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

2020-01-02 17:00:00

The full text has a total of 7008 words and is expected to last 21 minutes.

Source: why did the Pexels artificial intelligence project fail?

With the progress of the times and the development of science and technology, the advanced technology of artificial intelligence has penetrated into human resources, supply chain, multi-level marketing and other fields. The overall development prospects and situation seem to be very good, bright.

However, people's attitudes towards carrying out their own artificial intelligence projects are usually mixed.

At first, when you hear the word "artificial intelligence", you will think it's great and amazing. Indeed, the "success stories" of artificial intelligence spread throughout the year, and examples of the application of artificial intelligence to increase sales and turnover abound. As a result, people may think that there must be many opportunities for artificial intelligence projects to succeed. On the other hand, people never think about what to do if the project fails. How to defuse the risk and avoid wasting time and money on a project that is not feasible at all? There are still many such problems, but we are not at a loss in the face of these problems, and there are still solutions.

At present, why artificial intelligence projects have been repeatedly frustrated?

This article will discuss this problem and explore the reasons behind its failure, such as lack of data and other factors.

The future of artificial intelligence is bright, isn't it?

Source: Pexels

However, according to a recent study released by Pactera Technology Co., Ltd., about 85% of artificial intelligence projects have failed in recent years.

At this point, people may say, "Don't worry! I will not fail. I will be 15% of the success." Responsibly, it may succeed or it may fail. After all, everything is still unknown. What people need to do now is to expect the best results and think strategically at the same time. All in all, the most important thing is to familiarize yourself with the relevant materials in advance, prepare well, and treat each step carefully.

There are thousands of reasons for failure.

There are thousands of reasons for the failure of artificial intelligence projects, if not more than one.

Artificial intelligence brings infinite possibilities to human beings, including the possibility of failure, and there are many. The failure of an artificial intelligence project may be due to a mistake in the data strategy, a problem in the docking of business and technology, or some human factors. Of course, in addition to the above list, there may be a lot of problems, the author is not to frighten you. Now, taking advantage of the hot atmosphere of the New year, the author is here to tell you a "ghost story"-the death of artificial intelligence, to remind you to be more careful when dealing with artificial intelligence in the future.

Remember, plan ahead in order to prevent problems before they happen.

1. "big data" is not big enough.

In recent years, the popularity of the word "big data" has increased unabated, however, the public also has all kinds of doubts about it. How big is the so-called "big"? How much "data" is needed? Indeed, "data" is indeed the crux of the problem, which is reflected not only in the lack of data, but also in many aspects, such as data marking and training data.

The success or failure of an artificial intelligence system mainly depends on the quality of input data. Therefore, if there is not enough data to support it, how can we get substantive results? But specifically, what's wrong with the data itself?

First of all, lack of data is a big problem. If you are running a small project and the relevant data is limited, you need to discuss in advance with an experienced artificial intelligence consultant or data scientist to understand your expectations and current status of the dataset. How much data is needed?

To tell you the truth, this question is difficult to answer because it depends on the specific situation. The amount of data required depends largely on use cases, data types, and expected results. However, sometimes people often hear people say, "of course, the more, the better." Anyway, as far as data science projects are concerned, that's true.

Source: Pexels

two。 Selection of data

Although it is true that a lot of data are collected sometimes, are these data appropriate? People may think that now that all the necessary data is available, the project is bound to be successful!

Wait a minute, sometimes the data may seem like a lot, but it may not be appropriate. If you are an e-commerce business, there may be a lot of information about your customers, such as their names, addresses, invoices, and even their bank cards. So you know what they bought, when they bought it, what they browsed, and when and how they contacted you.

But which of these data is necessary? To put it simply, solving different problems requires different information. For example, when you want to implement a recommendation system, you don't need to use all the demographic data. Instead, you have to collect customer purchase records. However, if it is to be used to predict customer churn, various other factors need to be taken into account.

So even if all the data in the world is in your pocket (which is, in fact, impossible), consider what data is necessary. It is true that many people are very happy to collect all kinds of data crazily, even the more the better, but it is not necessary at all. In a word, only choose what is right, not more, because there is no point in choosing more.

