We develop brain-inspired artificial intelligence to make advances in theory-development of brain function and AI technology focusing on energy efficiency, explainability, robustness, causal representation learning, and continual learning.
Artificial Intelligence (AI) systems are transforming society, opening unprecedented opportunities for innovation and new fundamental understanding. While AI systems are inspired by neural networks in the brain, they are far from reaching the brain’s robustness, flexibility, small need for training data, and low energy expenditure. In the bioAI group, we aim to implement brain principles into AI systems to inspire new architectures. These should be more robust when facing adversarial attacks, more flexible in transferring learning from one context to another and allow systems to continuously learn new tasks while reducing the energy demands. Parallel studies in neuroscience and AI will open for improved explainability, a significant societal challenge.
While AI has made tremendous progress for specific tasks, in particular within deep learning, the brain is still far superior to machines in many areas. AI systems are surprisingly brittle compared to the animal brain. Most AI models are vulnerable to adversarial attacks, where tiny changes e.g. in an image, which are not noticeable by the human eye, can lead to an abrupt change in how the image is categorized - for example, changing a stop sign to a 50 km/h sign - with possible catastrophic effects. Another challenge in AI systems is catastrophic forgetting in continual learning. Learning new tasks with the same AI machinery destroys previously learned knowledge, which again seems inherent to AI methods. These challenges have been resolved by the brain, where biological neural systems provide structures and ways of learning tuned by evolution. While the consensus of the AI community seems to be that neuroscience has little to offer AI in terms of algorithmic improvements, we argue that this is only in terms of incremental improvements. We believe that implementing AI counterparts has the potential to move AI towards brain flexibility, robustness, and performance by introducing new approaches that are fundamentally different from current practices and, therefore, may bypass their inherent challenges. At the same time, technological advances in neuroscience are transforming the way we interpret data and build models of the brain. These advances call for new theoretical approaches and analysis methods to deal with large-scale neural recordings. We focus on applications and fundamental development of machine learning methods – creating a symbiotic relationship between understanding natural and artificial intelligence.