Deep Image Clustering with Model-Agnostic Meta-Learning
Kim Bjerge
1 a
, Paul Bodesheim
2 b
and Henrik Karstoft
1 c
1
Department of Electrical and Computer Engineering, Aarhus University, Finlandsgade 22, 8200 Aarhus N, Denmark
2
Computer Vision Group, Friedrich Schiller University, Ernst-Abbe-Platz 2, 07743 Jena, Germany
Keywords:
Deep Clustering, Episodic Training, Few-Shot Learning, Multivariate Loss, Semi-Supervised Learning.
Abstract:
Deep clustering has proven successful in analyzing complex, high-dimensional real-world data. Typically,
features are extracted from a deep neural network and then clustered. However, training the network to extract
features that can be clustered efficiently in a semantically meaningful way is particularly challenging when
data is sparse. In this paper, we present a semi-supervised method to fine-tune a deep learning network using
Model-Agnostic Meta-Learning, commonly employed in Few-Shot Learning. We apply episodic training
with a novel multivariate scatter loss, designed to enhance inter-class feature separation while minimizing
intra-class variance, thereby improving overall clustering performance. Our approach works with state-of-the-
art deep learning models, spanning convolutional neural networks and vision transformers, as well as different
clustering algorithms like K-means and Spectral clustering. The effectiveness of our method is tested on
several commonly used Few-Shot Learning datasets, where episodic fine-tuning with our multivariate scatter
loss and a ConvNeXt backbone outperforms other models, achieving adjusted rand index scores of 89.7% on
the EU moths dataset and 86.9% on the Caltech birds dataset, respectively. Hence, our proposed method can
be applied across various practical domains, such as clustering images of animal species in biology.
1 INTRODUCTION
In real-world applications such as biology and
ecology, training supervised networks to classify
rare species from images presents significant chal-
lenges (Mora et al., 2011; Binta Islam et al., 2023). In
many cases, images of these species are either scarce
or non-existing, making unsupervised methods such
as deep image clustering a valuable alternative.
Clustering is a core problem in unsuper-
vised learning, with traditional methods like K-
means (Macqueen, 1967) and Spectral cluster-
ing (Von Luxburg, 2007) used to group data into clus-
ters where similar data points are close together and
dissimilar points are far apart. However, these clas-
sical methods often operate in the feature space of
hand-crafted features, which may not capture the un-
derlying structure of the data, limiting their effec-
tiveness. To overcome this, deep clustering tech-
niques have emerged (Bo et al., 2020; Sun et al.,
2022; Huang et al., 2024; Lu et al., 2024), leverag-
ing deep neural networks to learn an embedding that
a
https://orcid.org/0000-0001-6742-9504
b
https://orcid.org/0000-0002-3564-6528
c
https://orcid.org/0000-0003-3739-8983
better reflects the intrinsic data structure before ap-
plying clustering. For instance, deep embedded clus-
tering uses auto-encoders and Kullback-Leibler (KL)
divergence minimization to improve clustering accu-
racy (Xie et al., 2016), while other methods use Con-
volutional Neural Networks (CNN) to extract latent
representations (Yang et al., 2016). Since deep clus-
tering is an unsupervised method it still may be misled
by noisy or complex data in the absence of labels.
A natural solution to this limitation is to incorpo-
rate some supervised information, leading to semi-
supervised deep clustering (Ren et al., 2019; Qin
et al., 2019; Cai et al., 2023). In real-world appli-
cations, individual labels are often hard to obtain, but
pairwise relations between data points are more ac-
cessible. For example, in face recognition, while the
identity (label) of a person may be unknown, it is of-
ten easy to determine whether two images represent
the same individual (Chopra et al., 2005). The learn-
ing process minimizes a contrastive loss function that
drives the similarity metric to be small for pairs of
faces from the same person, and large for pairs from
different persons. Semi-supervised methods use pair-
wise constraints — whether two data points belong to
the same class — to guide clustering. For example,
286
Bjerge, K., Bodesheim, P. and Karstoft, H.
Deep Image Clustering with Model-Agnostic Meta-Learning.
DOI: 10.5220/0013114600003912
Paper published under CC license (CC BY-NC-ND 4.0)
In Proceedings of the 20th International Joint Conference on Computer Vision, Imaging and Computer Graphics Theory and Applications (VISIGRAPP 2025) - Volume 2: VISAPP, pages
286-297
ISBN: 978-989-758-728-3; ISSN: 2184-4321
Proceedings Copyright © 2025 by SCITEPRESS – Science and Technology Publications, Lda.

Vilhagra et al. proposed a Siamese network (Vilhagra
et al., 2020), which has been applied with Spectral
clustering for single-cell RNA sequencing data (Jiang
et al., 2022). While this approach can enhance per-
formance, its success largely depends on the quality
of the pairwise constraints, and selecting them appro-
priately is a difficult challenge.
In this paper, we propose using Model-Agnostic
Meta-Learning (MAML) (Finn et al., 2017), a
technique commonly applied in Few-Shot Learning
(FSL) (Wang et al., 2020; Song et al., 2023), to ad-
dress the challenge of clustering with limited labeled
data using a set of data points. FSL, which involves
training models using only a few labeled examples,
employs a support and query set for episodic train-
ing (Li et al., 2019), and has gained substantial at-
tention for its ability to enable models to learn from
sparse data (Wang et al., 2020). FSL bridges the
gap between traditional deep learning, which requires
large labeled datasets, and the growing need for sys-
tems that can quickly adapt to new tasks or domains.
In our approach, we utilize a small amount of
labeled data, such as similar images from a limited
set of categories, to improve clustering performance.
This method is particularly relevant in domains like
biology, where clustering rare or unseen species is
crucial (Bjerge et al., 2023). By merging FSL with
clustering, we aim to develop models that can gener-
alize beyond their training data, addressing the critical
need for adaptability in real-world applications.
In domains where labeled data is available for
some classes but absent for others, our proposed
method offers a viable solution by clustering data, un-
like supervised classifiers, which struggle to handle
unknown data samples.
Contribution. Our paper introduces a novel
methodology specifically tailored for clustering
samples within the framework of few-shot learning,
utilizing episodic training to achieve domain gen-
eralization (Li et al., 2019; Ren et al., 2018). The
approach is designed to cluster samples from novel
classes not included in the FSL support set.
Our proposal uses transductive inference with
the Prototypical Network during meta-learning with
episodic training of models to tackle the challenges
posed by limited labeled data.
