Disentangling the independently controllable factors of variation by interacting with the world
read the original abstract
It has been postulated that a good representation is one that disentangles the underlying explanatory factors of variation. However, it remains an open question what kind of training framework could potentially achieve that. Whereas most previous work focuses on the static setting (e.g., with images), we postulate that some of the causal factors could be discovered if the learner is allowed to interact with its environment. The agent can experiment with different actions and observe their effects. More specifically, we hypothesize that some of these factors correspond to aspects of the environment which are independently controllable, i.e., that there exists a policy and a learnable feature for each such aspect of the environment, such that this policy can yield changes in that feature with minimal changes to other features that explain the statistical variations in the observed data. We propose a specific objective function to find such factors, and verify experimentally that it can indeed disentangle independently controllable aspects of the environment without any extrinsic reward signal.
This paper has not been read by Pith yet.
Forward citations
Cited by 15 Pith papers
-
Towards inertial-mode helioseismology: Direct sensing of solar rotation at 75 deg latitude and 0.8 Rsun
The m=1 high-latitude inertial mode frequency implies solar rotation of 365.3 nHz at 75° latitude and 0.8 R_sun, exceeding the p-mode reference by 8.1 nHz.
-
Millimeter dust continuum and polarization in protoplanetary disks with scattering: A slab model
Common analytic approximations underestimate protoplanetary disk millimeter continuum emission by 10-15%, causing overestimates of optical depth, mass, and temperature in SED analyses.
-
A new Gaia census of OB associations within 1 kpc
A new catalogue of 56 OB associations within 1 kpc is produced from Gaia data using HDBSCAN, with most showing expansion and connections to local galactic features.
-
The heart of NGC 5253 as seen with MUSE-NFM: nitrogen enrichment through stellar chemical feedback at parsec scales
High-resolution mapping reveals nitrogen enrichment around super star clusters in NGC 5253, attributed to Wolf-Rayet stars with material expelled to surrounding gas.
-
Linear toroidal-inertial waves on a differentially rotating sphere with application to helioseismology: Modeling, forward and inverse problems
Develops a fourth-order scalar equation for purely toroidal linear waves on a differentially rotating sphere, proves well-posedness under explicit rotation conditions, and demonstrates local unique identifiability plu...
-
CO(7-6) and [C I](2-1) survey in z > 6 quasars
ALMA survey of CO(7-6) and [C I](2-1) in 18 z~6 quasars yields molecular gas mass estimates from multiple tracers and indicates that PDR models cannot fully account for the observed excitation, pointing to additional ...
-
Dynamical Cluster Assembly Framework (D-CAF): The Link Between Star Cluster Formation and Expansion Rates
D-CAF simulations show that ongoing gas collapse during star formation shortens stellar crossing times, rendering gas expulsion more adiabatic and thereby regulating the survival and expansion rates of young stellar systems.
-
Tidal pre-conditioning and ram-pressure stripping in NGC 1427A. Deep VLT/MUSE spectroscopy and FUV-to-radio observations trace a Fornax Cluster dwarf in transformation
Multi-phase observations of NGC 1427A indicate tidal torquing from a dwarf fly-by has pre-conditioned its gas for ram-pressure stripping by the Fornax intracluster medium, placing the galaxy at the onset of environmen...
-
The emerging timescale of young star clusters regulated by cluster stellar mass
Massive young star clusters clear their natal gas faster than lower-mass clusters, based on HST and JWST imaging of four galaxies.
-
Scattered light signatures of flyby-induced warps in protoplanetary discs
Flyby-induced warps create broad, time-varying shadows in scattered light that persist for most of a low-viscosity disc lifetime and could be used to constrain disc viscosity.
-
The ${}^{13}\mathrm{CO}(2{-}1)/^{12}\mathrm{CO}(2{-}1)$ Line Ratio from 100 Molecular Clouds in the Large Magellanic Cloud
Observational study of 100 LMC GMCs finds median 13CO(2-1)/12CO(2-1) line ratio of 0.078, nearly linear with luminosity, and higher in clouds hosting IR-bright young stellar objects.
-
Modeling (Sub-)millimeter Scattering Properties of Fractal and Consolidated Porous Particles: Applications to Protoplanetary Disks
Numerical simulations of porous fractal and consolidated particles show stronger forward scattering, broader polarization peaks, and lower absorption per unit mass than compact spheres, implying larger dust masses fro...
-
An Improved HDBSCAN-based Detection and Tracking Method for Solar Active Regions in Magnetograms
HARDAT applies HDBSCAN clustering with differential-rotation tracking and SVM-based polarity inversion line extraction to magnetograms, claiming superior sensitivity, accuracy and identity continuity over DBSCAN-based...
-
Nuclear Activity and Host Galaxy Properties of Low-Luminosity AGN Identified from VLA Observations
VLA radio-selected LLAGN show 84% optical, 63% X-ray, and 13% infrared detection rates, with black holes ~0.7 dex smaller, accretion rates ~4.2 dex lower, and host galaxies ~0.3 dex lower in stellar mass with ~0.5 dex...
-
Tidal pre-conditioning and ram-pressure stripping in NGC 1427A. Deep VLT/MUSE spectroscopy and FUV-to-radio observations trace a Fornax Cluster dwarf in transformation
NGC 1427A exhibits ram-pressure stripping that has reached its ISM after tidal torquing by a nearby dwarf, marking the onset of cluster-driven gas loss and declining star formation.
discussion (0)
Sign in with ORCID, Apple, or X to comment. Anyone can read and Pith papers without signing in.