{"about":{"site":"https://codewithpapers.app","non_affiliation":"Code with Papers and Syntology are not affiliated with, endorsed by, or sponsored by Papers with Code, Meta, or the pwc-archive mirror.","licence":"CC BY-SA 4.0","licence_url":"https://creativecommons.org/licenses/by-sa/4.0/legalcode","attribution":"https://codewithpapers.app/attribution","modified":"archive material modified by Syntology; see the attribution page"},"url":"/paper/variational-walkback-learning-a-transition","title":"Variational Walkback: Learning a Transition Operator as a Stochastic Recurrent Net","arxiv_id":"1711.02282","date":"2017-11-07","proceeding":"NeurIPS 2017 12","authors":["Anirudh Goyal","Nan Rosemary Ke","Surya Ganguli","Yoshua Bengio"],"abstract":"We propose a novel method to directly learn a stochastic transition operator\nwhose repeated application provides generated samples. Traditional undirected\ngraphical models approach this problem indirectly by learning a Markov chain\nmodel whose stationary distribution obeys detailed balance with respect to a\nparameterized energy function. The energy function is then modified so the\nmodel and data distributions match, with no guarantee on the number of steps\nrequired for the Markov chain to converge. Moreover, the detailed balance\ncondition is highly restrictive: energy based models corresponding to neural\nnetworks must have symmetric weights, unlike biological neural circuits. In\ncontrast, we develop a method for directly learning arbitrarily parameterized\ntransition operators capable of expressing non-equilibrium stationary\ndistributions that violate detailed balance, thereby enabling us to learn more\nbiologically plausible asymmetric neural networks and more general non-energy\nbased dynamical systems. The proposed training objective, which we derive via\nprincipled variational methods, encourages the transition operator to \"walk\nback\" in multi-step trajectories that start at data-points, as quickly as\npossible back to the original data points. We present a series of experimental\nresults illustrating the soundness of the proposed approach, Variational\nWalkback (VW), on the MNIST, CIFAR-10, SVHN and CelebA datasets, demonstrating\nsuperior samples compared to earlier attempts to learn a transition operator.\nWe also show that although each rapid training trajectory is limited to a\nfinite but variable number of steps, our transition operator continues to\ngenerate good samples well past the length of such trajectories, thereby\ndemonstrating the match of its non-equilibrium stationary distribution to the\ndata distribution. Source Code: http://github.com/anirudh9119/walkback_nips17","url_abs":"http://arxiv.org/abs/1711.02282v1","url_pdf":"http://arxiv.org/pdf/1711.02282v1.pdf","source":{"archive":"pwc-archive (Hugging Face), CC BY-SA 4.0","snapshot":"2025-07-28","licence_url":"https://creativecommons.org/licenses/by-sa/4.0/legalcode","row_kind":"abstracts"},"code_links":[{"paper_slug":"variational-walkback-learning-a-transition","repo_url":"https://github.com/anirudh9119/walkback_nips17","is_official":1,"mentioned_in_paper":1,"mentioned_in_github":0,"framework":"none","reach":{"status":"ok"}}],"tasks":[],"methods":[],"datasets_introduced":[],"methods_introduced":[],"results":[],"syntology":{"atlas_url":"https://app.syntology.ai/?focus=1711.02282","mcp":null,"developers":"https://syntology.ai/developers"},"arxiv_metadata":null,"syntology_extracted_results":null}