The algorithm defines the hardware, ignoring putting AI back in the black box.
Fish and sheep are full of color from the concave templeQuantum bit | WeChat official account QbitAI
These days, AI applications in daily life, that’s reallyAs thin as hair.It’s in.
At the gas station, someone secretly smoked a cigarette? Such a dangerous behavior, AI’s small eyes can be targeted at the first time.
In residential areas, throwing objects at high altitude is a big headache, but with AI standing guard, you can directly hit the scene and give an alarm.
AI can cover all the details such as the old man falling down and even the kitchen trash can cover without a cover.
Needless to say, the safety factor of the construction site needs to be full, from smoke detection to seeing who is not wearing a helmet and forgetting to wear reflective clothes, which is not a problem for AI.
Most importantly, I want to introduce such an AI "inspector", nowThere is no need to install a new smart camera.And other equipment.
You may not believe it, but all the tasks in all the above scenarios, even a "box", can be hold.
The magic of an AI "box"
Yes, it is such a 10-inch real box:
This "box" comes from contempt, and all these AI applications mentioned at the beginning are its actual landing cases.
From the hardware point of view, this device named "Rubik’s Cube Intelligent Analysis Box" has an AX630 AI chip built in, and the computing power of INT4 reaches 28.8 tops-
A proper edge computing product.
But if you add a lot of AI algorithms behind it, things will change:
With the algorithm blessing, it can quickly connect with the front-end camera products through RTSP (Real-time Streaming Protocol), national standard and other protocols, and it can complete many different types of detection tasks such as "face detection", "smoke detection" and "dangerous behavior detection", and adapt to various application scenarios such as gas stations, communities, parks and construction sites.
In other words, the original camera does not need to move, and with such a Rubik’s cube box, the intelligent upgrade of various detection systems can be completed.
Is it a little convenient?
Speaking of this, you may also see that, unlike the common IoT smart devices on the market, the focus of this Rubik’s cube box is not on hardware, but onalgorithm.
It can even be said that the final product type of this hardware isDefined by an algorithm——
Put the core solutions on the algorithm level, get rid of the dependence on hardware, make the hardware as universal as possible, and make a set of hardware products can cope with many scenarios.
"Algorithm definition hardware"
Then the question comes, why should we find another way and use "algorithm to define hardware"?
This matter has to start from the status quo of the Internet of Things industry.
Friends who have been paying attention to AI for a long time know that security is one of the earliest application scenarios of AI. Whether it is AI company or Hikvision, an established technology company, aroundAIoTThe concept of (artificial intelligence Internet of Things), the most common case is in the security scene.
However, the Internet of Things scene with intelligent requirements is far more than security.
According to the data in 2021, in the field of AIoT, the penetration rate of AI is only 4%, and 96% of the scenes are not infiltrated by AI.
Behind the data, it reflects one of the biggest pain points that AIoT is facing now:Fragmentation of demand.
Explaining this problem from a technical point of view is that in a business with large demand and relatively simple and concentrated scenes, the algorithm can be reused and iterated continuously.
However, if it is replaced by more fragmented scenes such as battery car detection, garbage sorting, high-altitude parabolic, etc., it is limited by the current AI technology, and there are great difficulties in both data acquisition and algorithm reuse.
And such a fragmented scene is the big head of AIoT.
Traditional solution: massive hardware+customized algorithm
How to solve it?
The traditional solution is simple and rude:Massive hardware+customized algorithm.
In other words, the camera dedicated to detecting the battery car is used to detect the high-altitude parabolic objects, and the hardware dedicated to detecting the high-altitude parabolic objects is used. Under such a product system, tens of thousands of different cameras can even be produced.
Specific to a scene, taking a smart community as an example, if a community has the needs of both battery car detection and intelligent alert and outdoor channel occupancy detection, the whole workflow will look like this:
First of all, it is necessary to assign different points in advance to determine which point to deploy which hardware equipment. Then according to these specific needs, we will place orders and purchase different types of products respectively.
