206 316 This article is about 1400 words Estimated reading time 5 minutes Since ChatGPT became popular all over the world, artificial intelligence has regained everyone's attention. As a product designer in the field of artificial intelligence, I have naturally been following the development of this field. Recently, I finished reading the book "Generative Artificial Intelligence" written by Dr. Ding Lei, and through this book, I share the ideas in the book and give my own understanding, so that I can get familiar with it with readers and friends.
The author Ding Lei is not Ding Lei of NetEase, but another person with the same name and surname: former chief data scientist of finance and former head of PayPal's scientific data science department. He also holds a Ph.D. in artificial intelligence from The Ohio State University.
Before understanding generative AI, we can first divide the current mainstream AI technology into two categories, namely decision-making AI and generative AI.
The specific differences can be seen in the table below
First of all, for decision-making AI, the book explains it like this:
Decision-making AI (also known as discriminative AI) learns the conditional probability distribution in the data, that is, the probability that a sample belongs to a specific category, and then judges, analyzes, and analyzes new scenarios. There are several main application areas of decision-making AI: face recognition, recommendation systems, risk control systems, other intelligent decision-making systems, robotics, and autonomous driving.To put it simply, the current artificial intelligence model is a formula, and what parameters this function has, such as how to identify cats and dogs, through logical derivation, there is almost no way to know the parameters of this formula, then it is necessary to provide a large number of cats and dogs to train this AI model, tell it what cats and dogs should look like, let the computer learn by itself, extract features, and this uses the concept of combining brain science and computer science, by simulating the mechanism of brain neurons, An artificial neural network based on deep learning has been built, and by constantly feeding it to train **or**, the AI model has become more and more aware of how to recognize cats and dogs, and thus become as accurate as humans. It took millions of years for humans to learn the ability to quickly recognize cats and dogs, and so did others. And computers catch up with this recognition ability by training them in numbers and painstakingly.For example, in the field of face recognition, decision-making AI extracts feature information from face images obtained in real time, and then matches them with feature data in the face database to achieve face recognition.
Specific recognition generates specific artificial neural network implementation, for example, object detection is usually used to identify cats and dogs, and the YOLO series based on convolutional neural networks is realized. Decision-making AI is usually a small model, and through a small amount of data training, it can quickly obtain a domain-specific recognition model, but it can only do one thing well.
Generative AI, on the other hand, has the ability to generate new content, in addition to still using deep learning models to train concepts. The first outbreak is LLM, a large language model, which has trained a model with knowledge of all walks of life by feeding artificial intelligence a large number of human books, knowledge and other data, and reorganizing knowledge through algorithms, and the so-called reorganization of computers is to carry out conditional probability statistics on the next word of each word, in essence, statistics. For example, if you ask it, who was the first emperor in Chinese history? It will understand every word you say one by one, and then through the knowledge base it masters, find the probability of the next word of similar words, such as finding for example the first emperor is Qin, the probability of the second word is the beginning is higher, and the probability of the third word being the emperor is still very high, and it can continue to expand the content and continue to fill in the life and deeds of Qin Shi Huang. This is different from a search engine, which only finds the web pages and texts of the corresponding keywords in a massive database, and only some of them can be found in the database. Generative AI, on the other hand, can be understood as breaking up all knowledge and combining it at will, so that it has the ability to make up stories, provided that you tell the large language model what you want it to do, define its identity, and don't spread to the point where there are no boundaries. If you ask about medical knowledge, tell it to play the role of a doctor, so that it can refocus the scattered knowledge base and classify and reorganize medical-related knowledge.
In the same way, for AI generation, it is to disassemble each character and replace it with the pixel of dismantling each character, and calculate the probability distribution of adjacent pixels, so as to obtain a new ** based on text or input.
Based on the time relationship, let's start with more content today, and let's talk slowly in the next few days. If you have any questions about artificial intelligence, please leave a message to discuss and learn. More content, welcome to pay attention to WeChat*** Wu Yan is silent 0123 below, you can enjoy a few more AI-generated paintings, creation tools include: Midjourney, Shang Tang second painting, Ali Tongyi Wanxiang, Wenxin Yige, etc., sometimes it feels like generative AI painting is like a blind box game, even if you limit the prompt word, it is still different every time it is generated**, and it can be true, it is unique, and this is one of its charms: prompt words: artificial intelligence, recognize cats, Dogs, everything, humans benefit from it, masterpieces, rich in detail, high quality
For more information on the use of SenseTime and Ali Tongyi, please refer to the article:
Domestic AI drawing tools big PK: SenseTime second painting, Ali Tongyi Wanxiang and Wenxin Yige, whose drawing effect is more amazing?
For more information on how to use Wenxin Yige, please refer to the article:
Recommend a free and easy-to-use AI drawing artifact.