Meta-Learning Sparse Compression Networks

Jonathan Richard Schwarz; Yee Whye Teh

RECORD

Meta-Learning Sparse Compression Networks

Machine Learning · Submitted 18 May 2022

abstract compiled · 2026-09-23 · no language-model review

The authors' main claim The generality of this formalism allows us to present results on diverse data modalities such as images, manifolds, signed distance functions, 3D shapes and scenes, several of which establish new state-of-the-art results.

how it is stated · asserted 1.00 read the record →

Comment (0) Read on arXiv ↗

Cite ▾

Claim paper

The compiled record

The paper's own sentences, arranged. Each label is drawn from a closed vocabulary; the number beside it is the probability the classifier gave that label. Numbers in the tables are the authors', checked by code. No sentence in this record was written by a model.

Open the compiled record

Compiled by classifier-1.13.0 via TypeSafe API, question bank abstract-p1-20260923, 2026-09-23: 1 request, 5,044 input tokens.

What the abstract itself states, as probabilities: claims first 0.87 · states limitation 0.13 · human participants 0.02 · controlled comparison 0.51 · randomized assignment 0.01 · states quantitative result 0.05 · compares with prior results 0.91 · mentions code or data availability 0.03.

The abstract, sentence by sentence

arXiv listing abstract v2. Results and the main claim are underlined; show every role.

Recent work in Deep Learning has re-imagined the representation of data as functions mapping from a coordinate space to an underlying continuous signal. When such functions are approximated by neural networks this introduces a compelling alternative to the more common multi-dimensional array representation. Recent work on such Implicit Neural Representations (INRs) has shown that - following careful architecture search - INRs can outperform established compression methods such as JPEG (e.g. Dupont et al., 2021). In this paper, we propose crucial steps towards making such ideas scalable: Firstly, we employ state-of-the-art network sparsification techniques to drastically improve compression. Secondly, introduce the first method allowing for sparsification to be employed in the inner-loop of commonly used Meta-Learning algorithms, drastically improving both compression and the computational cost of learning INRs. The generality of this formalism allows us to present results on diverse data modalities such as images, manifolds, signed distance functions, 3D shapes and scenes, several of which establish new state-of-the-art results.

background contribution method result interpretation main claim

What the classifier reports for each sentence. The number is the probability it gave the label.
#	sentence	role	the authors' own work	states the main conclusion	strength
s0	Recent work in Deep Learning has re-imagined the representation of data as functions mapping from a coordinate space to an underlying continuous signal.	background 1.00	0.04	0.02	asserted 1.00
s1	When such functions are approximated by neural networks this introduces a compelling alternative to the more common multi-dimensional array representation.	background 1.00	0.06	0.02	asserted 0.93
s2	Recent work on such Implicit Neural Representations (INRs) has shown that - following careful architecture search - INRs can outperform established compression methods such as JPEG (e.g. Dupont et al., 2021).	background 1.00	0.03	0.05	asserted 0.99
s3	In this paper, we propose crucial steps towards making such ideas scalable: Firstly, we employ state-of-the-art network sparsification techniques to drastically improve compression.	contribution 0.82	0.84	0.13	asserted 0.97
s4	Secondly, introduce the first method allowing for sparsification to be employed in the inner-loop of commonly used Meta-Learning algorithms, drastically improving both compression and the computational cost of learning INRs.	contribution 0.94	0.94	0.36	asserted 1.00
s5	The generality of this formalism allows us to present results on diverse data modalities such as images, manifolds, signed distance functions, 3D shapes and scenes, several of which establish new state-of-the-art results.	result 0.95	0.92	0.34	asserted 1.00

What kind of paper the classifier reports this is: method or tool 1.00 · evidence from observation or experiment 0.63.

What was examined, and what was not

sources read

arXiv listing abstract v2

not examined

full text
figures and tables
references and citing sentences
relations to other papers

levels run

P abstract

not run

S sections
T body sentences
C citations
R cross-paper relations

Every quoted sentence on this record is the authors'. Every label is from Pith's closed vocabulary, with the probability shown beside it. Every number in the measurements table is the authors', checked by code. Record as JSON

The paper

Original abstract

Recent work in Deep Learning has re-imagined the representation of data as functions mapping from a coordinate space to an underlying continuous signal. When such functions are approximated by neural networks this introduces a compelling alternative to the more common multi-dimensional array representation. Recent work on such Implicit Neural Representations (INRs) has shown that - following careful architecture search - INRs can outperform established compression methods such as JPEG (e.g. Dupont et al., 2021). In this paper, we propose crucial steps towards making such ideas scalable: Firstly, we employ state-of-the-art network sparsification techniques to drastically improve compression. Secondly, introduce the first method allowing for sparsification to be employed in the inner-loop of commonly used Meta-Learning algorithms, drastically improving both compression and the computational cost of learning INRs. The generality of this formalism allows us to present results on diverse data modalities such as images, manifolds, signed distance functions, 3D shapes and scenes, several of which establish new state-of-the-art results.

