Don't fall into the 9 misunderstandings of machine learning

Machine learning has proven to be very useful, but it is also easy to assume that it solves all problems and applies to all situations. Like other tools, machine learning is useful in specific areas, especially those that you always know, but never hire enough people to solve; or those that have clear goals but don't have a clear implementation.

Still, every business organization can use machine learning in some way. In a recent Accenture survey, 42% of executives said they expect artificial intelligence to be behind almost all of our innovative projects by 2021. stand by. However, if you remove the hype, you can get better results by understanding what machine learning can and can't do to avoid common pitfalls.

Myth 1: Machine learning is artificial intelligence

Machine learning and artificial intelligence are often used as synonyms. Machine learning is the most successful technology in the real world from research laboratories, while artificial intelligence is a broad field covering areas such as computer vision, robotics and natural language processing. And methods that do not include constraints on machine learning. You can see artificial intelligence to see everything that can make a machine smart. All of this is not the kind of "artificial intelligence" that some people worry about competing with humans or even attacking humans.

You should be cautious about all kinds of popular vocabulary and try to be precise. Machine learning is about learning patterns and predicting big data sets; the results may seem "smart", but the core is the use of statistics at an unprecedented speed and scale.

Myth 2: All data is useful

Data is needed to do machine learning, but not all data is available for machine learning. In order to train the system, you need representative data to cover the patterns and results that the machine learning system needs to process. The data you need should not contain unrelated patterns (for example, the photo shows all the men standing up and all the ladies sitting, or all the vehicles are in the garage, all the bikes are in the muddy ground) because you created The machine learning model will reflect those patterns that are too specific and look for these patterns in the data center you are using. All the data used for training needs to be labeled and marked with features that match your questions to the machine learning system, which requires a lot of work.

Don't assume that you already have clean, clear, representative or easily tagged data.

Myth 3: You always need a lot of data

Recent advances in image recognition, machine reading comprehension, language translation, and other areas are mainly due to the fact that we now have better tools, computing hardware such as GPUs that can process large amounts of data in parallel, and large data sets that have been tagged. , including ImageNet and Stanford Question Answering Datase. However, because there is a technique called transfer learning, you don't always need a lot of data to get good results in a particular area; instead, you can train a machine learning system to learn with a large data set and then transfer it. Go to your own small training data set. That's how Salesforce and Microsoft Azure's custom visual API works: you only need 30-50 images to display what you want to categorize to get good results.

Transfer learning allows you to customize a pre-trained system for your problem with relatively little data.

Myth 4: Anyone can build a machine learning system

There are many open source tools and frameworks for machine learning, and countless courses teach you how to use machine learning. But machine learning is still a proprietary technology; you need to know how to prepare and partition, train, and test your data. You need to know how to choose the best algorithm and what heuristics to use, and how to turn it into reliable. production system. You also need to monitor the system to ensure that results remain relevant over time; whether your market changes or your machine learning system is good enough, you end up with a different customer base and you need to constantly check if the model is with you. The problem is consistent.

Proper use of machine learning requires experience; if you are just getting started, you can use APIs to pre-train models that can be called from code, while hiring data science experts and machine learning experts to build custom systems.

Myth 5: All patterns in the data are useful

The survival rate of pneumonia in asthma patients, chest pain or heart disease patients, and any elderly over 100 years old is much higher than expected. Yes, in fact, a simple machine learning system designed to automatically send hospitalization notifications may notify them to go home (a rule-based system that trains with the same data, just like a neural network). The reason why the survival rate is so high is because pneumonia is very dangerous and patients will be taken to the hospital immediately.

The system sees an effective model from the data; this is not a useful model for selecting who needs hospitalization (but it can help insurance companies predict treatment costs). Even more dangerous is that you don't know that this dataset has this useless anti-dataset unless you already know it exists.

