What is the problem/issue being addressed?
Algebraic Machine Learning (AML) is a new paradigm built on abstract algebra and model theory. Unlike deep learning and other popular algorithms, AML is not statistical. Instead, it produces generalising models from semantic embeddings of data into discrete algebraic structures. The result is a paradigm that:
1. is far less sensitive to the statistical characteristics of training data and does not fit (or even use) parameters;
2. can flexibly combine unstructured and complex information with formal specifications of human knowledge, constraints and task goals;
3. offers higher mathematical transparency than deep networks and other optimisation-based methods, using sets and graphs to generate human-understandable descriptions of what, why, and how has been learned; and
4. can be implemented in a distributed way that avoids centralised, privacy-invasive data collections in favour of collaboration among many local learners.
Why is it important for society?
Humans do not think in a vacuum but need substantial accumulated knowledge to resolve even simple real-life problems. High-impact areas such as healthcare, where AI could do the most good, require enormous knowledge to tackle challenges. AI can perform tasks humans cannot, but for this to happen, machines must learn what we know, communicate with each other, and collaborate with us. Current machine learning struggles to combine heterogeneous knowledge because it is statistical and lacks processes for encoding formal knowledge.
AML has the right properties to address these issues. It is less affected by training data statistics and provides mechanisms to encode heterogeneous knowledge. It can learn and produce complex models with multiple interrelated variables while also generalising and memorising. Its non-statistical nature and ability to encode constraints make AML suited to overcoming bias in training data and aligning machine intelligence with human ethical principles. Compared to statistical “black boxes,” AML models offer greater human control and understanding in both directions of communication. Ambitiously, AML makes augmentation of human intelligence possible, enabling true “human-computer partnerships” based on mutual understanding of world models, goals, and ethics.
What are the overall objectives?
The aim is to leverage AML properties for a new generation of interactive, human-centric ML systems. Interactive means humans and intelligent machines jointly learn and reason. AML enables users not only to reflect upon learning but also to drive it, enhancing cognitive powers through interaction. In line with Human-Centric AI, we will:
- reduce bias and prevent discrimination by reducing dependence on statistical data (P1), integrating formalised human knowledge (P2), and exploring the how and why of learning (P3);
- facilitate trust and reliability by respecting hard human-defined constraints (P2) and enhancing explainability (P3);
- integrate complex ethical constraints into Human-AI systems by going beyond bias prevention to interactively shape ethics between humans and machines (P3);
- enable distributed, incremental collaborative learning beyond centralised and offline approaches (P4).
Conclusions of the action
The project successfully delivered on these objectives. AML was formalised through new theorems and the AML Description Language (D2.1–D2.4 D3.1–D3.4). Collaborative learning was realised via AML-IP, a distributed integration framework with DDS Router and Dashboard support (D6.1–D6.4). Transparency and explainability were advanced through human-AI interaction methods and prototypes (D4.1–D4.11). Ethical and cultural world models were embedded into AML, demonstrating its ability to integrate sociocultural constraints (D5.3–D5.8). Finally, the release of an open-source AML engine and applied use cases in robotics and creative HCI confirmed AML as a viable and explainable alternative.
