Compositionality Unlocks Deep Interpretable Models

Thomas Dooms; Ward Gauderis; Geraint Wiggins; Jose Antonio Oramas Mogrovejo

Compositionality Unlocks Deep Interpretable Models

Thomas Dooms, Ward Gauderis, Geraint Wiggins, Jose Antonio Oramas Mogrovejo

Onderzoeksoutput: Conference paper

Samenvatting

We propose $\chi$-net, an intrinsically interpretable architecture combining the compositional multilinear structure of tensor networks with the expressivity and efficiency of deep neural networks. $\chi$-nets retain SoTA accuracy across tasks. Our novel, efficient diagonalisation algorithm, ODT, reveals linear low-rank structure in a multilayer SVHN model. We use this structure for weight-based interpretability and model compression.

Originele taal-2	English
Titel	Compositionality Unlocks Deep Interpretable Models
Uitgeverij	AAAI
Aantal pagina's	10
Status	Published - 4 mrt. 2025
Evenement	Connecting Low-Rank Representations in AI : At the 39th Annual AAAI Conference on Artificial Intelligence - Pennsylvania Convention Center, Philadelphia, United States Duur: 3 mrt. 2025 → 4 mrt. 2025 Congresnummer: 1 https://april-tools.github.io/colorai/

Workshop

Workshop	Connecting Low-Rank Representations in AI
Verkorte titel	CoLoRAI
Land/Regio	United States
Stad	Philadelphia
Periode	3/03/25 → 4/03/25
Internet adres	https://april-tools.github.io/colorai/

Bibliografische nota

Both authors, Thomas Dooms and Ward Gauderis, contributed equally.

Toegang tot document

https://openreview.net/forum?id=bXAt5iZ69lLicentie: CC BY

1 Poster

Compositionality Unlocks Deep Interpretable Models
Gauderis, W. & Dooms, T., 4 mrt. 2025. 1 blz.
Onderzoeksoutput: Poster

Open Access
Bestand

1 Participation in conference
1 Participation in workshop, seminar
1 Talk or presentation at a workshop/seminar

Introducing X-nets: Deep Interpretability without Compromise
Thomas Dooms (Speaker) & Ward Gauderis (Speaker)
4 mrt. 2025
Activiteit: Talk or presentation at a workshop/seminar
The 39th Annual AAAI Conference on Artificial Intelligence
Ward Gauderis (Participant)
25 feb. 2025 → 4 mrt. 2025
Activiteit: Participation in conference
Connecting Low-Rank Representations in AI
Ward Gauderis (Participant)
4 mrt. 2025
Activiteit: Participation in workshop, seminar

Citeer dit

@inproceedings{bf6e870924d24efeb04d04a4f8256a67,

title = "Compositionality Unlocks Deep Interpretable Models",

abstract = "We propose $\chi$-net, an intrinsically interpretable architecture combining the compositional multilinear structure of tensor networks with the expressivity and efficiency of deep neural networks. $\chi$-nets retain SoTA accuracy across tasks. Our novel, efficient diagonalisation algorithm, ODT, reveals linear low-rank structure in a multilayer SVHN model. We use this structure for weight-based interpretability and model compression.",

keywords = "Neural Network, Tensor Network, Mechanistic Interpretability, Compositional AI, Model Interpretation, Model Compression, low-rank and joint-sparse model",

author = "Thomas Dooms and Ward Gauderis and Geraint Wiggins and {Oramas Mogrovejo}, {Jose Antonio}",

note = "Both authors, Thomas Dooms and Ward Gauderis, contributed equally.; Connecting Low-Rank Representations in AI : At the 39th Annual AAAI Conference on Artificial Intelligence, CoLoRAI ; Conference date: 03-03-2025 Through 04-03-2025",

year = "2025",

month = mar,

day = "4",

language = "English",

booktitle = "Compositionality Unlocks Deep Interpretable Models",

publisher = "AAAI",

url = "https://april-tools.github.io/colorai/",

}

TY - GEN

T1 - Compositionality Unlocks Deep Interpretable Models

AU - Dooms, Thomas

AU - Gauderis, Ward

AU - Wiggins, Geraint

AU - Oramas Mogrovejo, Jose Antonio

N1 - Conference code: 1

PY - 2025/3/4

Y1 - 2025/3/4

N2 - We propose $\chi$-net, an intrinsically interpretable architecture combining the compositional multilinear structure of tensor networks with the expressivity and efficiency of deep neural networks. $\chi$-nets retain SoTA accuracy across tasks. Our novel, efficient diagonalisation algorithm, ODT, reveals linear low-rank structure in a multilayer SVHN model. We use this structure for weight-based interpretability and model compression.

AB - We propose $\chi$-net, an intrinsically interpretable architecture combining the compositional multilinear structure of tensor networks with the expressivity and efficiency of deep neural networks. $\chi$-nets retain SoTA accuracy across tasks. Our novel, efficient diagonalisation algorithm, ODT, reveals linear low-rank structure in a multilayer SVHN model. We use this structure for weight-based interpretability and model compression.

KW - Neural Network

KW - Tensor Network

KW - Mechanistic Interpretability

KW - Compositional AI

KW - Model Interpretation

KW - Model Compression

KW - low-rank and joint-sparse model

M3 - Conference paper

BT - Compositionality Unlocks Deep Interpretable Models

PB - AAAI

T2 - Connecting Low-Rank Representations in AI

Y2 - 3 March 2025 through 4 March 2025

ER -

Compositionality Unlocks Deep Interpretable Models

Samenvatting

Workshop

Bibliografische nota

Toegang tot document

Onderzoekersoutput

Compositionality Unlocks Deep Interpretable Models

Activiteiten

Introducing X-nets: Deep Interpretability without Compromise

The 39th Annual AAAI Conference on Artificial Intelligence

Connecting Low-Rank Representations in AI

Citeer dit