DEEPLAB: Deep Learning Research Infrastructure

Short Summary

The Deep Learning Research Infrastructure provides a versatile, high-performance computing environment for designing, training, evaluating, and deploying deep learning models across a broad range of architectures, including convolutional neural networks (CNNs), transformer-based models, recurrent networks, and hybrid approaches. Built on a combination of NVIDIA RTX 6000 ADA and H100 NVL GPUs, the testbed offers researchers and industry partners scalable GPU compute suited to experiments ranging from lightweight prototyping to large-scale model training. DEEPLAB supports the full deep learning research lifecycle, from dataset preparation and model architecture design through training, hyperparameter optimization, and performance benchmarking. The environment accommodates frameworks such as PyTorch and TensorFlow, and is well-suited to applications across domains including predictive analytics, signal processing, time-series forecasting, and representation learning.

Keywords: Deep Learning Infrastructure; GPU Computing; Neural Networks; Model Training; Predictive Analytics

Deeptech Area

Artificial Intelligence
Other

Hosting Institution and PI Info

Name of Host Organization	NOVA Information Management School (NOVA IMS), Universidade Nova de Lisboa
Department or Lab	MagIC (Information Management Research Center) - the NOVA IMS research and development center
Name of Building	Manuel Vilares Building
Physical Address	Campus de Campolide, 1070-312 Lisboa
Website Links	https://www.novaims.unl.pt/
Institutional contact name	Cristina Oliveira
Institutional contact email	magic@novaims.unl.pt
Principal Investigator Name	Professor Ian James Scott
Position / institutional role	Assistant Professor
ORCID	0000-0001-9699-4473
Email	iscott@novaims.unl.pt
TestBed Responsible Name (if different from PI)
Funding source(s) for TestBed’s acquisition	This testbed benefits from the resources of the NOVA Data & Analytics Hub (NOVA DAH), hosted at NOVA Information Management School (NOVA IMS) of Universidade NOVA de Lisboa. The work is supported by national funds through FCT (Fundação para a Ciência e a Tecnologia) under project UID/04152/2025 (https://doi.org/10.54499/UID/04152/2025) (Centro de Investigação em Gestão de Informação (MagIC)/NOVA IMS); by the Plano de Recuperação e Resiliência (PRR) under projects UID/PRR/04152/2025 (https://doi.org/10.54499/UID/PRR/04152/2025) and EQUIPAR+2: UID/PRR2/04152/2025 (https://doi.org/10.54499/UID/PRR2/04152/2025); and by LISBOA2030 under project LISBOA2030-FEDER-01317500.

Application Domain

Manufacturing
Healthcare
Logistics
Agriculture
Maintenance & Inspection

Application Cases

Application case:	Short description:
Time-Series Forecasting with Deep Learning	Train CNN and LSTM-based models on multivariate time-series data for demand forecasting, anomaly detection, and predictive maintenance in industrial and logistics settings.
Transfer Learning for Domain-Specific Classification	Fine-tune pre-trained CNN architectures (e.g., ResNet, EfficientNet) on custom labelled datasets for document classification, product categorisation, or sensor-based fault detection.
Transformer-Based Tabular and Sequential Data Modelling	Develop and benchmark transformer architectures adapted for structured tabular data and sequential business data, exploring attention mechanisms beyond NLP tasks.
Student Deep Learning Research Projects	The testbed supports MSc and PhD students in conducting deep learning experiments as part of dissertations and research projects, providing GPU compute that would otherwise be inaccessible, accelerating the research cycle from hypothesis to published result.

Potencial Stakeholders

Non-academic stakeholders	Industrial partners, SMEs, Startups, Government bodies, Professional associations, Public agencies and municipalities
Academic stakeholders	MSc students, PhD students, Researchers, Visiting researchers, Seconded researchers
Other types of stakeholders	R&I support professionals, R&I infrastructure operators, Innovation intermediaries, Technology transfer actors

Possible TRL and Exploitation Scenarios

TRL application range	4
Internal academic research	Yes
Collaborative research with external academic partners	Yes
Contract research / Proof-of-Concept for industry	Yes
Pilot / DeepTech Deployment in operational environment	No
Training services (courses, workshops, certification)	Yes
Service provision (testing, benchmarking, validation)	Yes
Open access for walk-in users (e.g. open days / hackathons)	No
Other (Secondments / sponsored access for visiting researchers under project-based or institutionally approved arrangements)	Yes

Formal access conditions and prerequisites

Type of contractual relationship	Academic partner	Industrial partner
No contract (direct access)	No	No
Direct contract between parties (e.g., research agreement)	Yes (See Note 1)	Yes (See Note 1)
Indirect contract between parties (e.g., project framework)	Yes (See Note 1)	Yes (See Note 1)

Note 1: All access is subject to terms and conditions.

Type of prerequisites	Description of prerequisites
Agreements	Confidentiality agreement for proprietary algorithms	In some cases (See Note 2)
	Data sharing agreement for datasets generated	In some cases (See Note 2)
	IP agreements	In some cases (See Note 2)
	Other	In some cases (See Note 3 and 4)
Insurance	Users must have appropriate liability coverage through their home institution	Yes

Note 2: Intellectual property, confidentiality, and exploitation conditions are governed by the applicable NOVA regulations, the CITADELS consortium framework, and any project- or service-specific agreements. Background IP remains with the original rightsholders. Foreground generated through collaborative or service activities will be managed according to the applicable contractual framework, including provisions on ownership, access rights, confidentiality, dissemination, and exploitation. Additional NDAs, data-processing agreements, or specific IP clauses may apply depending on the nature of the data, software, models, or other assets involved.