3. Data mark

Tagging humans-- of course; labeling data-- never.

When completing an artificial intelligence project, we not only need the existence of the data, but also need to mark the data in order to make it meaningful. If the data collected is disorganized, humans need to spend another amount of time to complete the tedious task of data tagging. The task of data tagging is so boring and tedious that many companies don't pay attention to this important task at all. JenniferPrendki, a data scientist, published an article on Amazon AWS's official blog that read:

"although there is such a huge elephant standing in the room, even the most powerful technology companies do not seem to see it, or are selectively blind. This elephant is the data marker."

For many machine learning models that train by supervised learning, data tagging is particularly important. The model requires that the data must be marked, otherwise the data will not make any sense.

Because data tagging is extremely time-consuming and laborious, data scientists usually choose to use off-the-shelf data that has been tagged. For example, nowadays, when people implement machine vision projects, although they can obtain a wide range of high-quality images from various sources, they usually mainly choose ImageNet database. Because ImageNet database is the largest tagged image database at present, there are about 14 million images in existence.

Today, human beings continue to generate more and more data every day. The amount of data uploaded to Facebook every day is as high as 50 megabytes, and Facebook is not the only source of data. It is conceivable that, taking into account all this data, we humans have reached an awkward position, that is, there are not so many hands to mark the data.

4. Unable to fully imitate human beings.

Source: Pexels

Usually, people always expect artificial intelligence to perform an intelligent task at a level comparable to, or even better than, human beings. It is also reasonable to think so, because we all know that artificial intelligence performs better than humans in more and more tasks. Indeed, not long ago, artificial intelligence even beat the go champion. However, in terms of flexibility, artificial intelligence systems are still far behind human thinking.

To further illustrate this point, "Smart recommendation" is an excellent example. Suppose you meet a very interesting person at a startup event (suppose his name is "John"). John enjoys talking to you and admires your profound business and technical knowledge. As he also wants to know about it, he wants you to recommend a related book to him. Then, you may quickly search the relevant books in your mind, such as A, B, C, D, E, and so on. So you replied, "John! I know which one you should read! you can read XX." So the question is, how do you know which book to recommend to John?

In fact, first of all, your brain scans the relevant information that has been stored so far, such as John's knowledge, his interest in talking to you, and his personal style. At this time, even if you don't know his true preference for books, you can recommend the most suitable bibliography based on the above information, because you always feel that he will like the book. Indeed, human feelings are often accurate.

Now let's change the scene. John "encountered" an artificial intelligence system this time. John opened an online bookstore website, and a dazzling array of bestsellers immediately appeared to him. But John never saw what he was interested in, so he kept clicking on "next page". What causes it?

Because the artificial intelligence system does not store background information about John. Professionally, this is a typical "ColdStart" case, in which personalized recommendations cannot be generated because John's information is not stored in the system. However, when John clicks on the search box and enters "Entrepreneurship" to search, a series of books related to "Entrepreneurship" will pop up. So John continued to browse through these search results. At this point, the artificial intelligence system will understand that "entrepreneurship" is a topic that John is interested in, and will then be able to recommend relevant content based on this topic.

Although the artificial intelligence system does not fully understand John, it can also analyze their personal preferences according to other users who have also browsed or bought entrepreneurial books. But what if no one else has ever looked for entrepreneurial books at all? In this case, John will not be able to get the relevant recommendation because the system has not obtained any relevant data to learn.

Finally, the books you and artificial intelligence recommend to John may be different. However, your recommendations may all be right, or they may all be wrong, or one by one. However, the human brain never complains about "insufficient data", and all judgments are made on an ad hoc basis. By contrast, artificial intelligence cannot do this. Therefore, as the "master" of artificial intelligence, we human beings do not have to be alarmist, because artificial intelligence can never perfectly reproduce the complex human brain.

