How to overcome challenges in Edge AI implementation

What do Face Recognition, Sound Recognition, and Gesture Detection have in common? They are among the core functions needed by intelligent devices to autonomously perceive and interact with the world around them. That is the promise of Edge AI in which Intelligent devices that can act and think for themselves.

The Future of AI in Everyday Devices

AI has become possible on small low-power devices. And that means an avalanche of innovative products we can’t even imagine right now, headed our way in the coming years.

If you thought inventions like folding screens and flying cars at CES (Consumer Electronics Show) 2024 were exciting, you haven’t seen anything yet. As new low-power AI-native chips become available, and core functions get built out, we are in for an incredible array of smart device innovations in every industry from healthcare to farming.

However, building AI applications on small devices is not like in the cloud, nor flying cars for that matter, where we have the luxury of unlimited compute power. So how is this done?

Challenges in Edge AI Implementation

For small devices equipped with the latest MCUs with integrated Neural Processors, performance and resources are finite, and core functions like face recognition, do not yet exist. Without these technology primitives for customers to build applications upon, Silicon Vendors can’t sell chips.

That’s why Semiconductor companies are scrambling to collaborate with embedded systems firms such as embedUR to develop these essential functions—or technology primitives – for their chips.

They will only succeed if they lower the barrier to adoption, and to do that, eventually they need to give customers SDKs containing the core functions they need to create smart devices. This way anyone can use AI Vision, without having to know how to process an image.

Automotive companies investing in flying cars, may have the teams to build what they need from scratch, but most companies can only begin when the core functions they need, exist.

Building Edge AI Applications

To illustrate: Once you can identify an object, recognize it as say a “chicken” and follow it as it moves around in front of a camera…what else can you do? You can teach AI to distinguish movement patterns because healthy chickens move differently than sick ones do.

Here are some key differences:

1. Posture and Mobility: Healthy chickens tend to stand upright, explore their environment, forage, and engage in social behaviours with other chickens. Whereas unhealthy chickens might appear lethargic, sit more often, or huddle in a corner.

2. Head Movements: Healthy chickens bob their heads frequently as they walk. This head bobbing is a normal part of their motion. A lack of head bobbing could suggest illness.

3. Feeding Behavior: Healthy chickens are more vigorous eaters. They will peck at food with enthusiasm and are more competitive when feeding in a group.

4. Social Interaction: Healthy chickens actively participate in communal activities like dust bathing and sunning, whereas sick chickens show little interest in the flock.

Point is: Now you have a health screening application, which could run on a tiny camera on each bird cage, leveraging object recognition without having to build it. That removes so much R&D time and risk from the equation.
Beyond the obvious autonomous vehicle market, there are hundreds of completely different applications that could be built on-top of object recognition in devices of all shapes and sizes.

Edge AI Technology Partnering

Object Identification is just one of several sub-disciplines of one core function, AI Vision. To unleash all the incredible applications people are dreaming about, there are still dozens of core functions that innovators need, before they can confidently choose an Edge AI compute platform to build solutions around.

Due to the absence of SDKs with these core functions already built, most companies must wait. However, those unwilling to delay, can mitigate risks by partnering with embedded systems firms that collaborate closely with Silicon Vendors. These partnerships help acquire the necessary core functions, thus accelerating product development and reducing time to market. 

Rajesh C Subramaniam

---

Rajesh C Subramaniam is Founder & CEO of embedUR Systems.