The next generation of artificial intelligence: logical understanding? Physical understanding?

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 came to a successful conclusion at the Shanghai World Expo Center. As one of the board members of the LF AI & Data sub-foundation, Liam Zheng, senior technical expert of OPPO's digital intelligence engineering system, published a titled "The Next Generation of Artificial Intelligence: Logical Understanding?" at the GOTC "AI Big Data and Digital Economy" sub-forum. Physical understanding? "Speech. This article shares his views and understanding of the next generation of artificial intelligence through an interview with Liam.

Q1: What is the background of the speech "The core of the next generation of artificial intelligence is logical understanding and physical understanding" in this GOTC sub-forum?
OPPO has joined the LF AI & Data sub-foundation, and we look forward to working with it on open source projects. Before our open source projects come out, we also need some warm-up. In addition, the development of artificial intelligence to the current stage, everyone found that there were a lot of bad cases after the actual deployment and online, but it was not very convenient to modify the model, usually a large amount of calibration data was required to retrain the model. In order to solve the above problems, my view is that the next generation of artificial intelligence should deeply transform the algorithm from the logical level and the physical level, rather than simply adding data or making the model larger.

Q2: Then what aspects do you mainly introduce this problem from?
At that time, I mainly talked about 4 parts. The first part is the bottleneck currently faced by artificial intelligence; the second part is to introduce some of the views of the industry's majors on the next generation of artificial intelligence; the third part is my comparative analysis of human intelligence and artificial intelligence ; Finally came to the point that the core of the next generation of artificial intelligence should be logical understanding and physical understanding.

Q3: Compared with human intelligence, at which stage do you think artificial intelligence is?
Human intelligence actually has eight fields, and artificial intelligence currently only involves work in 2-3 of them. Data representation in most areas has not yet been involved. Therefore, the current artificial intelligence is actually in its infancy, and it is far from reaching a particularly comprehensive and complete stage.

Q4: Just mentioned the bottleneck of artificial intelligence, what do you think is the biggest bottleneck?
The two main points, one is the poor robustness, I gave an example: for example, to classify a panda picture, add some random noise, and then it becomes another category. It is a small disturbance, the results of the model judgment will be very different, and the model can even be controlled to misjudge a certain category.
Another point is the lack of interpretability. For example, sometimes the model may perform very well, and in some cases it may perform poorly, but it is not possible to locate the specific feature and layer, which causes the model to perform poorly.
The above two points were also mentioned in the lecture of "Safe and Trustworthy Artificial Intelligence" by Academician He Jifeng.

Q5: What do you think of the next-generation artificial intelligence industry?
In the speech, I introduced the opinions of several big coffees.

One is that Geoffrey E. Hinton proposed the perspective of the capsule network. He believes that the cv model should not be Invariant, it should be Equivariant, that is, it can reflect the structure of the image; the current convolution model cannot reflect the structural information in the image. Put a certain part in any position, and the result will not change. For example, if a person's eyes are randomly placed, it will be a face. But if a person sees it by himself, if the position of the person's eyes is shifted a lot, they will not look like a face at all.

The second is Yan LeCun. He 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 through self-supervised learning. Currently, machine learning mainly relies on supervised learning. Come on, this is just a small part of machine learning

The third is the view of Professor Zhu Songchun, who believes that the next generation of artificial intelligence should be the crow paradigm, solving practical problems through multi-task learning with small samples;

Finally, there is Yoshua Bengio. He believes that artificial intelligence is currently in the perception stage, and the next stage is the cognitive stage, but I think the perception stage is far from over.

Q6: Why is it said that "the core of the next generation of artificial intelligence is logical understanding and physical understanding"?
The machine learning training set and test set are based on the assumption of IID (Independent and Identical Distribution). In fact, the data estimated after the line is often OOD (different from the training set distribution). Both IID and OOD refer to the distribution of representations, and a good representation will have a good OOD effect. Although the generalization ability of deep learning is better than traditional machine learning, it also faces the problem of OOD. 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. Doing simple supervised learning on the tiny training set will only learn the local pattern of the training sample, because only the local pattern representation can achieve the IID effect of the training set and the test set, and the local pattern representation and the local pattern are far from satisfying the online After the OOD situation. In short, OOD is the essential reason for the poor robustness of current artificial intelligence.
The next generation of artificial intelligence urgently needs to solve the robustness of perception. The key lies in the logical and physical understanding of representation and training, rather than the super-large model and super-large data.

For more exciting content, follow the [OPPO Internet Technology] public account