=Paper=
{{Paper
|id=Vol-2451/paper-19
|storemode=property
|title=FAIRnets Search - A Prototype Search Service to Find Neural Networks
|pdfUrl=https://ceur-ws.org/Vol-2451/paper-19.pdf
|volume=Vol-2451
|authors=Anna Nguyen,Tobias Weller
|dblpUrl=https://dblp.org/rec/conf/i-semantics/NguyenW19
}}
==FAIRnets Search - A Prototype Search Service to Find Neural Networks==
https://ceur-ws.org/Vol-2451/paper-19.pdf
FAIRnets Search - A Prototype Search Service
to Find Neural Networks
Anna Nguyen? and Tobias Weller?
Karlsruhe Institute of Technology (KIT), Karlsruhe, Germany
{firstname.lastname}@kit.edu
Abstract. Research on neural networks has gained significant momen-
tum over the past few years. A vast number of neural networks is current-
ly being developed and trained on available data in research as well as
in industry. As the number of neural network architectures increases,
we want to support people in the field of machine learning by making
existing architectures easier to find and reuse.
In this Demo, we support the findability and reusability of Neural Net-
works by using the FAIRnets Search. Attendees will learn how to use
the FAIRnets Search web service to search the FAIRnets dataset. The
FAIRnets dataset is an RDF dataset containing information about alrea-
dy modeled neural networks. By applying RDF and OWL, our system
can be queried using SPARQL queries indicating the desired character-
istics of the neural network. As a result, all neural networks fulfilling
the search query are returned to the user. The returned search results
support users to gain insights into existing neural networks. Furthermore,
we give the possibility to get more detailed information about the archi-
tecture of the networks, as well as further links. The demo is available
at http://km.aifb.kit.edu/services/fairnets/.
Keywords: Neural Network · Ontology · Reusability · FAIR.
1 Introduction
Neural networks (NNs) have become an important tool in research to make
predictions on measured data. Apart from preparing the data, it is use case and
data specific which neural network to use regarding architecture and parame-
ters. Despite the huge amount of available neural network architectures online,
finding one that fits your problem is quite challenging because of information
overload. There are approaches in machine learning such as neural architecture
search [1] for neural network architecture design based on human knowledge
and trial-and-error. However, these methods are time and memory consuming
following a brute-force approach. Therefore, we want to focus on finding exist-
ing archi- tectures given a use case to amplify the reuse of pre-trained neural
networks. Until now, there are neural network architecture repositories from
?
These authors contributed equally to the work.
Copyright © 2019 for this paper by its authors. Use permitted under Creative Commons License
Attribution 4.0 International (CC BY 4.0).
2 A. Nguyen and T. Weller
the Berkeley Artificial Intelligence Research Lab Caffe Model Zoo1 , Keras2 and
Wolfram Alpha3 containing a variety of architectures. However, these reposi-
tories represent only a fraction of the already developed neural networks. An
overview of the neural networks implemented in practice is not given. With
this demo, we want to give the user the possibility to access and query im-
plemented neural networks. For this purpose, we make FAIRnets Search avail-
able to search for neural networks in the FAIRnets dataset which is available at
https://zenodo.org/record/3228378. This dataset contains over 500 publicly
available neural networks under a license uploaded to GitHub. It is modeled in
RDF and uses the Neural Network Ontology for modeling the information which
is available at https://w3id.org/nno/ontology. These two resources are de-
scribed in a previous work [2] on representing neural networks according to the
FAIR guiding principles [4].
Based on these two existing resources, we present in this work FAIRnets
Search, a search service to query, search and find neural networks. The following
three use cases are covered by our demo:
– Search for neural networks
– Search for used datasets
– Fine-grained Search by exploiting the SPARQL Endpoint.
2 FAIRnets Search
FAIRnets Search is a service provided by us to make neural networks searchable
and findable. The service is available at the following URI http://km.aifb.
kit.edu/services/fairnets/. It represents an attempt to search for all neural
network architectures or neural network instances that fulfill specific require-
ments (e.g., used for specific tasks, having a specific architecture, etc.). For this
purpose, the Web service uses the FAIRnets dataset. This dataset currently con-
tains more than 500 neural networks. The FAIRnets dataset is modeled in RDF
and uses the Neural Network Ontology to structure the information. Figure 2
shows an overview of the framework. We collected neural networks from GitHub
and retrieved the information. The data is annotated using the Neural Network
Ontology and represented in RDF. For each of the neural networks in the FAIR-
nets dataset, the relevant properties according to the Neural Network Ontology
such as the description and the architecture are stored. That way, the FAIRnets
dataset can be queried with a set of desired properties and responds with a set
of neural networks that have these properties with the SPARQL Endpoint. The
FAIRnets Search combines these implementations by a browser-based frontend
to the SPARQL endpoint.