5. What is artificial intelligence bias?

Artificial intelligence bias, or algorithmic bias, refers to systematic, repeatable errors in computers that can lead to unfair results, such as gender discrimination, racial discrimination, or other discriminatory colors. Although from the name point of view, artificial intelligence discrimination seems to imply the fault of artificial intelligence, but in the final analysis, it is our own human beings who are wrong.

Google's chief decision scientist CassieKozyrkov once wrote:

"No technology can exist completely without its creator. although human beings express the best visions in science fiction, truly independent machine learning or artificial intelligence systems do not exist. because we humans are its creators, and all technologies more or less reflect the purpose and will of the creator."

No matter where it is used, artificial intelligence bias will usually have a negative impact. For example, for computer vision, recruitment tools, and so on, artificial intelligence bias can make them unfair and moral, or even break the law. What is even more unfortunate, however, is that it is not the fault of artificial intelligence, but that of human beings. Because it is human beings who are prejudiced, it is human beings who spread stereotypes, and it is human beings who are afraid of dissidents.

Therefore, in order to develop a more fair and responsible artificial intelligence system, human beings must break the shackles of personal opinions and beliefs, so as to ensure that the data in the training database is more rich, diverse and fair. It sounds simple, but in fact it's not simple at all. But to achieve this, human efforts are definitely worth it.

Source: Pexels

6. Algorithm VS. Justice

喜悦 Boulamwini (hereinafter referred to as Joey) is a researcher at the Massachusetts Institute of Technology and led the creation of the Alliance for algorithmic Justice (AlgorithmicJustice League). In 2017, Joey gave a speech on "algorithmic bias" on TED, which began with the following software experiment, which is as follows:

"Hi! camera! I have a face. Can you see my face? without glasses? now that you have seen it, what does my face look like? I'll wear another mask. Can you see my mask?"

In the end, the camera failed to detect Joey's face, only Joey's colleague and the white mask she was wearing, not her face. In fact, similar results have occurred more than once. When Joey was an undergraduate at GeorgiaTech, she was working on social robots and needed to complete the task of "teaching robots to play hide-and-seek (Peek-a-boo)." In the end, the robot failed to recognize her because she "borrowed" a roommate's face to muddle through. Later, a similar plot was performed again. A start-up company launched a social robot at a start-up competition in Hong Kong. The robot used the same facial recognition software and also failed to identify Joey.

Why did this happen? To that end, Joey went on to explain:

"computer vision uses machine learning technology for facial recognition. So how does it work? First, you need to create a training data set about face instances. This is a face, this is also a face, and this is not. Gradually, computers will learn how to recognize other faces. However, if the face data covered by the training data set is not rich enough, it will be difficult for the computer to recognize any face that deviates too much from the established standard. It was for this reason that the robot could not see me. "

Even so, what's the problem? People might ask.

You know, if the impact of algorithmic bias becomes wider and wider, it will no longer be as simple as facial recognition. It is true that the following example is too extreme, but its dangers cannot be ignored. If the police use such software to find suspects, facial recognition bias may put a small number of people at a disadvantage and even injustice them. If the machine makes a direct mistake in the process of identification, the consequences are even more unthinkable.

Now that we have talked about the fairness of machines, it is necessary to mention COMPAS here again. In fact, in a previous article on "trusting AI", the author has already described COMPAS. COMPAS is actually a prediction algorithm, which is used in the United States to predict the probability of a criminal committing a crime again and sentencing accordingly.

You know, such an algorithm, which relies entirely on historical data, will directly determine that black criminals have a higher recidivism rate.

In addition, Amazon has also launched a "notorious"AI recruiter". As a result, the system shows a preference for men, and since most office workers are men, it is entirely logical to have such a choice.

7. The department executives don't pay attention to it.

At present, the application of artificial intelligence is facing a variety of challenges, one of which is the lack of attention of department executives. They don't value the value of these emerging technologies, so they don't want to invest, or the department where you want to "Augment" with artificial intelligence is not interested at all.

Indeed, this is also human nature. Today, artificial intelligence is still regarded as a high-risk thing, which is not only expensive, but also difficult to operate and maintain. Nevertheless, the popularity of artificial intelligence continues unabated. In fact, people should use the right method when applying artificial intelligence, propose a business problem that can be solved by artificial intelligence in the initial stage, design a good data strategy, and record appropriate indicators and return on investment.