To further improve clustering performance, we
propose a novel multivariate scatter loss func-
tion, extending the univariate scatter loss introduced
by (Bjerge et al., 2024). This function is strategi-
cally applied during episodic training to disentangle
and separate the distribution of classes within the sim-
ilarity space. The paper evaluates four state-of-the-art
deep learning networks fine-tuned with episodic train-
ing and the multivariate scatter loss evaluated with the
K-means and Spectral clustering algorithms.
2 RELATED WORK
2.1 Deep Image Clustering
Deep clustering has emerged as a powerful technique
for analyzing complex, high-dimensional data. A
variety of methods have been proposed to enhance
clustering performance, especially in cases where la-
beled data is scarce or non-existent. Deep Embed-
ded Clustering (DEC), introduced by Xie et al. (Xie
et al., 2016), is one of the state-of-the-art approaches
in deep clustering. However, DEC does not leverage
prior knowledge to guide the learning process. To
address this limitation, Ren et al. (Ren et al., 2019)
propose Semi-Supervised Deep Embedded Clustering
(SDEC), which incorporates labeled data to improve
clustering outcomes. This highlights the importance
of integrating supervision in deep clustering for im-
proved performance. Bo et al. (Bo et al., 2020) further
advanced the field by introducing the Structural Deep
Clustering Network (SDCN), which integrates struc-
tural information into deep clustering. They point out
that most deep clustering methods rely on the neural
network ability to learn effective feature representa-
tions, often through auto-encoders.
While deep clustering methods have shown great
promise, Sun et al. (Sun et al., 2022) note that the ab-
sence of labels often results in unreliable clustering.
To mitigate this, they propose Deep Active Cluster-
ing, which actively selects key data points for human
labeling. This novel approach improves the clustering
process by intelligently choosing the most informa-
tive samples for annotation, addressing one of the lim-
itations of conventional semi-supervised deep cluster-
ing methods that rely on fixed, pre-labeled data.
The evolution of deep clustering methods has also
been tracked by Lu et al. (Lu et al., 2024), who of-
fer a comprehensive review of prior knowledge used
in deep clustering. Huang et al. (Huang et al., 2024)
provide a general framework for Deep Image Cluster-
ing Networks, outlining the key stages: image pre-
processing, embedding, feature processing, cluster-
ing, and result processing. The challenges posed by
big data clustering are also highlighted by Fahad et
al. (Fahad et al., 2014). They offer a survey of ex-
isting clustering algorithms and a comparison of their
theoretical and empirical performance.
Rodriguez et al. (Rodriguez et al., 2019) per-
formed a systematic comparison of nine well-known
Deep Image Clustering with Model-Agnostic Meta-Learning
287

clustering methods, offering insights into their rel-
ative performance across different datasets. Addi-
tionally, various classic clustering algorithms such as
Density-Based Clustering (Ester et al., 1996; Ankerst
et al., 1999), EM Algorithms (Dempster et al.,
1977), Hierarchical Density-Based Clustering (HDB-
SCAN) (McInnes et al., 2017) and Spectral Clus-
tering (Von Luxburg, 2007) have been explored in
the context of clustering of large-scale data. Jain et
al. (Jain, 2010) summarize these methods, discussing
key challenges like feature selection, semi-supervised
clustering, and clustering at scale.
2.2 Few-Shot Learning
Few-Shot Learning comprising diverse approaches
tailored to address the challenges inherent in learning
from limited labeled data. Few-shot learning can be
categorized into two distinct branches: inductive FSL
and transductive FSL. The former involves the pre-
diction of test samples individually, while the latter
addresses the prediction of test samples collectively.
For a comprehensive overview of the evolving
FSL landscape, we refer to surveys such as those pre-
sented in works by Wang et al. (Wang et al., 2020)
and Song et al. (Song et al., 2023). A notable cat-
egory of methods relevant to our proposed approach
leverage euclidean distance and cosine similarity as
the fundamental measure. This includes Prototyp-
ical Networks (Snell et al., 2017), Finetune (Chen
et al., 2019), Transductive Information Maximiza-
tion (Boudiaf et al., 2020), and Prototypical Rectifi-
cation (Liu et al., 2020).
Meta-learning is a key paradigm in the FSL land-
scape, and a comprehensive survey on the subject
is presented by Hospedales et al. (Hospedales et al.,
2022). Additionally, the integration of episodic train-
ing for domain generalization, as discussed by Li
et al. (Li et al., 2019) and Model-Agnostic Meta-
Learning (Finn et al., 2017), emerges as a crucial as-
pect in enhancing the adaptability and robustness of
FSL models.
In few-shot classification, we are given a small
support set of (N · K) labeled examples S =
{(X
(1)
1
, y
1
), ..., (X
(K)
N
, y
K
)} where each X
(k)
n
∈ R
D
is
the D-dimensional embedded feature vector of an
example and y
k
∈ {1, ..., K} are the correspond-
ing labels (Bjerge et al., 2024). The set S
k
=
{(X
(k)
1
, y
k
), ..., (X
(k)
N
, y
k
)} denotes the subset of exam-
ples labeled with class k and the number of class la-
bels in the support set is denoted K-way where we
have N-shots of examples in each S
k
. A query set Q
contains sample images q
i
that belong to classes in
the support set where the goal is to match the query
samples to the correct class label. When training with
a dataset that comprises more classes than those in-
cluded in the support set, a random subset of K classes
is selected for each few-shot task, which encompasses
both the support and query sets. We propose to use
the Prototypical Network (Snell et al., 2017), which
employs the Euclidean distance to compare the center
point of the support set with the query samples.
In conclusion, FSL and deep clustering has made
significant strides, leveraging powerful feature ex-
traction methods from deep learning and incorporat-
ing structural and active learning techniques to en-
hance clustering performance. Despite these advance-
ments, challenges remain, particularly in large-scale
data clustering and the integration of prior knowledge,
which continue to drive innovation in the field.
3 METHOD
3.1 Deep Clustering with Few-Shot
Learning
We propose clustering images by utilizing feature em-
beddings derived from the output vector of a deep
learning (DL) model. The DL model is fine-tuned
through episodic training, a technique commonly
used in FSL (Li et al., 2019). We introduce a multi-
variate loss function to enhance feature clustering dur-
ing episodic training. The pipeline, illustrated in Fig-
ure 1, consists of three steps.
Step 1. involves training and fine-tuning a pre-
trained DL model to extract features that effectively
represent the images, ensuring that the embedding
space is spread across diverse clusters, as visualized
in Figure 2. To enhance the quality of these feature
embeddings, we leverage MAML (Finn et al., 2017;
Hospedales et al., 2022), a technique commonly used
in FSL. MAML uses a training and validation dataset
with classes of images that are not present in the test
dataset, allowing the model to generalize well to new,
unseen classes.
