Summary
- Profile Type
- Technology offer
- POD Reference
- TODE20241120007
- Term of Validity
- 20 November 2024 - 20 November 2025
- Company's Country
- Germany
- Type of partnership
- Commercial agreement with technical assistance
- Investment agreement
- Research and development cooperation agreement
- Targeted Countries
- All countries
Contact the EEN partner nearest to you for more information.
Find my local partner
General information
- Short Summary
- A German university has developed a new neural-algebraic learning (NAL) concept that significantly enhances machine learning with neural networks. This innovative technology offers remarkable improvements in model accuracy and efficiency, achieving high performance with substantially less training data.The University is looking for cooperation partners, licensee and/or investors to further develop the new technology to series maturity.
- Full Description
-
The current state-of-the-art machine learning and neural network approaches face several significant challenges. One of the primary issues is the high computational cost associated with traditional neural networks, which often require numerous training epochs to achieve high accuracy. This translates to substantial demands on computational resources and time.
Data inefficiency presents another hurdle, as conventional approaches generally require vast amounts of training data to perform well. This can be particularly problematic in scenarios where data is scarce or expensive to acquire. Furthermore, the limited interpretability of most highly accurate and complex models, including neural networks, poses a significant challenge. These "black box" models often function in ways that are not easily understandable to humans due to their intricate nature. Moreover, generalization to unseen data remains a persistent challenge, especially when models are trained on limited datasets. This limitation in adaptability can significantly impact a model's real-world performance and utility.
The training process itself is fraught with potential pitfalls, including the risks of overfitting and underfitting. Models may perform exceptionally well on training data but struggle with unseen data, or conversely, fail to capture the underlying patterns in the data altogether. Determining the optimal point to halt training – a process known as early stopping – can be particularly challenging, as it requires striking a delicate balance between avoiding overfitting and ensuring the model has learned sufficiently.
In this context, a German university invented an innovative approach called neural-algebraic learning (NAL), which promises to revolutionize machine learning with neural networks. This new technology offers several significant advantages over traditional methods. NAL demonstrates remarkable improvements in model accuracy, achieving 85% accuracy after just 10 training epochs, compared to the 150 epochs required by conventional neural networks to reach the same level of performance. Moreover, NAL models exhibit dramatically faster convergence rates, approaching near-perfect accuracy around 38 epochs, while traditional networks struggle to match this performance even after 200 epochs.
This enhanced efficiency translates directly into reduced computational costs, as NAL requires significantly fewer training epochs to achieve high accuracy. The technology also addresses the common criticism of neural networks as "black boxes" by transforming them into interpretable algebraic function representations. NAL's flexibility is another key strength, as it can be applied to various algebraic structures like Boolean functions and polynomials, making it versatile for different types of problems. Additionally, the ability to train NAL networks using synthetically generated examples efficiently creates large, diverse training datasets.
In essence, NAL technology represents a significant leap forward in machine learning, offering a method to create more accurate and efficient models with less training data and computational resources while enhancing model interpretability. This advancement has the potential to greatly impact the field of artificial intelligence and its practical applications across various domains. It can be used in the following areas: Autonomous vehicles, Traffic prediction and management, Predictive maintenance, route optimization, Quality control and defect detection, Predictive maintenance, supply chain optimization, and Process optimization. - Advantages and Innovations
-
Advantages of the innovation are:
• Significantly improved model accuracy (nearly perfect after 38 epochs)
• Dramatically faster convergence rates (8 times faster than the conventional model)
• Reduced computational costs
• Efficient synthetic data generation
• Broad applicability across domains
• Improved efficiency with less training data (100% accuracy in only 38 epochs) - Stage of Development
- Available for demonstration
- Sustainable Development Goals
- Goal 17: Partnerships to achieve the Goal
- Goal 9: Industry, Innovation and Infrastructure
- IPR status
- IPR applied but not yet granted
Partner Sought
- Expected Role of a Partner
- The German university is actively seeking innovative companies and industry leaders interested in licensing this groundbreaking neural-algebraic learning (NAL) technology to enhance their existing products or develop new, cutting-edge solutions across various sectors, including autonomous vehicles, predictive maintenance, and process optimization.
- Type and Size of Partner
- SME 11-49
- SME 50 - 249
- Big company
- Other
- Type of partnership
- Commercial agreement with technical assistance
- Investment agreement
- Research and development cooperation agreement
Dissemination
- Technology keywords
- 01003012 - Imaging, Image Processing, Pattern Recognition
- 01003013 - Information Technology/Informatics
- 01003003 - Artificial Intelligence (AI)
- 01003017 - Speech Processing/Technology
- 01003020 - Building Automation Software
- Market keywords
- 02007011 - Manufacturing/industrial software
- 02007020 - Artificial intelligence programming aids
- 08002005 - Machine vision software and systems
- 02007013 - Banks/financial institutions software
- 02007016 - Artificial intelligence related software
- Targeted countries
- All countries