Break into the bottom layer of AI! NUS Youyang's team uses diffusion model to construct neural network parameters, LeCun likes it

The diffusion model has ushered in a major new application -

Just like Sora generates videos, it generates parameters for the neural network and directly penetrates into the bottom layer of AI!

This is the latest open source research result of Professor You Yang’s team at the National University of Singapore, together with UCB, Meta AI Laboratory and other institutions.

Specifically, the research team proposed a diffusion model p(arameter)-diff for generating neural network parameters.

Use it to generate network parameters, the speed is up to 44 times faster than direct training, and the performance is not inferior.

After the model was released, it quickly aroused heated discussions in the AI ​​community. Experts in the circle expressed the same amazing attitude towards it as ordinary people did when they saw Sora.

Some people even directly exclaimed that this is basically equivalent to AI creating new AI.

Even AI giant LeCun praised the result after seeing it, saying it was really a cute idea.

In fact, p-diff does have the same significance as Sora. Dr. Fuzhao Xue (Xue Fuzhao) from the same laboratory explained in detail:

Sora generates high-dimensional data, i.e. videos, which makes Sora a world simulator (close to AGI from one dimension).

And this work, neural network diffusion, can generate parameters in the model, has the potential to become a meta-world-class learner/optimizer, moving towards AGI from another new important dimension.

Getting back to the subject, how does p-diff generate neural network parameters?

Combining the autoencoder with the diffusion model

To clarify this problem, we must first understand the working characteristics of the diffusion model and the neural network.

The diffusion generation process is a transformation from a random distribution to a highly specific distribution. Through the addition of compound noise, the visual information is reduced to a simple noise distribution.

Neural network training also follows this transformation process and can also be degraded by adding noise. Inspired by this feature, researchers proposed the p-diff method.

From a structural point of view, p-diff is designed by the research team based on the standard latent diffusion model and combined with the autoencoder.

The researcher first selects a part of the network parameters that have been trained and performed well, and expands them into a one-dimensional vector form.

An autoencoder is then used to extract latent representations from the one-dimensional vector as training data for the diffusion model, which can capture the key features of the original parameters.

During the training process, the researchers let p-diff learn the distribution of parameters through forward and reverse processes. After completion, the diffusion model synthesizes these potential representations from random noise like the process of generating visual information. .

Finally, the newly generated latent representation is restored to network parameters by the decoder corresponding to the encoder and used to build a new model.

The following figure is the parameter distribution of the ResNet-18 model trained from scratch using 3 random seeds through p-diff, showing the differences between different layers and the same layer. distribution pattern among parameters.

To evaluate the quality of the parameters generated by p-diff, the researchers used 3 types of neural networks of two sizes each on 8 data sets. taking the test.

In the table below, the three numbers in each group represent the evaluation results of the original model, the integrated model and the model generated with p-diff.

As can be seen from the results, the performance of the model generated by p-diff is basically close to or even better than the original model trained manually.

In terms of efficiency, without losing accuracy, p-diff generates ResNet-18 network 15 times faster than traditional training, and generates Vit-Base 44 times faster.

Additional test results demonstrate that the model generated by p-diff is significantly different from the training data.

As can be seen from the figure (a) below, the similarity between the models generated by p-diff is lower than the similarity between the original models, as well as the similarity between p-diff and the original model.

It can be seen from (b) and (c) that compared with fine-tuning and noise addition methods, the similarity of p-diff is also lower.

These results show that p-diff actually generates a new model, rather than just memorizing training samples. It also shows that it has good generalization ability and can generate new models that are different from the training data.

Currently, the code of p-diff has been open sourced. If you are interested, you can check it out on GitHub.

Paper address: https://arxiv.org/abs/2402.13144

GitHub: https ://github.com/NUS-HPC-AI-Lab/Neural-Network-Diffusion

The above is the detailed content of Break into the bottom layer of AI! NUS Youyang's team uses diffusion model to construct neural network parameters, LeCun likes it. For more information, please follow other related articles on the PHP Chinese website!