In our study, we selected four DL models that span
both convolutional neural networks (CNNs) and vi-
sion transformers (ViTs). We chose ResNet50v2 (He
et al., 2016) and EfficientNetB3 (Tan and Le, 2019)
as they are widely used CNNs, achieving top-1
accuracies of 76.0% and 81.6% on the ImageNet
dataset (Russakovsky et al., 2015), respectively. Fur-
thermore, we included ConvNext-B (Todi et al.,
2023), a recently published model that achieves an
ImageNet top-1 accuracy of 85.3%, and the vision
VISAPP 2025 - 20th International Conference on Computer Vision Theory and Applications
288

Figure 1: Illustrates the pipeline of our proposed deep clustering method for effectively grouping images in few-shot scenarios.
A training and validation dataset with different classes is used to fine-tune a deep learning model for feature extraction. In
step 1. the model is fine-tuned with few-shot learning and episodic training where the best model is chosen based on K-means
clustering of the validation dataset. In Step 2 and 3, the extracted features are fed into a clustering algorithm, which groups a
set of new test images into N clusters.
(a) Pretrained DL model
(b) Fine-tuned DL model
Figure 2: T-SNE projection of feature embeddings for clus-
ters with a pretrained DL model and a fine-tuned DL model
with episodic training and multivariate scatter loss. The
scatter plot illustrates the clustering of the first and second
components of 11 classes in the embedding space.
transformer ViT-B/16 (Dosovitskiy et al., 2021) with
an accuracy of 88.6%.
Step 2. concerns extracting features from the fine-
tuned DL model. The feature vector is extracted from
the backbone of the trained DL model. The backbone
architecture includes multiple convolutional layers or
vision transformer layers, with the final head of fully
connected layers removed. The test dataset contains
image categories that are not present in the training or
validation datasets. Additionally, the categories in the
training and validation datasets are distinct from each
other as well. The size of the feature vector depends
on the used DL model.
Step 3. involves clustering of feature embeddings
extracted with the DL model. Clustering analysis in-
volves grouping a set of objects such that those within
the same cluster are more similar to each other than
to objects in other clusters. In our study, the objects
being clustered are unlabeled images, which could in-
clude images of various unlabeled categories. In eco-
logical studies, this could involve species such as an-
imals, birds, or insects for which labels are sparse or
entirely absent. The primary challenge is to learn ef-
fective feature embeddings of training images that can
serve as input to a clustering algorithm.
For this purpose, we utilize K-means (Macqueen,
1967) and Spectral clustering (Von Luxburg, 2007),
assuming the number of clusters is known. These two
clustering methods were selected based on a compar-
ative analysis of nine common clustering methods by
Rodriguez et al. (Rodriguez et al., 2019). K-means is
a well-known and widely used method, while Spec-
tral clustering demonstrated superior performance,
achieving an adjusted rand index of 68.16% as the
best out of nine different selected clustering algo-
rithms.
Unlike K-means, which works well for convex
clusters, spectral clustering can capture more com-
plex, non-linear relationships in the data. Spectral
clustering is a graph-based approach and leverages the
spectral (eigenvalue) properties of a similarity graph
constructed from the data to identify clusters, rather
than relying on traditional distance-based metrics like
K-means. It effectively detects clusters of arbitrary
Deep Image Clustering with Model-Agnostic Meta-Learning
289

shapes and sizes, particularly in cases where data
points are not well-separated by traditional distance
metrics.
While K-means and Spectral clustering are cen-
tral to our demonstration, other clustering meth-
ods could also be employed. These include Fast
Density-Based Clustering (DBSCAN) (Ester et al.,
1996), Expectation-Maximization (EM) (Dempster
et al., 1977), or hierarchical methods such as ag-
glomerative and divisive clustering, as explored by
Jain et al. (Jain, 2010). For unsupervised cluster-
ing where the number of clusters is unknown, meth-
ods like Ordering Points To Identify the Clustering
Structure (OPTICS) (Ankerst et al., 1999) and HDB-
SCAN (McInnes et al., 2017) could be utilized.
A solution to clustering of images not included
in training would benefit many real-life applications
where the number of samples for each class is sparse
or labelled data was non-existent during training.
3.2 Episodic Training with Multivariate
Scatter Loss
In few-shot learning, episodic training with domain
generalization (DG) (Li et al., 2019), also known
as meta-learning (Hospedales et al., 2022), involves
training with a set of tasks during each epoch. Each
task comprises several episodes, with each episode
containing a labeled support set and a query set drawn
from the training dataset. After each epoch, the model
accuracy is evaluated using the validation dataset,
which contains tasks with class categories different
from those in the training dataset, thereby assessing
the model ability to generalize across domains.
A Prototypical Network uses Euclidean distance
as a similarity function to predict the relationship
between query samples and class labels in the sup-
port set. Experiments have demonstrated that training
with Euclidean distance, rather than cosine similar-
ity, yields superior results (Li et al., 2019). In our
approach, we use the cross entropy loss together a
new a multivariate scatter loss function during train-
ing. This new loss function is designed to minimize
within-class variance while simultaneously maximiz-
ing the mean separation between classes, enhancing
the model’s discriminative ability. The method is in-
spired by the work of Bodesheim et al. (Bodesheim
et al., 2013) and Bjerge et al. (Bjerge et al., 2024).
For an illustration of class distribution after training
with multivariate scatter loss, see Figure 2. The mul-
tivariant scatter loss is defined as
Lm(θ) =
∑
K
k=1
∑
N
k
n=1
(X
n
− X
k
)
T
(X
n
− X
k
)
∑
K
i=1
∑
K
j=i+1
(X
i
− X
j
)
T
(X
i
− X
j
)
(1)
with X
n
being the embedding vector for each sample
in the support set S
k
of class k. N
k
is the number of
samples in each class of the support set and θ is the
parameters of the DL model. With X
k
, X
i
, and X
j
we
denote the mean center point of samples in the classes
k, i, and j of the support set. The numerator of Eq. (1)
encourages the norm of the data points, centred with
respect to their cluster, to be small, and the denomi-
nator encourages the cluster centres to be far apart.