In other cases, a system can learn an effective model (such as a controversial facial recognition system that can accurately predict sexual orientation from a selfie) because it has no clear and obvious explanation, so it is useless (in In this case, the photo will show some social cues, such as a photo gesture, rather than showing some of the natural features.

The "black box" model is effective, but it is not clear what patterns they have learned. Things like the generic add-on model are more transparent, and the understandable algorithm allows us to better understand the learning content of the model and thus determine whether it is suitable for deployment.

Myth 6: You can use reinforcement learning at any time.

Today, almost all machine learning systems in use use supervised learning; in most cases, the system is trained based on clearly labeled data sets, and humans are involved in the preparation of these data sets. Organizing these data sets requires time and effort, so people have a great interest in unsupervised forms of learning, especially reinforcement learning (RL). Reinforced learning means that learners pass trial and error, interact with the environment, and predict correctly. Behavior is rewarded. DeepMind's AlphaGo system combines enhanced learning and supervised learning to defeat advanced Go players, while the Libratus system built by Carnegie Mellon University team enhances learning and two other artificial intelligence technologies in No Limit Hold'em. Defeated the world's top players. Researchers are experimenting with enhanced learning for everything from robots to testing security software.

But outside the lab, reinforcement learning is not common. Google DeepMind reduces data center power consumption by learning how to cool more efficiently; Microsoft has adopted a specific and limited enhanced learning version called Co ntextual Bandit, MSN. Visitors to com display personalized news headlines. The problem is that few real-world environments have rewards and immediate feedback that can be easily discovered, especially when the agent takes a number of actions before anything happens.

Myth 7: There is no prejudice in machine learning

Because machine learning learns from data, it replicates any bias in the data set. Searching for images of CEOs may result in photos of white male CEOs as more CEOs are white males. But it turns out that machine learning also magnifies bias.

The COCO dataset, often used to train image recognition systems, has male and female photos; but more of the female images appear next to kitchen appliances, and more male images are associated with computer keyboards, mice or tennis rackets and snowboards. appeared. Train the system on COCO, which binds men and computer hardware more closely.

A machine learning system may also bias another machine learning system. Use a popular framework to train a machine learning system, using words to express the vector of relationships between them, learning like "men are relative to women as computer programmers versus housewives", or "doctors are like nurses The stereotype of the boss relative to the receptionist. If you use this system, the system will translate he and she (English) into a language with gender neutral pronouns (such as Finnish or Turkish), "they are doctors" and become "he is a doctor", "they are nurses "Become "she is a nurse."

It would be useful to get similar advice on the shopping site, but it can cause problems when it comes to sensitive forests, and it will generate feedback loops; if you join Facebook against vaccinated organizations, Facebook's recommendation engine will recommend other Focus on conspiracy theories or think that the earth is a flat organization.

It is important to understand the prejudice issues in machine learning. If you can't eliminate the bias in the training dataset, you can use techniques such as regularizing gender associations between pairs of words to reduce bias or add irrelevant entries to the suggestion to avoid "filtering bubbles."

Myth 8: Using machine learning is a good side

Machine learning provides powerful capabilities for anti-virus tools, further focusing on the behavior of new attacks to discover these behaviors as quickly as possible. But in the same way, hackers are also using machine learning to study the defenses of anti-virus tools, and launch targeted phishing attacks on a large scale by analyzing large amounts of public data or previously successful phishing incidents.

Myth 9: Machine learning will replace humanity

People are often worried that artificial intelligence will grab human work and will definitely change the way we work; machine learning systems can increase efficiency and compliance and reduce costs. In the long run, artificial intelligence will create new roles in the business and eliminate some of the current positions. But many machine learning tasks that automate were previously unimaginable, whether it's complexity or scale. For example, you can't hire enough people to view every image posted to social media to see if they contain you. The brand identity of the company.

Machine learning has now begun to create new business opportunities, such as improving customer experience through predictive maintenance and providing advice and support to business decision makers. Like previous generations of automation, machine learning gives employees the freedom to use their expertise and creativity.