Fig. 2: Learning the dynamics of the lattice active-matter model.
From: Learning stochastic dynamics and predicting emergent behavior using transformers
a Training of a transformer in Mode 1 (unrestricted rates) to maximize the log-likelihood \({U}_{\omega }^{({{{{{{{\boldsymbol{\theta }}}}}}}})}\), Eq. (1), of the training trajectory ω. The horizontal black line denotes the value of the path weight associated with the original model. b Dependence of \({U}_{\omega }^{({{{{{{{\boldsymbol{\theta }}}}}}}})}\) for a transformer trained in Mode 2, in which it is asked to identify \({N}_{W}^{({{{{{{{\boldsymbol{\theta }}}}}}}})}\) distinct classes of move. This procedure allows us to identify the existence of \({N}_{W}^{\star }=4\) distinct rates. Inset: Evolution of the rates during training in Mode 2, with \({N}_{W}^{({{{{{{{\boldsymbol{\theta }}}}}}}})}=4\). The horizontal black lines denote the values of the rates in the original dynamics.