The cross-entropy loss function in Eq. (2) ensures
that a query sample q is classified correctly according
to the support set during training:
Lc
q
(θ) = − ˆy
j
log(
exp(−d
q,k
j
)
∑
K
i=1
exp(−d
q,k
i
)
) (2)
Here, d
q,k
j
is the euclidean distance between the sup-
port center point k
j
and the query sample q and ˆy
j
is
the one-hot encoded vector for the correct label of the
query sample. K is the number of classes (K-way)
in the support set. Finally a combined loss function
is defined in Eq. (3) to prioritize between the cross-
entropy and scatter loss:
L(θ) = αLm(θ) + (1 − α)Lc(θ) . (3)
The goal is to increase the distance between class dis-
tributions in the support set while increasing the num-
ber of correctly classified query samples related to the
support classes. The loss function L(θ) will prioritize
between minimizing the multivariate scatter loss (Lm)
and the cross-entropy loss for the batch of query sam-
ples (Lc) by adjusting α ∈ [0, 1]
3.3 Fine-Tuning with Episodic Training
A DL model was used as a backbone to extract fea-
ture embeddings. The outputs from the backbone
of the DL model were flattened and used as embed-
dings, leading to D-dimensional feature vectors. In
our work, we have trained larger models than used
in FSL. The models were fine-tuned using four DL
architectures, generating feature vectors with the fol-
lowing dimensions: 2048 (ResNet50v2), 1536 (Ef-
ficientNetB3), 1024 (ConvNeXt-B), and 768 (ViT-
B/16). For fine-tuning, we used classical pre-trained
weights from the ImageNet dataset (Russakovsky
et al., 2015). These pre-trained DL models were then
fine-tuned with episodic training on a new domain-
specific dataset. During all episodic training sessions,
a 5-shot K-way support set was utilized. The value
of K ranged between 15 and 30 constrained by the
DL model and a maximum of 50GB of GPU mem-
ory. Training on the datasets was performed using
data augmentation, including image scaling, horizon-
tal flip and adding color jitter for brightness, contrast
and saturation.
VISAPP 2025 - 20th International Conference on Computer Vision Theory and Applications
290

The stochastic gradient descent optimizer (SGD)
was used during episodic training and fine-tuning.
The SGD was configured with the momentum of 0.9
and a weight decay of 5.0 · 10
−4
using a multi-step
scheduler to lower the learning rate at two specified
milestones specified by epochs. The first milestone
was set to 3 epochs and the second to 6 epochs with
a total of 9 epochs. The limited number of epochs
is a result of episodic training, with each epoch com-
prising 300 few-shot tasks, each consisting of 6 query
samples per class from the support set. SGD was
tested with the initial learning rate of 1.0 · 10
−3
for
pre-trained models during fine-tuning.
K-means clustering was performed on all images
in the validation dataset after each epoch to select the
best performing model. The clustering was applied to
the validation datasets described in section 4.1 which
included 20, 40, 50, or 97 classes. After each epoch,
cluster accuracy (CA) was used to select the optimal
model for clustering the validation dataset with K-
means.
3.4 Performance Metrics for Clustering
In this section, we briefly outline the criteria used
for performance evaluation based on commonly used
clustering validation indices (Wu et al., 2019; Huang
et al., 2024; Lu et al., 2024). These criteria are anal-
ogous to accuracy or recognition rate in supervised
learning. Four key evaluation metrics are used in-
clude Cluster Accuracy (CA), Adjusted Rand Index
(ARI) (Hubert and Arabie, 1985), Normalized Mutual
Information (NMI) (Strehl and Ghosh, 2003) and Ad-
justed Mutual Information (AMI) (Vinh et al., 2009).
When the true class labels of a dataset are known,
these metrics allow us to assess how accurately a clus-
tering technique partitions the data relative to the cor-
rect class labels.
CA. measures the percentage of correctly classified
images in the clustering solution compared to pre-
defined image class labels (Ω). We use CA as defined
by (Fahad et al., 2014):
CA =
K
∑
i=1
max(C
i
|L
i
)
|Ω|
(4)
where C
i
is the set of instances in the ith cluster. L
i
is the class labels for all images in the ith cluster, and
max(C
i
|L
i
) is the number of instances with the major-
ity label in the ith cluster (e.g. if label l appears in
the ith cluster more often than any other label, then
max(C
i
|L
i
) is the number of instances in C
i
with the
label l).
ARI. is a measure used to evaluate the similarity
between two clustering results, taking into account
the possibility of chance. It is an improvement over
the basic Rand Index, which measures the proportion
of agreement between two clusterings (i.e., the pro-
portion of points that are either clustered together or
clustered apart in both clusterings). ARI takes into
account the number of instances that exist in the same
cluster and different clusters.
NMI. is a normalized version of Mutual Informa-
tion (MI) using the upper bound to score results be-
tween 0 (no mutual information) and 1 (perfect corre-
lation).
AMI. is an adjustment of the MI score to account
for chance. It accounts for the fact that the Mutual In-
formation is generally higher for two clusterings with
a larger number of clusters, regardless of whether
there is actually more information shared.
4 EXPERIMENTAL SETUP
The proposed method undergoes training and evalu-
ation using four distinct datasets for FSL that are de-
scribed in Sec. 4.1. Classic training and fine-tuning of
models has also been conducted to compare perfor-
mances with episodic training. The best model was
selected based on cluster accuracy using K-means,
evaluated after each epoch for both training methods.
The metrics outlined in the tables of Sec. 5 include
CA, NMI, AMI, and ARI for clustering of features
with classic and episodic training. Metrics in all ta-
bles are computed across 5 random runs of clustering
feature vectors from fine-tuned DL models, with av-
erage (AVG) and standard deviations (SD).
4.1 Datasets
We have selected four datasets with images typically
used for FSL, here we use the validation and test-
ing dataset for clustering. These datasets were cho-
sen over commonly used clustering datasets due to
their relatively high resolution and the clear separa-
tion into distinct classes for training, validation, and
testing, as demonstrated in previous studies (Ravi
and Larochelle, 2017; Wang et al., 2019; Bjerge
et al., 2024). Unlike FSL datasets, standard clustering
datasets typically do not provide such distinct class
separations.
Mini-ImageNet is a benchmark dataset and is a
subset of the larger ILSVRC-12 dataset (Russakovsky
Deep Image Clustering with Model-Agnostic Meta-Learning
291

Table 1: Shows the metrics (CA, NMI, AMI, ARI) for K-means and spectral clustering of features from fine-tuned models
with classic and episodic training with ConvNeXt-B on the four datasets, tested with clustering on test datasets. A best α
value above zero indicates that the multivariate scatter loss improved episodic training in finding the best model. The best
metric results for each clustering method and dataset is highlighted with bold.
