AliMei’s Guide: Machine learning is one of the cores of artificial intelligence, involving complex disciplines such as probability theory and statistics. It is indeed not easy for non-professionals to understand it.
Recently, Bai Ning from Ant AI Platform found the key to explaining “machine learning” while taking a walk, explaining it in an engaging and insightful way. Let’s learn together!
Article Author: Bai Ning (Zhu Baining), Senior Product Expert, Ant Financial AI Platform
One afternoon at the end of May, a few PD colleagues from the Ant AI Platform had just finished lunch and were taking a walk downstairs at Z Space.
During this time, one colleague mentioned a piece of news about a “child who did something silly due to trivial matters”.
As everyone was lamenting, a product expert vividly said, “This child’s model overfitting, poor generalization ability“.
This metaphor is indeed piercing and profound! Everyone expressed that with such insight, that colleague’s “promotion to researcher and taking charge of the organization department” is just around the corner!
Indeed, a pampered (overfitting) fragile mentality (model) often cannot correctly handle (inaccurate predictions, poor generalization ability) when facing various disappointments in the real world, easily leading to tragedy.
In machine learning, when training a model, if the samples are too homogeneous or the feature selection is inappropriate, overfitting will occur. This means treating special cases in the training samples as general cases. Thus, when facing new samples, it cannot be handled correctly.
We often say that test-oriented education and quality education have a huge difference in generalization ability.
Test-oriented education makes students do tests and practice question banks all day long, training models that are overfitted and have poor generalization ability, leading to the phenomenon of “high scores but low abilities”. In contrast, quality education focuses on ability training and is more diversified, resulting in models with strong generalization ability, so students will not face significant problems in life and work in the future.
In fact, if you carefully compare human growth with machine learning, you will find that both have more similarities.
A baby, when born, naturally possesses some abilities (self-contained algorithm library), such as heartbeat, breathing, crying, laughing, and fear.
These abilities are gifts from miraculous evolution and great genes. Otherwise, it would be too difficult for us to relearn these skills.
Of course, to live normally, merely having these initial algorithms and models is not enough; you also need to continuously acquire new skills (self-developed algorithms).
For example, eating, walking, speaking, etc.
Mastering these skills requires parents to tirelessly teach us day after day, year after year, training us to babble and take our first steps.
This is akin to machine learning, which requires a sufficient number of samples (thousands) and consumes a lot of computational resources, enduring minutes, hours, or even days of continuous training to obtain a model.
In fact, in the human brain, there are not only skills (models) acquired through years of training but also some strong rules.
Some of these rules are taught by parents, while others are formed by social moral norms and laws.
For instance, parents might tell children: “Don’t eat food offered by strangers” or “Look both ways before crossing the street”.
In life, we should respect the elderly and love the young, maintain courtesy in interactions, abide by the law, and practice the core socialist values.
For these rules, we remember them and just follow them in the future.
Thus, we can see that our brain is like a decision center or decision hub, containing countless rules and models.
Each decision is made by combinatorially using these rules and models.
In actual machine learning application scenarios, in various business lines, there are also various similar “decision centers”, such as UCT in big security, AGDS in micro-loan, and DecisionX as a “universal decision center”.
This decision center contains hundreds or thousands of complex rules (or strategies) and models obtained through training. For any specific case, a combination of strategies and/or models is used for judgment and decision-making.
Moreover, it is often the case that strategies are judged first; if a certain strategy is not satisfied, a decision is made directly.
As we age, in addition to basic skills like eating and drinking, our decision center also needs to master more skills, such as language, mathematics, music, dance, sports, etc.
In this regard, we often see some children who are “gifted”, meaning they come with super awesome algorithms inherited from their parents.
However, ordinary people should not be discouraged; we can put in more effort (large samples), continuously train hard (models continuously retrain and evolve), and still achieve good results.
Life is a marathon, and model training is similar.
In contrast, in this marathon, the learning methods we use at different ages are also different.
When we are young, most of us use supervised machine learning.
For example, parents would teach us various fruits: “This is an apple, this is an orange”.
Cartoons and storybooks often tell us who the good guys are and who the bad guys are.
