Turing winner Yoshua Bengio Posted by explaining the relationship between the generating flow network and the generation model

Recently, Yoshua Bengio, one of the top AI experts and one of the Turing Award winners, has introduced the connection between GFLOWNET and GFLOWNET (Generative Flow Network) and deep generation models.

GFLOWNET is a new network generation method proposed by Benjio, which involves "strengthening learning, deep -generating models and energy -based probability modeling", and it also has a certain connection with the change model and reasoning.

Benjio mentioned on his personal website that he rarely was so enthusiastic about the new research direction, one of which was GFLOWNET.

Picture | Yoshua Bengio (Source: Benji Personal Website)

This research paper was submitted on ARXIV on September 6th, with the title of "Unifying Generative Models with GFLOWNETs" (unified by GFLOWNET ".

First of all, make a brief introduction to Benjio. He is a professor at the Department of Computer and Operational Science at the University of Montreal, Canada, and the founder and scientific director of the Mira-Quebec Institute of Artificial Intelligence.

He is considered one of the most people who promote the development of deep learning in the 1990s and 2000, and became the highest computer scientist in the world in 2022.

In 2018, due to the pioneering work and important contributions in deep learning, Benji and the professor of computer science at the University of Toronto, Geoffrey Hinton, Vice President and Chief AI Scientist of Meta, Yann Lecun ) He won the Turing Award from the International Computer Society (usually called the "Nobel Calculation Award"). The three of them are sometimes called "artificial intelligence godfather" and "deep learning godfather".

It is understood that Benjio received a PhD in computer science from McGill University in 1991, and then served as a postdoctoral researcher at Massachusetts Institute of Technology and AT & T Bell Laboratory. Since joining the University of Montreal in 1993. The authors such as "Deep Learning" (Deep Learning) and "Adaptive Computation and Machine Learning)," Towards Biology Plasible Deep Learning ".

In 2021, Benjio published an important paper "GFLOWNET FOUNDATIONS" (GFLOWNET foundation) about GFLOWNET as a work.

At present, GFlowNet has been introduced into an active learning environment to sample various candidate collection. It also provides a new field of vision for non -parameter Bayesian modeling and abstract supervision learning. "Its training is to make them approximately sampling proportional to the given reward function." The paper mentioned in the paper.

In addition to the interpretation cause and effect factors and related mechanisms, GFLOWNET is particularly helpful for the implementation system induction deviation. GFLOWNET is also a new and difficult research field. In order to understand and apply it, appropriate optimization algorithms are still developing rapidly. The concept is gradually expanding.

In this study, the paper mentioned: "There are many frameworks in the depth of modeling modeling, and each framework has its own specific training algorithm and reasoning method. The connection between the GFLOWNET framework gives a unified point of view. This provides a method for unified training and reasoning algorithms, and provides a path for constructing a formation model aggregation. "

From the perspective of probability modeling, GFLOWNET is a generating model. The purpose is to sample X based on the proportion of a given reward function R (x).

Specifically, a GFLOWNET will sample the trajectory of a Malcov with a length N τ = (S0, S1, ..., SN). If there is no special specification, the symbol X = SN will be used to represent the final state of the trajectory.

This process has a natural connection with strengthening learning. All state S constructs a non -circulating diagram in the potential state space. Each trajectory starts from the same (abstract) initial state S0 and runs to a different endpoint SN. Ideally, the traffic that hopes to go X is equal to given rewards.

In the "Learning Reward Function Function Function" in the paper, the research team mentioned: "Energy-based models (EBM, Energy-Based Model) can be used as a (negative) reward function for GFLOWNET training. We can use any GFLOWNETt Modeling, and two models (EBM and GFLOWNET) joint training. "

In addition, the GANRATIVE Adversarial Network is closely related to EBM, but the calculation efficiency of its algorithm is higher. However, although it may be reasonable at first glance, it cannot be used directly as a reward for GFLOWNET training. If so, at the end of a perfect training, a optimal discriminator and optimal GFLOWNET generator will be obtained. In order to fill this gap, Benjio has designed some more meaningful algorithms.

(Source: Benji Personal Website)

The picture above illustrates why the word "flow" is used in GFLOWNET. This takes into account the flow of standardized probability, similar to the amount of water flowing from the initial state (0) from the initial state (on the left) in the computer, which may be in the computer. All possible action sequences (that is, actions that determine the state conversion) in order to build complex objects in order, such as molecular diagrams, causality, explanation of scenes, or our minds.

The final conclusion of the paper mentioned: "Today's generating model can be understood as GFLOWNETs with differentiated strategies on the sample trajectory. This is the overlapping part of the existing modeling framework, and the relationship between the general algorithm trained by training them. , Provided some views.

This unity means a method of generating modeling different types of clusters. Due to the superiority of reasoning and training, GFLOWNET can be used as a general adhesive. "

Reference materials: https:////abs/2209.02606https: //yoshuabngio.org/2022/03/05/genlant- flow- netWorks: //en.wikiped/wiki/yhua_bangio

- END -