Dataset Cluster Training Best CA NMI AMI ARI
method α AVG (SD) AVG (SD) AVG (SD) AVG (SD)
EU Moths K-means Classic - 0.773 (0.022) 0.872 (0.009) 0.785 (0.014) 0.653 (0.026)
EU Moths K-means Episodic 0.1 0.837 (0.029) 0.912 (0.005) 0.851 (0.008) 0.748 (0.009)
EU Moths Spectral Classic - 0.874 (0.009) 0.922 (0.003) 0.868 (0.006) 0.787 (0.009)
EU Moths Spectral Episodic 0.1 0.940 (0.011) 0.962 (0.004) 0.935 (0.007) 0.897 (0.012)
CUB K-means Classic - 0.844 (0.019) 0.885 (0.010) 0.864 (0.011) 0.759 (0.026)
CUB K-means Episodic 0.1 0.834 (0.025) 0.885 (0.005) 0.864 (0.006) 0.745 (0.038)
CUB Spectral Classic - 0.936 (0.000) 0.932 (0.000) 0.919 (0.000) 0.881 (0.000)
CUB Spectral Episodic 0.0 0.928 (0.002) 0.930 (0.002) 0.917 (0.002) 0.869 (0.004)
mini-ImageNet K-means Classic - 0.895 (0.005) 0.942 (0.001) 0.941 (0.001) 0.853 (0.005)
mini-ImageNet
K-means Episodic 0.5 0.900 (0.010) 0.942 (0.001) 0.941 (0.001) 0.856 (0.009)
mini-ImageNet Spectral Classic - 0.903 (0.003) 0.941 (0.001) 0.941 (0.001) 0.832 (0.006)
mini-ImageNet Spectral Episodic - 0.893 (0.002) 0.937 (0.000) 0.936 (0.000) 0.813 (0.001)
tired-ImageNet K-means Classic - 0.975 (0.024) 0.984 (0.007) 0.984 (0.007) 0.965 (0.028)
tired-ImageNet K-means Episodic 0.5 0.984 (0.021) 0.986 (0.007) 0.986 (0.007) 0.977 (0.023)
tired-ImageNet Spectral Classic - 0.978 (0.018) 0.982 (0.007) 0.982 (0.007) 0.966 (0.023)
tired-ImageNet Spectral Episodic 0.2 0.982 (0.019) 0.984 (0.007) 0.984 (0.007) 0.971 (0.024)
et al., 2015). It has a total of 60,000 color images from
100 classes, where each class has 600 images of size
224x224 (Vinyals et al., 2016). In alignment with es-
tablished practices (Ravi and Larochelle, 2017; Wang
et al., 2019), we adopt a partitioning scheme consist-
ing of 60 training classes, complemented by 20 vali-
dation classes and 20 test classes.
Tiered-ImageNet (Ren et al., 2018) is a larger sub-
set of the larger ILSVRC-12 dataset (Russakovsky
et al., 2015). It contains 608 classes with 779,165
images partitioned into disjoint sets for training (351
classes), validation and testing. Images for evaluation
are split into 97 classes for validation and 160 classes
for testing.
Caltech Birds-200-2011 (CUB) (Wah et al., 2011)
is a fine-grained image classification dataset. We
adopt Chen et al. (Chen et al., 2019) for few-shot clas-
sification on CUB, which splits into 120 training, 40
validation, and 40 test classes with a total of 11,788
images for the experiments. To maintain consistency
and comparability with mini-ImageNet, the images
from CUB are uniformly resized to 224x224 pixels.
The EU Moths dataset, as introduced by B
¨
ohlke
et al. (B
¨
ohlke et al., 2021), encapsulates a collec-
tion of 200 moth species prevalent in Central Europe.
Notably, each species is delineated by few samples,
comprising a mere 11 images, resulting in a total of
2,205 images within the dataset. This dataset was
chosen for our evaluation of the multivariate scatter
loss because it presents the most challenging for clus-
tering, with only a few samples per class and species
with highly similar appearances. The images have
high resolution, even surpassing 1000x1000 pixels,
and have been resized to 224x224 pixels. The dataset
is split into 100 training, 50 validation, and 50 test
classes for experiments.
5 RESULTS AND DISCUSSION
Several experiments were performed on the datasets
to evaluate clustering and episodic training with the
multivariate scatter loss function. Here, results were
obtained with pre-trained models on ImageNet and
fine-tuning with classic and episodic training on
the four datasets. Since mini-ImageNet and tiered-
ImageNet are a subset of ImageNet we only expect to
see minor improvements when fine-tuning the models
with classic or episodic training.
Detailed results as shown in Table 1 for
training models with ResNet50v2, EfficientNetB3,
ConvNeXt-B and ViT-B/16 can be found on Github
1
with the source code for the experimental results.
5.1 Evaluation of the Multivariate
Scatter Loss
Spectral clustering of features from fine-tuned mod-
els, using both classic and episodic training on the EU
moth dataset, is shown in Figure 3. The scores repre-
sent the average of five runs. It is important to em-
phasize that the optimal model, along with the corre-
1
https://github.com/kimbjerge/few-shot-clustering
VISAPP 2025 - 20th International Conference on Computer Vision Theory and Applications
292

(a) ResNet50V2 feature clustering. (b) EfficientNetB3 feature clustering.
(c) ConvNeXt-B feature clustering. (d) ViT-B/16 feature clustering.
Figure 3: Shows the ARI and AMI metrics with Spectral clustering of features from fine-tuned models on the EU moth dataset
with different α values. Each green (AMI) and blue (ARI) dot represents a clustering result out of five random runs for each
metric. The red (AMI) and yellow (ARI) circles at Alpha=0 are the metrics for the classic trained models.
sponding α values, was determined through episodic
fine-tuning using the validation sets.
All four models achieved higher AMI and ARI
scores with episodic training compared to fine-
tuning with classic training. The highest relative in-
crease in performance with episodic training is ob-
served for ResNet50v2 and EfficientNetB3. However,
ConvNeXt-B and ViT-B/16 achieved the highest AMI
and ARI scores overall. As the value of α increases,
both AMI and ARI scores tend to decrease. Nonethe-
less, the highest ARI scores were observed with non-
zero α values at: α = 0.5 (ResNet50v2), α = 0.1 (Ef-
ficientNetB3), α = 0.1 (ConvNeXt-B) and α = 0.2
(ViT-B/16). This indicates that the multivariate scat-
ter loss positively impacts episodic training, leading
to improved Spectral clustering performance. How-
ever, there is only very little improvement compared
to α = 0 for all networks which limits the impact of
the scatter loss.
