AI is a buzz word almost everywhere, auto-pilot, face recognition, Siri, you name it. AI gains the abilities to do so many tasks from past human experience, in other word data. AI learns from human and creates a complex neuron network to do stuff. It still not very clear why the neuron network can do that but it just does. The scientists are still trying to decode the secret of the system.
AI also gives us panic as it can replace human in many jobs, especially these highly repetitive, routine jobs dealing with information, an assistant checking the X-ray, a paralegal get information from a massive amount of evidence. AI is a double-edged sword, at least to some extent.
AI learned a great deal from human but can we also benefit from AI Intelligence? Yes, we can. The math behind machine learning and deep learning network is not very complicated. The system might have up to millions of parameters. It keeps tuning these parameters by a piece of data with an algorithm called gradient descent. It finds a direction that can effectively reduce the cost and keeps moving in this direction. In 2-dimension space, it just as simple as that. Gets a little bit complex in 3-dimension spaces but think a bowl and get to the bottom of the bowl shape.
It's obvious but what if in the space of 10 dimensions, 20 dimensions, that so many factors can influence the final result, which is our daily life. We have to balance many things in our hand and try to reach an optimal result. I know everyone has different ways, but the essential AI algorithm also comes into play. It also has a name: Gradient Heuristics
You are trying to optimize something, to make it perfect. There're many factors. How do you find the best way to do that. Do a baby step in each direction and see one the more active, that's the gradient to optimize the cost. Then go for it. Got me very careful getting closer to the point. That's what the Gradient descent also goes. Significant steps and then baby steps.
Like making soup, if it tastes mild, you want to make it salty. How do you do? Add salt because this is the most effective way to make it salty. You may add a couple of spoons at the beginning, better but not enough. Then add half spoons and taste until getting to the point. In this way, you make a soup with a perfect level of salt.
Ok, This is the first we learned from AI algorithm. Test with a baby step and find the gradient, the most effective direction to do and go for it. But this is not enough. We may fall into local optimal. We are looking at this diagram.
If you start from point A, with the gradient algorithm, you will get to the point but missing the other end which is much higher. Think about you are working in IBM 20 years ago and got an offer from Facebook. You got a good career in the IBM looking forward. Are you going to consider the offer from Facebook? It makes a lot of sense to stay but the offer might a life-changing opportunity.
The AI solves this problem by randomly select some initial points. In life, it's called Error-Allowing Heuristic. We can't 'decide the initial positions in life, but we can get prepared for the unexpecting. We got to accept the unexpecting in our life. Might look doom but an excellent start to the other peak. Particularly for young people, trying new things increase your odds to get to a higher point. Let them try, fail and learn until finding a direction, and then you need to work hard and get to the optimal.
We learned two tips from AI:
1. Gradient Heuristics
2. Error-Allowing Heuristics
Hope it helps in your life. Thanks for listening