Gang Fang's Blog

What is energy-based modeling? My answer after brief reading: the fundamental idea of EBM is to interpret a system in terms of energy. The energy function, which outputs a scalar value, defines a system’s state. The lower the energy, the more stable, desirable or likely the system. In optimization, the energy function is similar to…

Note: a trained GFN is both a sampler (i.e., generating compositional objects) and an inference machine (i.e., answering questions and predicting probabilities) How does a parametrized energy function look like? TO_BE_ANSWERED. But the training of it can be done with classical maximum likelihood. we have shown how we can jointly train the GFlowNet sampler and…

The kind of sequential probabilistic inference that GFlowNets can perform is a powerful form of learned reasoning machinery, which could be used for interpretation of sensory inputs, interpretation of selected past observations, planning, and counterfactuals https://milayb.notion.site/The-GFlowNet-Tutorial-95434ef0e2d94c24aab90e69b30be9b3#208ee566b55048cda2c87fd5e0e93330 These are the potential applications of GFN. However, I think if they are all achieved then AGI is essentially…

Note: updated mental map of GFlowNet How does GFlowNet approximate human reasoning? (high level, from Bayesian and Variational inference perspectives) Note: training GFlowNet What does amortization mean in general CS and under the context of GFlowNet? My answer before reading: amortization in CS means spreading the computational cost among a series of computations. An example…

In regular GFlowNets, choosing a_t from s_t deterministically yields some s_{t+1}, which means that we can also write \pi(a_t|s_t)=P_F(s_{t+1}|s_t) for that policy Doesn’t choosing a_t from s_t deterministically mean P_F(s_{t+1}|s_t)=1? What do we still need \pi(a_t|s_t)=P_F(s_{t+1}|s_t)? My answer with ChatGPT help: \pi(a_t|s_t) remains a stochastic policy given P_F. The deterministic element is instead the equation…

How is this done and what do “theories shared across examples” refer to? This makes GFlowNets amortized probabilistic inference machines that can be used both to sample latent variables (as in [14]) or parameters and theories shared across examples (as in [5]). TO_BE_ANSWERED. How is approximate and amortized marginalization done with GFlowNet? They can also…