Human review

How Pith checks papers → · Full audit of this paper →

How the literature uses it

6 arXiv papers cite this in 9 sentences

Each citing sentence is taken from the citing paper's LaTeX source and typed by Pith's classifier. A label appears at probability 0.8 or more (0.95 for support), a contest at 0.7 or more; the quoted words are the citing authors'. How each label was checked · JSON

Papers citing it · 4 sentences without a confident label

Uses its method or tool 1 paper

To learn discrete binary masks through gradient optimization, we take inspiration from [ref] , and reparameterize the $m$-th binary mask using an underlying continuous distribution $q_ϕ_m$ governed by parameters $ϕ_m$.

Unleashing the Power of Meta-tuning for Few-shot Generalization Through Sparse Interpolated Experts · Meta-learning sparse modulation with controlled $L_0$ sparsity · 0.98

Cites it as support 1 paper

Compared to a fixed structure, FNS that is adjusted depending on the task may provide the following benefits: (i) alleviate overfitting caused by over-parameterization [ref] ; (ii) reduce the computational overhead caused by redundant neurons [ref] .

Neuromodulated Meta-Learning · Introduction · 0.98

Cites it as background 3 papers

Recent works [ref] focus on improving COIN via a meta-learning approach in which a base network is (pre-)trained over a large collection of datapoints, so that instance-dependent INRs can be found quickly.

L$_0$onie: Compressing COINs with L$_0$-constraints · Introduction · 0.93
COIN++ [ref] and MSCN [ref] applied meta-learning methods to compress time slices of climate data into latent vectors of several bits.

Compressing multidimensional weather and climate data into neural networks · Related work · 0.97
INRs have also been explored for compact signal representation and compression, often with meta-learning or sparsification to reduce per-sample fitting cost [ref] .

AIR: Amortized Image Reconstruction Framework for Self-Supervised Feed-Forward 2D Gaussian Splatting · Implicit Neural Image Representation · 0.92

What it cites · 20 sentences without a confident label

Uses the method or tool of 4 papers

Secondly, in order to tackle the computational cost of learning INRs, we follow recent work [ref] by adopting Meta-Learning techniques such as MAML [ref] which allow learning an INR representing a single signal by finetuning from a learned initialisation using only a handful of optimisation steps.

From data to functa: Your data point is a function and you can treat it like one · Introduction · 0.97
We thus employ recent state-of-the-art sparsity techniques [ref] to explicitly optimise INRs using as few parameters as possible, drastically improving their compression cost.

Learning Sparse Neural Networks through $L_0$ Regularization · Introduction · 1.00
In this setting, we found the use of aforementioned MetaSGD [ref] to be particularly effective.

Meta-SGD: Learning to Learn Quickly for Few-Shot Learning · Unstructured Sparse Gradients · 0.91

Compares with 1 paper

The closest comparison is Functa [ref] , which we take to be our baseline method.

From data to functa: Your data point is a function and you can treat it like one · Structured sparse modulations · 0.99

Cites as support 2 papers

First, we recognise insights of recent deep learning studies [ref] which show how only a small subset of parameters encode the predictive function.

The Lottery Ticket Hypothesis: Finding Sparse, Trainable Neural Networks · Introduction · 0.96
Indeed [ref] write “reconstructions are more sensitive to earlier modulation layers than later ones.

From data to functa: Your data point is a function and you can treat it like one · Structured sparse modulations · 0.99

Cites as background 6 papers

The compression perspective of INRs has recently been explored in the aforementioned COIN [ref] and in [ref] who focus on 3D shapes.

On the Effectiveness of Weight-Encoded Neural Implicit 3D Shapes · Implicit Neural Representations · 0.94
The compression perspective of INRs has recently been explored in the aforementioned COIN [ref] and in [ref] who focus on 3D shapes.

COIN: COmpression with Implicit Neural representations · Implicit Neural Representations · 0.94
Current techniques can be mainly categorised as the iterative sparsification of a densely initialised network [ref] or techniques that maintain constant sparsity throughout learning [ref] .

The State of Sparsity in Deep Neural Networks · Neural network sparsity · 0.97

Discussion

Know something about this paper? Leave a comment, or highlight any passage above to comment on it.