Causal Discovery with Energy-Based Models
Visualitza/Obre
Causal_discovery_with_energy_based_models_FernandezLopez_Christian.pdf (3,472Mb) (Accés restringit)
Estadístiques de LA Referencia / Recolecta
Inclou dades d'ús des de 2022
Cita com:
hdl:2117/383042
Realitzat a/ambNortheastern University
Tipus de documentProjecte Final de Màster Oficial
Data2022-05-26
Condicions d'accésAccés restringit per acord de confidencialitat
(embargat fins 2025-02-15T07:25:15Z)
Tots els drets reservats. Aquesta obra està protegida pels drets de propietat intel·lectual i
industrial corresponents. Sense perjudici de les exempcions legals existents, queda prohibida la seva
reproducció, distribució, comunicació pública o transformació sense l'autorització del titular dels drets
Abstract
Discovery of causal relations from observational data is a key step toward system's understanding. Standard causal discovery methods fit a new model whenever a new causal graph is presented. Recent works propose to take advantage of shared dynamics to infer causal relations to time-series data. Although this approach significantly improves causal discovery performance, it fails if a different dynamic is presented. Our approach draws from energy-based models and proposes to model the causality graph as an energy function. This allows for discriminating whether or not an input trajectory belongs to the trained data distribution or presents a different dynamic. We present a novel framework for causal discovery that uses an inference-time optimization procedure to identify the causal relations of input trajectories and generate trajectories given the identified underlying causal graph. Therefore a single model is enough to self-supervise the problem. We evaluate our framework in a simulated dynamic system where particles are connected by springs. Our approach has competitive and robust performance compared with state-of-the-art methods. In addition, we demonstrate that we can detect and identify outliers and verify if an observation belongs to the trained system.
MatèriesNeural networks (Computer science), Deep learning, Xarxes neuronals (Informàtica), Aprenentatge profund
TitulacióMÀSTER UNIVERSITARI EN TECNOLOGIES AVANÇADES DE TELECOMUNICACIÓ (Pla 2019)
Fitxers | Descripció | Mida | Format | Visualitza |
---|---|---|---|---|
Causal_discover ... rnandezLopez_Christian.pdf | 3,472Mb | Accés restringit |