In Table 1 and on Github
1
, we present the
CA, NMI, AMI, and ARI metrics for both clas-
sic and episodic fine-tuning of models, with cluster-
ing performed using K-means and Spectral cluster-
ing. For K-means clustering with α = 0.1, the best
results were achieved with episodic fine-tuning of
the ResNet50v2, ConvNeXt-B, and ViT-B/16 models.
However, for EfficientNetB3, the multivariate scatter
loss did not result in performance improvement for
K-means clustering.
Fine-tuning models on the mini-ImageNet dataset
demonstrates that the multivariate scatter loss indi-
cates minor improved model performance, with the
episodic fine-tuned models achieving the best aver-
age ARI of 0.977 and AMI of 0.984 with α = 0.5
and ConvNeXt-B. In two cases, using ResNet50v2
and ViT-B/16, the highest scores were obtained with
Spectral clustering, solely by applying multivariate
scatter loss (α = 1.0).
5.2 Clustering with the EU Moths and
CUB Datasets
Figure 4 summarizes the ARI scores for Spectral clus-
tering of features from models pre-trained on Ima-
geNet and fine-tuned using either classic or episodic
training on the EU moths and CUB datasets.
As expected, the fine-tuned models outperform
the pre-trained models, as the fine-tuning process
adapts the models to datasets within the same domain,
in our case different species of moths or birds.
Notably, models fine-tuned with episodic train-
ing consistently outperform those trained with clas-
sic fine-tuning. This suggests that episodic training
enhances the models’ ability to generalize and adapt
to new domains, such as switching between differ-
ent species in the moth and bird datasets. For the
Deep Image Clustering with Model-Agnostic Meta-Learning
293

(a) EU moths with Spectral clustering. (b) CUB dataset with Spectral clustering.
Figure 4: Shows the ARI score with Spectral clustering of features from pre-trained and fine-tuned models on the EU moth
and CUB dataset. The blue bars presents clustering of features from classic pre-trained models on ImageNet. The green bars
presents clustering of features from classic fine-tuned models on the respective dataset. The red bars presents clustering of
features from the best episodic fine-tuned models.
(a) Mini-ImageNet - Spectral clustering. (b) TieredImageNet - K-means clustering.
Figure 5: Shows the ARI score with Spectral and K-means clustering of features from pre-trained and fine-tuned models
on the mini-ImageNet and tiered-ImagneNet dataset. The blue bars presents clustering of features from classic pre-trained
models on ImageNet. The green bars presents clustering of features from classic fine-tuned models on the respective dataset.
The red bars presents clustering of features from the best episodic fine-tuned models.
CUB dataset, the performance of ConvNeXt-B is the
same for both classic and episodic training. However,
this is not the case for the EU moths dataset, which
has fewer samples per class. The stronger perfor-
mance of episodic training on the EU moths dataset
highlights its effectiveness, particularly when work-
ing with datasets that have limited samples per class.
Table 1 presents the CA, NMI, AMI, and ARI
metrics for K-means and Spectral clustering of fea-
tures extracted from fine-tuned ConvNeXt-B models
using both classic and episodic training of the EU
moths, CUB, mini-ImageNet and tiered-ImageNet
datasets. Metrics in all tables are computed across
5 random runs of clustering feature vectors from fine-
VISAPP 2025 - 20th International Conference on Computer Vision Theory and Applications
294

tuned DL models, with average (AVG) and standard
deviations (SD). The best results were obtained us-
ing Spectral clustering on features from a ConvNeXt-
B model trained with episodic learning. For both
datasets, Spectral clustering outperforms K-means
clustering with an increase of 8% - 14% on all clus-
tering metrics.
On the EU moths dataset, the best model achieved
NMI=0.962, AMI=0.935, and ARI=0.897 with mul-
tivariate scatter loss (α = 0.1). For the CUB
dataset, ConvNeXt-B model reached NMI=0.930,
AMI=0.917, and ARI=0.869 with α = 0.0. How-
ever, classic fine-tuning yielded slightly better results,
with an ARI of 0.881. All scores are reported as the
average of 5 runs. The CUB dataset contains more
samples compared to the EU moths dataset, suggest-
ing that the multivariate scatter loss has a greater im-
pact with episodic training on datasets with fewer data
samples per class.
5.3 Clustering with the mini-ImageNet
and tiered-ImageNet Datasets
Figure 5 summarizes the ARI scores for Spectral and
K-means clustering of features extracted from models
pre-trained on ImageNet and fine-tuned using either
classic or episodic training on the mini-ImageNet and
tiered-ImageNet datasets. As expected, both classic
and episodic fine-tuning outperform the pre-trained
models. While episodic fine-tuning generally yields
slightly better results than classic fine-tuning, Effi-
cientNetB3 is an exception, where classic fine-tuning
is the best.
The multivariate scatter loss enhances episodic
training even without the use of cross-entropy loss
(α = 1.0). Given that the models were pre-trained
on ImageNet, the multivariate scatter loss appears to
improve the distribution of features into well-defined
clusters, particularly when the model has already been
pre-trained on the same classes.
The test dataset for tiered-ImageNet contains 160
classes, and Spectral clustering requires several hours
(10-20) to process. As a result, only the performance
of ConvNeXt-B is included in the detailed results on
Github
1
. Interestingly, K-means clustering outper-
forms Spectral clustering, achieving an ARI of 0.856
compared to 0.813 with Spectral clustering. This
suggests that while Spectral clustering excels with
test datasets containing 20-50 classes, K-means per-
forms better when the number of classes increases
above 150, it might be the preferred method for larger
datasets.
6 CONCLUSION
In this study, we present a novel method for clus-
tering images using Model-Agnostic Meta-Learning
within the context of few-shot learning and episodic
training with multivariate scatter loss. The proposed
method was evaluated on four commonly used few-
shot learning datasets, employing four state-of-the-art
DL models for feature extraction. ConvNeXt-B out-
performed the other networks, achieving ARI scores
of 0.897 and 0.869 on the EU moths and Caltech
Birds datasets, respectively. On the mini-ImageNet
and tiered-ImageNet datasets, episodic fine-tuned
models with multivariate scatter loss further improved
clustering performance, with ARI scores of 0.977
and 0.856 (α = 0.5). The multivariate scatter loss
consistently enhanced clustering performance during
episodic fine-tuning across most experiments, par-
ticularly on the EU moths dataset, where its effec-
tiveness was demonstrated with only 11 samples per
class. This highlights the method’s potential in han-
dling datasets with limited samples.