At the same time, team members should be prepared to "work" with artificial intelligence systems and establish criteria for success and failure in a timely manner.

As you may have noticed, the author used the word "Augment" when talking about the task of artificial intelligence above. The reason is simple: the main task of artificial intelligence is to "assist" human work and support data-driven decision-making, rather than completely replacing human working roles. Of course, there are some artificial intelligence projects that do aim to automate as much as possible. But generally speaking, this is not the "main business" of artificial intelligence, because artificial intelligence mainly cooperates with human beings.

And research shows that the cooperation between human and artificial intelligence can produce better results. In an article in the Harvard Business Review (Harvard Business Review), writers James H.James Wilson and Paul R. Daugherty wrote:

"in a study of 1500 companies, we found that companies produce the highest benefits when humans work with machines."

However, as a leader, his role in artificial intelligence projects is to help employees understand why artificial intelligence technology is introduced and teach them how to use models to accomplish tasks. If not, even the most magical artificial intelligence system will only be reduced to a meaningless combination of numbers.

To further illustrate its importance, let's take a look at an example cited by the CIO CIO magazine. In order to improve customer service, a company called Mr.Cooper has introduced a recommendation system that provides answers to customer questions. However, after nine months of running the system, the company found that its employees were not using the intelligent system. After six months of research, the company finally found out what the problem was. Finally, the study found that because the training data are mainly some internal documents, and the description of the problem in these documents is full of professional terms, and ordinary users often use everyday language when describing the problem, so this makes the algorithm model incomprehensible, and finally recommends some irrelevant content.

The above examples fully demonstrate the importance of employees' understanding. They must understand why and how to work with artificial intelligence, and have the right to question the effectiveness of the system and report relevant issues if necessary. In addition, this example also tells us how important reliable training data is!

Source: Pexels

8. "die young"

In the actual implementation of artificial intelligence projects, some people may have finished before they even started.

It is no exaggeration to say that this could really happen. The reason for this is that people are eager to start the project before they have the necessary resources such as data, budget, team, strategy, and so on. If these elements are not prepared in advance, everything will turn into an unrealistic fantasy.

It is precisely because of this that we have repeatedly stressed the importance of strategic approaches. Before embarking on an artificial intelligence project, you must make sure that you have prepared the elements, found the right business use cases, conceived the right data strategy, and set goals. If you do not think about the specific strategy at the beginning, the subsequent steps will be difficult and the risk will be greatly increased.

When creating an artificial intelligence project, especially your first project, you should set a large overall goal to guide the direction, as well as some phased goals.

In this way, while proving the feasibility of the project, it can also effectively reduce the risk of failure, thus avoiding wasting the company's money on a completely meaningless tool. When implementing the first artificial intelligence project, you should not immediately roll it out across the company. Instead, you can choose to experiment with the PoC project first, so that the whole organizational structure can adapt to this "new normal" of the future ahead of time.

Over time, the whole company and artificial intelligence systems will develop: artificial intelligence systems will become more and more advanced, while company teams will become more efficient and data-driven will increase.

In the process of the project, if people can gradually achieve the phased goals, and always grasp the overall goal and general direction, then mutual benefit and win-win will be the inevitable result. In a word, artificial intelligence is only a tool that human beings use to achieve their goals, not the goals themselves.

How to avoid failure

Of course, failure is not inevitable.

Now that so many organizations have failed in artificial intelligence, we can learn from their mistakes and prevent our companies from making the same mistake again.

In addition, we should also follow the laws of the market, avoid being limited to the immediate competition, and focus on the whole technological world. Only in this way can we set realistic goals, find promising use cases, and find our own limitations in time.

Source: Pexels

Human vision, guidance and investment eventually become an important part of the success of artificial intelligence projects. Now that we have firmly established the road of artificial intelligence, please stick to it until the end. I believe that one day we can achieve the perfect overall situation of "winning again and again".

Https://www.toutiao.com/i6777018959519547908/

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.