1
https://github.com/BVLC/caffe/wiki/Model-Zoo, last accessed 2019-06-18
2
https://keras.io/applications/, last accessed 2019-06-18
3
https://resources.wolframcloud.com/NeuralNetRepository, last accessed 2019-
06-18
FAIRnets Search - A Prototype Search Service to Find Neural Networks 3
Link to GitHub
Reuse
Repository
Search Engine Query in
SPARQL
Transformation Neural Network
into Ontology Ontology
Data Extraction Annotation
in RDF
Neural Network Neural Networks
Repository in GitHub
Fig. 1: The FAIRnets Search Framework
3 Demonstration of Use Cases
The attendees of the demo will learn how FAIRnets Search can be used to
gain insights into the usage of existing neural networks and datasets in machine
learning. In the online demo, the users are encouraged to use the search engine to
find and access neural network architectures. With the FAIRnets Search Demo
we will tackle the following three scenarios:
Search for Neural Networks. The FAIRnets Search engine allows users to
search keyword-based for neural networks. The FAIRnets dataset is searched
using SPARQL. Multiple keywords are supported in the search. The results are
sorted based on the number of hits counted, i.e. how often the keywords appear
in title and description. The attendee of the demo can, for example, search for
the terms image and classification and will get a list of neural networks that are
related to these terms (see Figure 2). Existing neural networks in this area can
thus easily be retrieved. Detailed information on the individual neural networks
can be accessed on the model sites of the neural network. Information such as the
publisher, links, architecture information and the latest update of the network
are provided and shown by our demo. The attendees of the demo can choose
based on the information and links provided by us if an already modeled neural
network fits their use case. We support the reusability of neural networks with
the FAIRnets Search Demo.
Search for used datasets. Another use case is the usage of datasets. Attendees
of the demo can search for specific datasets (e.g. search for mnist). FAIRnets
Search lists neural networks that are related to the searched dataset. This gives
the attendees the possibility to find out which neural network architectures have
been applied to a given dataset. Additional information such as the link to the
GitHub repository is available on the respective pages. This allows for getting
4 A. Nguyen and T. Weller
Fig. 2: Returned hits of the FAIRnets Search Demo, based on the entered key-
words. In the search presented here, the user was interested in image classifica-
tion.
more information about the performance of the architectures on the datasets. Be-
sides identifying already applied neural network architectures on a given dataset,
the search can also be used to identify new datasets. This information is implicit
in the descriptions of neural networks. Searching for image classification lists all
available neural networks in this domain. In the description of the neural net-
work or on the corresponding GitHub repository page further information about
the used datasets for training can be found. This supports the attendees of the
demo to find new datasets suitable for their use case.
Fine-grained Search by exploiting the SPARQL Endpoint. Besides the
search functionality, we offer the attendees of the demo the possibility to post in-
dividual SPARQL queries to the FAIRnets endpoint. We use YASGUI [3] to dis-
play the results of the queries. The interface to the provided endpoint can be ac-
cessed via the following link: https://km.aifb.kit.edu/services/fairnets/
sparql. The endpoint allows for answering individual requests upon the dataset.
We already offer some pre-selected SPARQL queries, such as a list of all neural
networks with a maximum number of layers (see figure 3) and an overview of
the frequencies of the activation functions used. Further ad-hoc requests during
the demo are welcome.
FAIRnets Search - A Prototype Search Service to Find Neural Networks 5
Fig. 3: List of all GitHub links, which provide neural networks that have at most
10 layers.
4 Conclusions
This demo presents FAIRnets Search, a web service that allows users to query
for publicly available neural networks. FAIRnets Search allows for making neural
networks better findable, searchable and accessible. Also, it allows for gaining a
better understanding of the used architectures of neural networks. In three use
cases, we allow the attendees for querying neural networks, finding the use cases
for datasets and analyzing the structure of neural networks.
References
1. Elsken, T., Metzen, J.H., Hutter, F.: Neural architecture search: A survey. Journal
of Machine Learning Research 20(55), 1–21 (2019)
2. Nguyen, A., Weller, T., Sure-Vetter, Y.: Making neural networks fair (2019)
3. Rietveld, L., Hoekstra, R.: The YASGUI family of SPARQL clients. Semantic Web
8(3), 373–383 (2017)
4. Wilkinson, M.D., Dumontier, M., Aalbersberg, I.J., et al.: The FAIR Guiding Prin-
ciples for scientific data management and stewardship. Scientific Data 3 (2016)