change in light signal instead of an absolute value which is how the human eye works.
How will such systems be advantageous for AI?
The ability to compress and reduce the amount of information being processed by an AI system allows for that system to be reduced in complexity and in power consumption. The ability to do both of these at the same time not only makes the final design smaller but also more energy-efficient.
Since Qualcomm specialises in mobile technologies, the ability to integrate AI systems that can process visual data without compromising on power consumption and silicon space used will allow for visual AI to run on the edge with ease. Furthermore, the ability to run AI locally on a device improves privacy concerns by not requiring data sent to a remote server. It also helps to save on power consumption by reducing the use of wireless technologies (Wi-Fi is notorious for energy consumption).
Using low-cost, high-efficiency AI on edge devices also allows for the possibility of data obscuration. A security camera system can be directly connected to an edge AI device that extracts important information. This information is then sent to a remote server for further processing. While the exact image is lost, the data in that image is not (i.e. points of interest). Thus, future AI systems can be used to read such information (i.e. the number of people, key facial features), without compromising privacy.
Overall, AI can bring a lot of power to a device, but using remote services to handle private information is by far the least desirable option. If Qualcomm can create an AI device that intelligently reads data to only use what is relevant while keeping energy costs down, AI could be very easily integrated into everyday devices.
Read More
