abstracts[] |
{'sha1': 'b27fe1925239d6f5552423450f4948eaa52d25be', 'content': 'Inspired by the adaptation phenomenon of biological neuronal firing, we\npropose regularity normalization: a reparameterization of the activation in the\nneural network that take into account the statistical regularity in the\nimplicit space. By considering the neural network optimization process as a\nmodel selection problem, the implicit space is constrained by the normalizing\nfactor, the minimum description length of the optimal universal code. We\nintroduce an incremental version of computing this universal code as normalized\nmaximum likelihood and demonstrated its flexibility to include data prior such\nas top-down attention and other oracle information and its compatibility to be\nincorporated into batch normalization and layer normalization. The preliminary\nresults showed that the proposed method outperforms existing normalization\nmethods in tackling the limited and imbalanced data from a non-stationary\ndistribution benchmarked on computer vision tasks. As an unsupervised attention\nmechanism given input data, this biologically plausible normalization has the\npotential to deal with other complicated real-world scenarios as well as\nreinforcement learning setting where the rewards are sparse and non-uniform.\nFurther research is proposed to discover these scenarios and explore the\nbehaviors among different variants.', 'mimetype': 'text/plain', 'lang': 'en'}
|
container |
|
container_id |
|
contribs[] |
{'index': 0, 'creator_id': None, 'creator': None, 'raw_name': 'Baihan Lin', 'given_name': None, 'surname': None, 'role': 'author', 'raw_affiliation': None, 'extra': None}
|
ext_ids |
{'doi': None, 'wikidata_qid': None, 'isbn13': None, 'pmid': None, 'pmcid': None, 'core': None, 'arxiv': '1902.10658v4', 'jstor': None, 'ark': None, 'mag': None, 'doaj': None, 'dblp': None, 'oai': None, 'hdl': None}
|
files[] |
{'state': 'active', 'ident': 'rh5rssoyrvfgbj42my24i3gph4', 'revision': 'f3078898-95be-4065-ab9b-459e077e6b93', 'redirect': None, 'extra': None, 'edit_extra': None, 'size': 278658, 'md5': '0f48300560d137c1322df4c513b4c236', 'sha1': '7d0c49df8a744c37de418c4345883a2dcd30c8a5', 'sha256': '9da2f74acd2a3dadc30b3f980e42aad26b426e2d3a8347784daa42a1310b33c1', 'urls': [{'url': 'https://arxiv.org/pdf/1902.10658v4.pdf', 'rel': 'repository'}, {'url': 'https://web.archive.org/web/20200823211747/https://arxiv.org/pdf/1902.10658v4.pdf', 'rel': 'webarchive'}], 'mimetype': 'application/pdf', 'content_scope': None, 'release_ids': ['g2icobif55garboe5qzsryyp6m'], 'releases': None}
|
filesets |
[]
|
issue |
|
language |
en
|
license_slug |
ARXIV-1.0
|
number |
|
original_title |
|
pages |
|
publisher |
|
refs |
[]
|
release_date |
2019-03-30
|
release_stage |
submitted
|
release_type |
article
|
release_year |
2019
|
subtitle |
|
title |
Regularity Normalization: Constraining Implicit Space with Minimum
Description Length
|
version |
v4
|
volume |
|
webcaptures |
[]
|
withdrawn_date |
|
withdrawn_status |
|
withdrawn_year |
|
work_id |
2ulnzgzuyzelrhbls23xmvmyt4
|