We explored two commonly used clustering algo-
rithms: K-means and Spectral clustering. On the EU
moths and Caltech Birds datasets, Spectral clustering
outperformed K-means, with an 8% to 14% improve-
ment across all clustering metrics. However, on the
tiered-ImageNet dataset, which contains a large num-
ber of classes (160), K-means clustering delivered the
best results. Future experiments could explore addi-
tional clustering algorithms, especially when applied
to datasets with few samples in each class.
Our solution represents an advancement in tack-
ling the challenge of image clustering, especially in
real-world scenarios where class samples are sparse.
The demonstrated accuracy and effectiveness of our
proposed method highlight its potential as a valuable
tool in settings with limited labeled data, offering
promising applications across a wide range of prac-
tical domains such as clustering of images of animal
species in biology.
ACKNOWLEDGEMENT
During the preparation of this work, the first author
utilized ChatGPT (OpenAI, 2023) to enhance the clar-
ity and formulation for parts of the written text. How-
ever, the authors takes full responsibility for the con-
tent of the publication.
Deep Image Clustering with Model-Agnostic Meta-Learning
295

REFERENCES
Ankerst, M., Breunig, M. M., Kriegel, H. P., and Sander,
J. (1999). OPTICS: Ordering Points to Identify the
Clustering Structure. SIGMOD Record (ACM Special
Interest Group on Management of Data), 28(2).
Binta Islam, S., Valles, D., Hibbitts, T. J., Ryberg, W. A.,
Walkup, D. K., and Forstner, M. R. (2023). Animal
Species Recognition with Deep Convolutional Neural
Networks from Ecological Camera Trap Images. Ani-
mals, 13(9).
Bjerge, K., Bodesheim, P., and Karstoft, H. (2024). Few-
shot learning with novelty detection. In Fred, A.,
Hadjali, A., Gusikhin, O., and Sansone, C., editors,
Deep Learning Theory and Applications, pages 340–
363, Cham. Springer Nature Switzerland.
Bjerge, K., Geissmann, Q., Alison, J., Mann, H. M., Høye,
T. T., Dyrmann, M., and Karstoft, H. (2023). Hier-
archical classification of insects with multitask learn-
ing and anomaly detection. Ecological Informatics,
77:102278.
Bo, D., Wang, X., Shi, C., Zhu, M., Lu, E., and Cui, P.
(2020). Structural Deep Clustering Network. In The
Web Conference 2020 - Proceedings of the World Wide
Web Conference, WWW 2020.
Bodesheim, P., Freytag, A., Rodner, E., Kemmler, M., and
Denzler, J. (2013). Kernel null space methods for nov-
elty detection. In Proceedings of the IEEE Computer
Society Conference on Computer Vision and Pattern
Recognition.
B
¨
ohlke, J., Korsch, D., Bodesheim, P., and Denzler, J.
(2021). Exploiting Web Images for Moth Species
Classification. In Lecture Notes in Informatics (LNI),
Proceedings - Series of the Gesellschaft fur Informatik
(GI), volume P-314.
Boudiaf, M., Masud, Z. I., Rony, J., Dolz, J., Piantanida,
P., and Ayed, I. B. (2020). Transductive informa-
tion maximization for few-shot learning. In Advances
in Neural Information Processing Systems, volume
2020-December.
Cai, J., Hao, J., Yang, H., Zhao, X., and Yang, Y. (2023).
A review on semi-supervised clustering. Information
Sciences, 632.
Chen, W. Y., Wang, Y. C. F., Liu, Y. C., Kira, Z., and Huang,
J. B. (2019). A closer look at few-shot classification.
In 7th International Conference on Learning Repre-
sentations, ICLR 2019.
Chopra, S., Hadsell, R., and LeCun, Y. (2005). Learn-
ing a similarity metric discriminatively, with appli-
cation to face verification. In Proceedings - 2005
IEEE Computer Society Conference on Computer Vi-
sion and Pattern Recognition, CVPR 2005, volume I.
Dempster, A. P., Laird, N. M., and Rubin, D. B. (1977).
Maximum Likelihood from Incomplete Data Via the
EM Algorithm . Journal of the Royal Statistical Soci-
ety Series B: Statistical Methodology, 39(1).
Dosovitskiy, A., Beyer, L., Kolesnikov, A., Weissenborn,
D., Zhai, X., Unterthiner, T., Dehghani, M., Minderer,
M., Heigold, G., Gelly, S., Uszkoreit, J., and Houlsby,
N. (2021). An image is worth 16X16 words: Trans-
formers for image recognition at scale. In ICLR 2021
- 9th International Conference on Learning Represen-
tations.
Ester, M., Kriegel, H. P., Sander, J., and Xu, X. (1996). A
Density-Based Algorithm for Discovering Clusters A
Density-Based Algorithm for Discovering Clusters in
Large Spatial Databases with Noise. In Proceedings -
2nd International Conference on Knowledge Discov-
ery and Data Mining, KDD 1996.
Fahad, A., Alshatri, N., Tari, Z., Alamri, A., Khalil, I.,
Zomaya, A. Y., Foufou, S., and Bouras, A. (2014).
A survey of clustering algorithms for big data: Tax-
onomy and empirical analysis. IEEE Transactions on
Emerging Topics in Computing, 2(3).
Finn, C., Abbeel, P., and Levine, S. (2017). Model-agnostic
meta-learning for fast adaptation of deep networks. In
34th International Conference on Machine Learning,
ICML 2017, volume 3.
He, K., Zhang, X., Ren, S., and Sun, J. (2016). Deep resid-
ual learning for image recognition. In Proceedings of
the IEEE Computer Society Conference on Computer
Vision and Pattern Recognition, volume 2016-Decem.
Hospedales, T., Antoniou, A., Micaelli, P., and Storkey, A.
(2022). Meta-Learning in Neural Networks: A Sur-
vey. IEEE Transactions on Pattern Analysis and Ma-
chine Intelligence, 44(9).
Huang, H., Wang, C., Wei, X., and Zhou, Y. (2024).
Deep image clustering: A survey. Neurocomputing,
599:128101.
Hubert, L. and Arabie, P. (1985). Comparing partitions.
Journal of Classification, 2(1).
Jain, A. K. (2010). Data clustering: 50 years beyond K-
means. Pattern Recognition Letters, 31(8).
Jiang, H., Huang, Y., and Li, Q. (2022). Spectral clus-
tering of single cells using Siamese nerual network
combined with improved affinity matrix. Briefings in
Bioinformatics, 23(3).
