國立虎尾科技大學 |

Towards Aligning Artificial and Biological Vision Systems With Self-Supervised Representation Learning.

紀錄類型:	書目-語言資料,手稿 : Monograph/item
正題名/作者:	Towards Aligning Artificial and Biological Vision Systems With Self-Supervised Representation Learning./
作者:	Parthasarathy, Nikhil.
面頁冊數:	1 online resource (160 pages)
附註:	Source: Dissertations Abstracts International, Volume: 85-09, Section: B.
Contained By:	Dissertations Abstracts International85-09B.
標題:	Neurosciences. -
電子資源:	click for full text (PQDT)
ISBN:	9798381850772

Towards Aligning Artificial and Biological Vision Systems With Self-Supervised Representation Learning.
Parthasarathy, Nikhil.

Towards Aligning Artificial and Biological Vision Systems With Self-Supervised Representation Learning. - 1 online resource (160 pages)

Source: Dissertations Abstracts International, Volume: 85-09, Section: B.

Thesis (Ph.D.)--New York University, 2024.

Includes bibliographical references

This dissertation investigates self-supervised learning methods for training deep neural network (DNN) models that better align with both human behavior and neural responses in early visual areas. We first propose VITO, an attention-guided contrastive video-pretraining method, that improves dramatically on prior work to learn general, robust, and more human-aligned representations from natural video data. We specifically demonstrate that dynamic temporal content is required for the improved robustness and human-alignment. We next explore a complementary line of work focused on improving the alignment of intermediate DNN representations with early visual areas. We first provide a simple demonstration that selectivity for visual texture can be learned via optimizing a single-layer objective in a biologically-inspired architecture modeled off of areas V1 and V2. We then refine and extend this study to a more general layerwise learning paradigm, capable of learning features simultaneously in a two-layer network. We do so by leveraging a novel self-supervised layerwise complexity-matched learning paradigm. Our trained model provides better predictions of neural responses in early visual areas and particularly achieves state-of-the-art predictions for cortical area V2. Finally, we provide some preliminary analyses probing the limitations of current regression-based evaluations for measuring alignment with neural responses. Taken together, this thesis lays the foundation for future research in using learned DNNs to reveal new organizing principles for how selectivities are formed in visual hierarchies, with potential implications for both neuroscience and machine learning.

Electronic reproduction.
Ann Arbor, Mich. :
ProQuest,
2024

Mode of access: World Wide Web

ISBN: 9798381850772Subjects--Topical Terms:

593561
Neurosciences.
Subjects--Index Terms:

computational neuroscienceIndex Terms--Genre/Form:

554714
Electronic books.

Towards Aligning Artificial and Biological Vision Systems With Self-Supervised Representation Learning.
LDR:03107ntm a22003977 4500 001 1149679
005 20241022112612.5
006 m o d
007 cr bn ---uuuuu
008 250605s2024 xx obm 000 0 eng d
020 $a 9798381850772
035 $a (MiAaPQ)AAI30814528
035 $a AAI30814528
040 $a MiAaPQ $b eng $c MiAaPQ $d NTU
100 1 $a Parthasarathy, Nikhil. $3 1476005
245 1 0 $a Towards Aligning Artificial and Biological Vision Systems With Self-Supervised Representation Learning.
264 0 $c 2024
300 $a 1 online resource (160 pages)
336 $a text $b txt $2 rdacontent
337 $a computer $b c $2 rdamedia
338 $a online resource $b cr $2 rdacarrier
500 $a Source: Dissertations Abstracts International, Volume: 85-09, Section: B.
500 $a Advisor: Simoncelli, Eero P.
502 $a Thesis (Ph.D.)--New York University, 2024.
504 $a Includes bibliographical references
520 $a This dissertation investigates self-supervised learning methods for training deep neural network (DNN) models that better align with both human behavior and neural responses in early visual areas. We first propose VITO, an attention-guided contrastive video-pretraining method, that improves dramatically on prior work to learn general, robust, and more human-aligned representations from natural video data. We specifically demonstrate that dynamic temporal content is required for the improved robustness and human-alignment. We next explore a complementary line of work focused on improving the alignment of intermediate DNN representations with early visual areas. We first provide a simple demonstration that selectivity for visual texture can be learned via optimizing a single-layer objective in a biologically-inspired architecture modeled off of areas V1 and V2. We then refine and extend this study to a more general layerwise learning paradigm, capable of learning features simultaneously in a two-layer network. We do so by leveraging a novel self-supervised layerwise complexity-matched learning paradigm. Our trained model provides better predictions of neural responses in early visual areas and particularly achieves state-of-the-art predictions for cortical area V2. Finally, we provide some preliminary analyses probing the limitations of current regression-based evaluations for measuring alignment with neural responses. Taken together, this thesis lays the foundation for future research in using learned DNNs to reveal new organizing principles for how selectivities are formed in visual hierarchies, with potential implications for both neuroscience and machine learning.
533 $a Electronic reproduction. $b Ann Arbor, Mich. : $c ProQuest, $d 2024
538 $a Mode of access: World Wide Web
650 4 $a Neurosciences. $3 593561
650 4 $a Computer science. $3 573171
650 4 $a Biomedical engineering. $3 588770
653 $a computational neuroscience
653 $a Deep learning
653 $a Self-supervised learning
653 $a Vision
653 $a Visual representations
655 7 $a Electronic books. $2 local $3 554714
690 $a 0317
690 $a 0984
690 $a 0541
710 2 $a ProQuest Information and Learning Co. $3 1178819
710 2 $a New York University. $b Center for Neural Science. $3 1180547
773 0 $t Dissertations Abstracts International $g 85-09B.
856 4 0 $u http://pqdd.sinica.edu.tw/twdaoapp/servlet/advanced?query=30814528 $z click for full text (PQDT)