Revolutionising Neural Network Training with Intelligent Hyperparameter Control

Información del proyecto

Paramorph

Identificador del acuerdo de subvención: 101330944

DOI

10.3030/101330944

Fecha de la firma de la CE 28 Julio 2026

Fecha de inicio 1 Agosto 2026

Fecha de finalización 31 Julio 2028

Financiado con arreglo a

• The European Innovation Council (EIC)

Coste total

Sin datos

Aportación de la UE

€ 2 441 694,94

Coordinado por

Inephany Ltd
United Kingdom

Objetivo

Paramorph is an advanced AI optimisation technology developed by Inephany to improve neural network training efficiency and efficacy. Training modern AI models is costly, time intensive, and energy demanding. Current practice relies on launching many separate runs with fixed hyperparameters to identify suitable configurations. This process consumes substantial GPU resources, increases financial and environmental costs and may yield suboptimal models.
Paramorph replaces this multirun trial approach with real time adaptive control within a single training run. Using a proprietary multi-agent reinforcement learning framework, it adjusts key hyperparameters at the level of individual model layers as training progresses. This enables fine-grained, continuous, learned optimisation during execution. No existing deployed solution provides real time, layer level, learned hyperparameter control within one training run.
The technology has demonstrated up to 3.5x faster convergence compared to widely used optimisers such as Adam under standard settings. Validation has been conducted on transformer models including GPT 2 and OLMo. Paramorph has reached TRL5 and is supported by an EU patent application, trade secrets, and a proprietary dataset covering more than 2,800 model configurations.
The target market includes AI research organisations, foundation model developers, quantitative finance firms, biotechnology companies, and enterprises operating compute intensive AI workloads. Paramorph reduces required experiments and shortens training time, lowering compute expenditure and energy consumption. Efficiency gains of up to 5x in large scale training could reduce carbon emissions by several thousand tonnes per cycle.
Support from the EIC will enable progression to TRL8 through large scale experimentation and structured enterprise pilots, de-risking the technology, unlocking private investment, and strengthening Europe’s position in sustainable, competitive AI infrastructure.

Ámbito científico (EuroSciVoc)

• ciencias naturales › informática y ciencias de la información › inteligencia artificial › aprendizaje automático › aprendizaje por refuerzo

Palabras clave

• Foundation models
• Reinforcement learning
• neural network optimisation
• hyper-parameter tuning
• meta-optimization
• LLM training optimisation

Programa(s)

• HORIZON.3.1 - The European Innovation Council (EIC) PROGRAMA PRINCIPAL
  
• HORIZON.3.1.2 - The Accelerator
  

Tema(s)

• HORIZON-EIC-2026-ACCELERATOR-02-OPEN - Accelerator Open
  

Régimen de financiación

HORIZON-EIC-ACC - HORIZON EIC Accelerator

Convocatoria de propuestas

HORIZON-EIC-2026-ACCELERATOR-02

Coordinador

Inephany Ltd

• Aportación neta de la UEn: € 2 441 694,94
• Dirección: WeWork, Aviation House, 125 Kingsway ; WC2B 6NH London ; Reino Unido
• Pyme Sí:
  Región
  London › Inner London — West › Camden and City of London
• Región: London › Inner London — West › Camden and City of London
• Tipo de actividad: Private for-profit entities (excluding Higher or Secondary Education Establishments)
• Coste total: Sin datos