Recently, the first “GOTC Global Open Source Technology Summit”, jointly organized by the Open Atom Open Source Foundation and the Linux Foundation and Open Source China, was successfully concluded at the Shanghai World Expo Center. As a member of the board of directors of the LF AI & Data Sub-Foundation, Liam Zheng, a senior technical expert of OPO Digital Intelligence Engineering Systems, published a paper titled “Next Generation Artificial Intelligence: Logical Understanding? Physical Understanding?” at the GOTC sub-forum on “AI Big Data and Digital Economy”. In the speech. In this interview, Liam shares his thoughts and understanding on the next generation of artificial intelligence.
Q1: What is the background of the speech “Logic Understanding and Physical Understanding are the Core of the Next Generation of Artificial Intelligence” in this GOTC sub-forum? Oppo has joined the LF AI &Data Sub-Foundation and we are looking forward to working with them on open source projects. We also need some warm-up before our open source projects come out. In addition, with the development of artificial intelligence at the current stage, people find that after the actual deployment, they encounter a lot of badcases, but it is not very convenient to modify the model, which usually requires a lot of calibration data and retraining of the model. In order to solve the above problems, my opinion is that the next generation of AI should deeply modify algorithms on the logical and physical level, rather than simply adding data or making models bigger.
Q2: What are your main aspects to introduce this problem? At that time, I mainly talked about four parts. The first part is the bottleneck of artificial intelligence. The second part is to introduce some views of the industry masters on the next generation of artificial intelligence. The third part is a comparative analysis of human intelligence and artificial intelligence. Finally, we conclude that the core of the next generation of artificial intelligence should be logical understanding and physical understanding.
Q3: Where do you think artificial intelligence is now compared to human intelligence? There are actually eight fields of human intelligence, and currently artificial intelligence is only involved in 2-3 of them. Most of the domain data representations have not yet been covered. So the current artificial intelligence is actually in its infancy, it is far from a particularly comprehensive, particularly complete stage.
Q4: I just mentioned the bottleneck of artificial intelligence. What do you think is the biggest bottleneck? One is poor robustness. I gave an example: for example, classifying a picture of a panda becomes another category by adding some random noise. Even small perturbations can make a big difference in the results of the model judgment, and can even control the model’s misjudgment into a particular category. Another is the lack of interpretability. For example, there may be times when the model performs particularly well and times when it performs poorly, but it is impossible to pinpoint the specific feature or layer that causes the model to perform poorly. The above two points were also mentioned in the speech “Safe and Trusted Artificial Intelligence” by Academician He Jifeng.
Q5: What is the industry’s view on the next generation of AI? In the speech, I introduced the views of several famous people.
First, Geoffrey E. Hinton proposed the perspective of capsule network. He believed that CV model should not be Invariant but equivariant, which can reflect the structure of images. The current convolution model cannot reflect the structural information in the image. If a part is placed in any position, the result will remain unchanged. For example, if a person’s eyes are randomly placed, a human face will be obtained. But if you look at it yourself, your eyes, if they are too far out of position, don’t look like a face at all.
The second is Yan LeCun, who proposed that the next generation of artificial intelligence mainly relies on self-supervised learning. I basically agree with this view. The logical and physical initialization model space is achieved through self-supervised learning. At present, machine learning mainly relies on supervised learning, but in my opinion, this is only a small part of machine learning
Third, Professor Zhu Songchun believes that the next generation of artificial intelligence should be the crow paradigm, which solves practical problems through small sample multi-tasking learning.
And finally, there’s Yoshua Bengio, who thinks that AI is now the perceptual stage, and the next stage is the cognitive stage, but I think the perceptual stage is far from over.
Q6: Why is it said that “the core of the next generation of AI is logical understanding and physical understanding”? The machine learning training set and test set are based on the IID (independent and equally distributed) assumption, in fact the data estimated behind the line is often OOD (different from the training set distribution). Both IID and OOD refer to the distribution of representation, and good representation leads to good OOD effect. Although the generalization ability of deep learning is better than that of traditional machine learning, it also faces the OOD problem. When the sample space is large, the training set is always only a small part of the whole, and the distribution of the whole will be very different. Simple supervised learning on a small training set can only learn the local patterns of training samples, because IID effects of training set and test set can be obtained only by local pattern representation, but local pattern representation and local pattern are far from satisfying the OOD situation after online. In short, OOD is the essential reason for the poor robustness of current AI. The next generation of artificial intelligence urgently needs to solve the robustness of perception. The key lies in the logical understanding and physical understanding of representation and training, rather than the large model and the large data.
Get more exciting content, pay attention to [OPPO Internet technology] public number