Li, D., Zhang, J., Yang, Y., Liu, C., Song, Y. Z., and
Hospedales, T. (2019). Episodic training for domain
generalization. In Proceedings of the IEEE Interna-
tional Conference on Computer Vision, volume 2019-
October.
Liu, J., Song, L., and Qin, Y. (2020). Prototype Rectifica-
tion for Few-Shot Learning. Lecture Notes in Com-
puter Science (including subseries Lecture Notes in
Artificial Intelligence and Lecture Notes in Bioinfor-
matics), 12346 LNCS:741–756.
Lu, Y., Li, H., Li, Y., Lin, Y., and Peng, X. (2024). A survey
on deep clustering: from the prior perspective. Vici-
nagearth, 1(1):1–17.
Macqueen, J. (1967). Some methods for classification and
analysis of multivarite observation. Preceeding of the
5th Berkeley symposium on mathematical statistics
and probability, Berkeley. University of california
press, 281.
McInnes, L., Healy, J., and Astels, S. (2017). hdbscan: Hi-
erarchical density based clustering. The Journal of
Open Source Software, 2(11).
VISAPP 2025 - 20th International Conference on Computer Vision Theory and Applications
296

Mora, C., Tittensor, D. P., Adl, S., Simpson, A. G., and
Worm, B. (2011). How many species are there on
earth and in the ocean? PLoS Biology, 9(8).
OpenAI (2023). ChatGPT 3.5.
Qin, Y., Ding, S., Wang, L., and Wang, Y. (2019). Research
Progress on Semi-Supervised Clustering.
Ravi, S. and Larochelle, H. (2017). Optimization as a model
for few-shot learning. In 5th International Confer-
ence on Learning Representations, ICLR 2017 - Con-
ference Track Proceedings.
Ren, M., Triantafillou, E., Ravi, S., Snell, J., Swersky, K.,
Tenenbaum, J. B., Larochelle, H., and Zemel, R. S.
(2018). Meta-learning for semi-supervised few-shot
classification. In 6th International Conference on
Learning Representations, ICLR 2018 - Conference
Track Proceedings.
Ren, Y., Hu, K., Dai, X., Pan, L., Hoi, S. C., and Xu, Z.
(2019). Semi-supervised deep embedded clustering.
Neurocomputing, 325.
Rodriguez, M. Z., Comin, C. H., Casanova, D., Bruno,
O. M., Amancio, D. R., Costa, L. d. F., and Rodrigues,
F. A. (2019). Clustering algorithms: A comparative
approach. PLoS ONE, 14(1).
Russakovsky, O., Deng, J., Su, H., Krause, J., Satheesh, S.,
Ma, S., Huang, Z., Karpathy, A., Khosla, A., Bern-
stein, M., Berg, A. C., and Fei-Fei, L. (2015). Ima-
geNet Large Scale Visual Recognition Challenge. In-
ternational Journal of Computer Vision, 115(3).
Snell, J., Swersky, K., and Zemel, R. (2017). Prototypical
networks for few-shot learning. In Advances in Neu-
ral Information Processing Systems, volume 2017-
December.
Song, Y., Wang, T., Cai, P., Mondal, S. K., and Sahoo, J. P.
(2023). A Comprehensive Survey of Few-shot Learn-
ing: Evolution, Applications, Challenges, and Oppor-
tunities. ACM Computing Surveys, 55(13s).
Strehl, A. and Ghosh, J. (2003). Cluster ensembles - A
knowledge reuse framework for combining multiple
partitions. In Journal of Machine Learning Research,
volume 3.
Sun, B., Zhou, P., Du, L., and Li, X. (2022). Active deep
image clustering. Knowledge-Based Systems, 252.
Tan, M. and Le, Q. V. (2019). EfficientNet: Rethinking
model scaling for convolutional neural networks. In
36th International Conference on Machine Learning,
ICML 2019, volume 97, pages 6105–6114.
Todi, A., Narula, N., Sharma, M., and Gupta, U. (2023).
ConvNext: A Contemporary Architecture for Convo-
lutional Neural Networks for Image Classification. In
Proceedings - 2023 3rd International Conference on
Innovative Sustainable Computational Technologies,
CISCT 2023.
Vilhagra, L. A., Fernandes, E. R., and Nogueira, B. M.
(2020). TextCSN: A semi-supervised approach for
text clustering using pairwise constraints and convo-
lutional siamese network. In Proceedings of the ACM
Symposium on Applied Computing.
Vinh, N. X., Epps, J., and Bailey, J. (2009). Information the-
oretic measures for clusterings comparison: Is a cor-
rection for chance necessary? In Proceedings of the
26th International Conference On Machine Learning,
ICML 2009.
Vinyals, O., Blundell, C., Lillicrap, T., Kavukcuoglu, K.,
and Wierstra, D. (2016). Matching networks for one
shot learning. In Advances in Neural Information Pro-
cessing Systems.
Von Luxburg, U. (2007). A tutorial on spectral clustering.
Statistics and Computing, 17(4).
Wah, C., Branson, S., Welinder, P., Perona, P., and Be-
longie, S. (2011). Caltech-UCSD Birds-200-2011
(CUB-200-2011). California Institute of Technology,
CNS-TR-2011-001.
Wang, Y., Chao, W.-L., Weinberger, K. Q., and van der
Maaten, L. (2019). SimpleShot: Revisiting
Nearest-Neighbor Classification for Few-Shot Learn-
ing. arXiv.
Wang, Y., Yao, Q., Kwok, J. T., and Ni, L. M. (2020). Gen-
eralizing from a Few Examples: A Survey on Few-
shot Learning. ACM Computing Surveys, 53(3).
Wu, J., Long, K., Wang, F., Qian, C., Li, C., Lin, Z., and
Zha, H. (2019). Deep comprehensive correlation min-
ing for image clustering. In Proceedings of the IEEE
International Conference on Computer Vision, vol-
ume 2019-October.
Xie, J., Girshick, R., and Farhadi, A. (2016). Unsupervised
deep embedding for clustering analysis. In 33rd In-
ternational Conference on Machine Learning, ICML
2016, volume 1.
Yang, J., Parikh, D., and Batra, D. (2016). Joint unsuper-
vised learning of deep representations and image clus-
ters. In Proceedings of the IEEE Computer Society
Conference on Computer Vision and Pattern Recogni-
tion, volume 2016-December.
Deep Image Clustering with Model-Agnostic Meta-Learning
297