Thus, we most commonly use binary classification algorithms: big/small, long/short, yes/no, good/bad.
As we grow older, we encounter more unsupervised or semi-supervised learning, where many things do not come with clear labels of right or wrong.
Therefore, we will use some clustering algorithms. After getting familiar with people over time, we categorize them into groups like “loyal”, “good drinkers”, “good at fixing computers”, “love to eat old mother-in-law’s”.
Although the methods of learning vary, the principles are simple and universal.
We know that a person who achieves success in one area can easily excel in other fields (transfer learning).
For example, Einstein was not only an outstanding scientist but also proficient in playing the violin. A C language expert can quickly become a Java master.
Of course, upon closer examination, there are many differences between the human brain and machine learning.
For instance, a child can recognize many different styles of cars just by playing with a few toy cars or looking at several pictures of cars.
In contrast, to enable a machine to achieve such seemingly simple abilities, it needs tens of thousands or millions of samples for training.
In recent years, one research direction in AutoML has been to solve how to train models with a small number of samples.
Moreover, we find that what humans can do instinctively and in an instant is often very difficult for artificial intelligence.
For example, recognizing objects and faces (image recognition ability), understanding emotions like happiness and sadness, and the ability to walk and run. The reason lies in the great evolution—our models have been developed over billions of years, going through countless iterations.
Conversely, what takes humans a long time to accomplish is a piece of cake for computers.
For example, summing up one million numbers in a short time or calculating pi to a million decimal places.
However, to find the true reasons behind the differences in abilities between the human brain and artificial intelligence, humanity still has a long way to go.
Despite the rapid advancement of technology and the invention of deep neural networks to achieve artificial intelligence (deep learning), enabling machines to recognize cats and play Go, we still know very little about how our own brains work.
It can be said that humans, with their magical brains whose principles are still unclear (non-explainable), have created various models whose mechanisms are also unclear (non-explainable).
In other words, even if you train a model that can recognize a cat using deep learning, that model lacks explainability. You cannot specify which features or principles this impressive model relies on to recognize a cat.
The human mind is composed of countless such obscure models and clear rules.
A person’s life is a process of continuously iterating old algorithms, retraining old models, developing new algorithms, and training new models.
As the saying goes, “reading thousands of books, traveling thousands of miles, conversing with thousands of people”—this helps us master more algorithms, have a more comprehensive sample, and thus train more diverse models.
However, unfortunately, unlike features or skills like “appearance, crying, laughing”, most of our models cannot be inherited by our children.
For example, a senior technical expert (P8) who is “proficient in Java/Python, good at debugging and core analysis” and a senior product manager (P6) who “deeply understands user experience and human nature” cannot guarantee that their offspring will be born able to code or design wireframes.
In other words, the various elegant models you have trained throughout your life will eventually go offline (model retirement).
But don’t be sad about this, as life is originally a process of experiencing beauty and excitement.
People often say that learning is a lifelong journey, and such a life is complete and worth looking forward to.
In summary, the Ant Financial AI Platform gathers many elites from the fields of machine learning, big data intelligence, and more from all over the world. The department’s products support many core products and businesses of Ant Financial.
This team is not only talented and knowledgeable but also passionate, loyal, and fun!
We urgently need students for the following positions:
Big Data Intelligence – Senior Product Expert:
https://job.alibaba.com/zhaopin/position_detail.htm?positionId=48713
AI/Machine Learning – Senior Product Expert:
https://job.alibaba.com/zhaopin/position_detail.htm?positionId=24972
Financial Intelligence Platform – Product Operations Expert:
https://job.alibaba.com/zhaopin/position_detail.htm?positionId=49973
If interested, please send your resume quickly (you can also send it directly to [email protected]).
“AI + Finance”, you cannot miss it!
Source: Self-Propagation Laboratory
You may also like
Click on the images below to read
How does an engineer dad prepare gifts for Children’s Day?
Do programmers rely on youth? Essentially depends on……
Why are Alibaba engineers posting their code on the intranet?
Follow “Alibaba Technology”
Keep up with cutting-edge technology trends