Shulou(Shulou.com)06/02 Report--

Independent thinking, seeking truth from facts, perseverance, diligence to make up for clumsiness.

-- A western blue flower

First, what is machine learning from AI technology?

Tom Michell defines that for a certain type of task T and performance measurement P, if the performance of a computer program measured by P on T improves itself with experience E, then we say that the computer program learns from experience E.

In general, in order to well define a learning problem, we must identify three characteristics: the type of task, the criteria for measuring task improvement, and the source of experience.

Take a handwriting recognition learning problem:

Task T: recognition and classification of handwritten text in images performance standard P: classification accuracy training experience E: handwritten text database machine learning and deep learning with known classification

Deep learning is a specific type of machine learning with strong ability and flexibility. It represents the world as a nested hierarchical conceptual system.

Deep learning is an algorithm inspired by the brain structure and function of artificial neural networks. due to its rise, it promotes the landing application of artificial intelligence technology, such as face recognition.

Second, from technology to user experience

With the development of technology, a large number of products based on AI technology have been created in the market.

Product design is a very difficult thing in itself.

AI products, because of the complexity of their application scenarios, make product design more difficult. For product managers, they need to think more thoroughly about the actual application scenarios when designing products.

The article "elements of user experience" begins with the following paragraph:

In product design, we pay too little attention to user experience: the products we produce are for people to use in the real world. In the process of product development, people pay more attention to what the product will do. User experience is another factor that is often overlooked-- how the product works-- and this factor is the key factor that determines the success or failure of the product.

User experience does not refer to how a product itself works. User experience refers to "how the product connects and works with the outside world", that is, how people "touch" and "use" it. When people ask you about a product or service, they ask about the experience of using it. Is it difficult to use? Is it easy to learn? How does it feel to use it?

In the above excerpt, there is a very interesting passage to be taken out separately:

In the process of product development, people pay more attention to what the product will do. User experience is another factor that is often overlooked-- how the product works-- and this factor is the key factor that determines the success or failure of the product.

Let me explain the meaning of the above sentence. There are two layers of understanding:

First, in the process of product development, the product manager pays more attention to what the product will do in the future. That is, ideally, the future function of the product. The second layer points out a point that the product manager ignores, when the product is applied to the real scene, the product interacts with the real scene, and how the product will work in such scenarios.

In product development, product managers tend to ignore the interaction between the product and the real scene. This is especially true in the design of AI products. The usage scenarios of AI products are complex, and it is often necessary for products to make judgments under different circumstances. If the product manager does not think clearly about all possible scenarios and make the design plan in advance, the user experience caused by the AI product may be very bad.

Use an example of an AI product to illustrate.

Second, AI scene, pedestrians running through a red light to capture.

Restore the real business scene of "pedestrians running through red lights":

Crossroads, in order to better regulate pedestrians and vehicles, so there are traffic lights. When the red light is on, neither pedestrians nor vehicles can pass or drive. When the green light is on, both pedestrians and vehicles can pass.

But there will always be people who disobey traffic rules and run red lights. Therefore, in this scenario, we can use AI to help us better monitor the behavior of pedestrians running red lights. We hope that if a pedestrian A runs a red light, our device can record that person and upload information about the user's running a red light, including time, place and pedestrian's face, to the back-end monitoring device.

The requirement description above is relatively broad, which makes us think about how to disassemble the requirement into AI + hardware implementations.

During the red light, the algorithm is integrated with the industrial camera to detect and capture people who appear at the zebra crossing during the red light; upload the event-related information, including time, place, pedestrian's face, to the back-end monitoring device. Third, from requirements documents to actual business scenarios

After the output of the requirements document, our product seems to be on the verge of success. But let's look at a picture:

(1) what is the above picture?

As a promotional advertisement, Ms. Dong Mingzhu, who was printed on the body of the bus, was caught running a red light. And boarded the local pedestrian red light exposure platform.

(2) Why did the above happen?

When outputting the requirements document, the product manager only describes that the requirement is to detect pedestrians running red lights. The situation of car body advertising is not taken into account, so the solution is not considered in advance, so the situation in the picture above appears.

In his new book, Yu Jun's Product Methodology, Mr. Yu Jun said:

Technology itself does not create value, and it must be applied to the product.

No matter how high-end the technology is, without the product, all

Fourth, how to do better? 4.1 restore the real scene

Let's think about the real business scenario. Traffic lights are used to maintain traffic order at crossroads.

When the red light in aisle An is on, pedestrians, electric cars and bicycles in aisle An are prohibited. At this time, vehicles in channel B are passable. If there is a portrait printed on the body of the vehicle in Channel B, what will happen?

Perhaps the idea of designing the product at the beginning did not go deep enough to ignore this scene, resulting in Dong Mingzhu printed on the car body as a pedestrian running a red light.

4.1 how to do better

We understand the real scene, based on the understanding of the real scene, if we do it all over again, how can we do it better?

Live detection: the scene requirement is to capture the pedestrian running a red light at the zebra crossing, so the area that the algorithm needs to detect is the "zebra crossing" area. The target to be detected is "pedestrian", and the time period is during the red light period. The detection area is "zebra crossing" and the detection target is "pedestrians running red lights". In this area, during the red light period, the car body face appears, so it is detected by the system in order to run the red light. In view of this scenario, we can think of the solution of "living detection" in advance. Boundary analysis: we need to set the detection boundary according to the three elements of time, place and character, corresponding to this scene, namely: red light time period, zebra crossing area, face + living body. Technical solution: in this scenario, how to do in vivo detection? Because of the openness of the scene, unlike online payment, we can ask users to look left and right. Then we can add a target detection, if a face is detected, and a car is detected, then we can rule it out. At the same time, pedestrians running red lights to cross the road is a continuous action, this process will last for a period of time, we can use this as a starting point. One is to consider increasing the "attitude estimation", and the other is to consider that after the "trajectory" detects the first face, it does not push, but continues to detect, and at the same time, it takes into account that the moving direction of the bus is vertical to that of pedestrians at the zebra crossing. To sum up, AI product managers / project managers need to think a lot about everything from algorithmic models to deployment to real business scenarios. In many cases, it is important to think about how the product interacts with real business scenarios, not just how our desired algorithms work. How the product interacts with the world is very important.

Reference:

"user experience elements" https://github.com/lonelygo/Shift-AI-models-to-real-world-products

Http://www.woshipm.com/ai/3187438.html

Welcome to subscribe "Shulou Technology Information " to get latest news, interesting things and hot topics in the IT industry, and controls the hottest and latest Internet news, technology news and IT industry trends.