In this way, it puts forward higher requirements for site planning and survey, and once the equipment is deployed, it will be more difficult to adjust the functions of different points.
Speaking of which, I think you can see the problems.
First, in the fragmentation networking scenario, "massive hardware+customized algorithm" lacks flexibility, and the hardware construction and maintenance costs are high. Basically, if you want to deploy new algorithms for new requirements, you have to change the hardware again.
Secondly, from the perspective of technical practice, because the solution depends on the hardware function to a great extent, the algorithm needs to make a compromise more or less, and balance between business requirements and hardware, resulting in the most suitable algorithm is often not the optimal solution.
New idea: software and hardware are integrated to maximize the advantages of the algorithm.
Faced with such a market situation, contempt, as a company that started with AI algorithm technology, has gradually explored a new idea of "algorithm defining hardware".
As mentioned earlier, that is, doing the opposite,Solve the problem of scene differentiation with algorithm as the core.Weaken the dependence on hardware features.
Taking the scene of smart community as an example, we only need to purchase unified hardware after determining the number of detection points in the case of choosing the intelligent analysis box of the Rubik’s Cube.
Then, according to the specific plan, different algorithm packages can be installed at different points. For example, if it is necessary to monitor whether the battery car enters the elevator in the elevator, then load and install the algorithm package for battery car detection.
In the future, if you need to use this point to detect fireworks, you don’t need to change the hardware, just change the algorithm package for fireworks detection.
To sum up briefly, in products such as Rubik’s cube box, what we do is to develop hardware based on the ability of software and hardware, starting from the demand of maximizing the algorithm ability.
By loading different algorithm packages, different products are formed on a hardware device, which makes the hardware itself more universal and standardized.
In this way, from the user’s point of view, on the one hand, the cost of intelligent transformation on the old system will become lower, and more potential needs can be realized at a lower cost.
On the other hand, the product itself starts from the algorithm, maximizes the advantages of the algorithm, and can achieve higher cost performance. For example, through the algorithm to optimize the new products of lower grade, so that it can reach a higher level of computing power and accuracy.
It is worth mentioning that this is only the first stage of "algorithm definition hardware".
It is revealed that with the improvement of the algorithm distribution platform, the hardware will further evolve into the carrier of the algorithm. Just like Tesla’s OTA, algorithm updates can bring new functions to hardware products.
AI company’s breakthrough opportunity
In fact, for a long time, in the security and other Internet of Things scenarios, although the importance of intelligent capabilities has become increasingly apparent, the outside world still has doubts: As a latecomer, what is the competitive advantage of AI company?
In the commercial fields such as supply chain and channels, traditional strong enterprises undoubtedly have the advantage of first-Mover market, which makes them establish their dominance in the field of hardware-led standardized products early.
But now it seems that, for this reason, in the fragmentation scene that more and more mass-produced hardware products are difficult to cover, there is a breakthrough opportunity for AI companies.
The contemptuous "algorithm defines hardware" is a representative breakthrough path of AIoT: the algorithm is essentially directly oriented to various application scenarios, which is naturally closer to the needs of users. With the algorithm as the core, it is possible for hardware to meet the needs of massive application scenarios of AIoT.
Compared with traditional strong enterprises, the core advantages of AI company are still reflected in the long-term research and in-depth insight into AI technology. This advantage is embodied in:
Long-term investment in basic model research has a deeper understanding of algorithm model;The algorithm accuracy in the leading position in the industry;The algorithm mass production ability can efficiently generate massive algorithms and lower the threshold of algorithm production;A more scalable platform can realize efficient and flexible algorithm iteration according to the changes of user scenarios;Have the product capability of software and hardware integration to maximize the advantages of the algorithm.……
From this point of view, in the AIoT era, AI companies that are good at AI algorithms already have the first-Mover advantage.
Reporting/feedback