Note 3: Access is granted on a project-based or institutionally approved basis, subject to feasibility assessment, resource availability, compliance with data protection and security requirements, and acceptance of the applicable terms and conditions. Special arrangements may apply for CITADELS secondments and other approved visiting researcher schemes. Where sensitive, proprietary, or regulated assets are involved, additional safeguards may be required before access is enabled.

Note 4: Gurobi licenses can only be used for academic applications.

Training and Safety

Mandatory technical training	N/A
Recommended technical training	Recommended training on cluster operation, job submission, queuing in SLURM.
Mandatory safety requirements	N/A

Technical Components for the Testbed

Components:

Description:

Hardware

(physical equipment available in this TestBed)

1) NOVA DAH01 System Specifications:

a) CPU: 32-Core CPU - This processor provides a significant amount of processing power, enabling users to run multiple demanding tasks simultaneously, such as simulations, data processing, and other compute-intensive workloads.

b) GPU: 2 x Nvidia RTX 6000 ADA - These high-performance GPUs are designed to accelerate AI, HPC, and other GPU-accelerated workloads. With two RTX 6000 ADA GPUs, users can leverage massive parallel processing capabilities, handling large amounts of data and providing a substantial boost to performance.

c) Storage: 7TB NVMe Storage - This high-capacity storage solution provides rapid data access and transfer speeds, ideal for applications that require high-performance storage, such as data analytics, scientific simulations.

2) NOVA DAH02 System Specifications:

a) CPU: 112 CPU cores, providing a substantial amount of processing power for compute-intensive tasks. This will enable users to run multiple simulations, data processing, and other tasks concurrently.

b) GPU: 2 x Nvidia H100 NVL (Next-Generation High-Performance Computing) GPUs, which offer significant performance boosts for AI, HPC, and other GPU-accelerated workloads. The H100 NVL GPUs are designed to handle massive amounts of data and provide high-performance computing capabilities.

3) NOVA DAH WS includes:

a) 16 Lenovo Thinkstations P5 units, each equipped with an Intel(R) Xeon(R) W3-2423 processor, 32 GB DDRS-4800 MHz ECC memory, NVIDIA RTX(R) 2000 GPU with 16 GB GDDR6 (Ada Generation), and 1 TB PCIe Neg4 SSD.

b) Operating system Windows 11 Education,

c) Broad range of licensed and open-source software for data science, analytics, modelling, and visualisation, including but not limited to a broad range of licensed and open-source software for data science, analytics, modelling, and visualization, including but not limited to Python, R, Power BI, Tableau, SPSS, SAS, QGIS, ArcGIS, Docker, Anaconda, Visual Studio Code, and Zotero.

c) Storage: 500 TB of storage, providing ample space for storing large datasets, applications, and other data. This storage capacity will enable users to work with big data and store the results of their computations.

4) Others

Software

(needed to run
the TestBed)

1) SSH client

2) File transfer tools recommended

3) Apptainer runtime to test locally

Standards and regulations
(relevant for the safe and compliant operation of this TestBed)

N/A

Testbed Documentation

Type	Short description:
Tutorial	How to connect to NOVA DAH (Tutorial NOVA DAH01)
Tutorial	How to connect to NOVA DAH (Tutorial NOVA DAH02)

Ethical and Societal Aspects

Ethical and societal aspect:	Short description:
Responsible AI Development and the EU AI Act	Deep learning models developed on this testbed may fall under the EU AI Act's classification of high-risk AI systems, particularly when applied to domains such as employment, education, or safety-critical systems. Researchers are expected to conduct conformity assessments, maintain technical documentation, and implement human oversight mechanisms in line with the Act's requirements.
GDPR Compliance and Data Minimisation	Training deep learning models on personal or sensitive data requires adherence to GDPR principles, including data minimisation, purpose limitation, and lawful basis for processing. Where personal data is used, a Data Protection Impact Assessment (DPIA) should be conducted prior to commencing experiments, and appropriate anonymisation or pseudonymisation techniques applied.
Transparency and Explainability	Deep learning models are often perceived as black boxes, raising concerns about accountability in decision-making. Researchers using the testbed are encouraged to incorporate explainability methods (e.g., SHAP, Grad-CAM) to improve model interpretability, particularly in applications affecting individuals, such as credit scoring, hiring, or medical triage.
Democratisation of Advanced AI Research	By providing accessible GPU infrastructure to academic researchers, SMEs, and startups, DEEPLAB reduces the dependency on large commercial cloud providers, lowering barriers to entry for organisations with limited resources and supporting a more diverse and competitive European AI research ecosystem.
Environmental Responsibility	Large-scale deep learning training is computationally intensive and carries a significant energy footprint. Users of the testbed are encouraged to adopt efficient training practices (such as early stopping, mixed-precision training, and model pruning) to minimise energy consumption and align with NOVA IMS's broader sustainability commitments.

Funding Source

This testbed benefits from the resources of the NOVA Data & Analytics Hub (NOVA DAH), hosted at NOVA Information Management School (NOVA IMS) of Universidade NOVA de Lisboa. The work is supported by national funds through FCT (Fundação para a Ciência e a Tecnologia) under project UID/04152/2025 (https://doi.org/10.54499/UID/04152/2025) (Centro de Investigação em Gestão de Informação (MagIC)/NOVA IMS); by the Plano de Recuperação e Resiliência (PRR) under projects UID/PRR/04152/2025 (https://doi.org/10.54499/UID/PRR/04152/2025) and EQUIPAR+2: UID/PRR2/04152/2025 (https://doi.org/10.54499/UID/PRR2/04152/2025); and by LISBOA2030 under project LISBOA2030-FEDER-01317500.

More info

(TBD)