Workshop on SBA:
Synergising the Human Brain and Artificial Neural Networks
April 26, 2024, Hybrid Workshop, Birmingham, UK
Split-room Discussion, Towards the Roadmap: Retrospective and Prospective
The participants will discuss and propose the roadmap for synergising human brains and ANNs, to improve the understanding of the human brain, and also integrate inspirations from the human brain to help ANN design. Specifically, example topics could be discussed including:
- How can the development of ANN benefit from insights from human neuroscience?
- How can human neuroscience benefit from ANN developments?
- What are the challenges and opportunities of ANN+Brain?
Discussion keypoints
The participants had a heated discussion towards the roadmap for synergising human brains and ANNs. A summarisation of the discussion with several interesting keypoints is as follows:
ANN benefit from human neuroscience
- The network structure matters (e.g. CNN, Transformers, etc.)
- Use less data for training, wrt the energy consumption in human brains
- Human visual system understanding
- Massive data-driven approaches in ANNs vs. learn quickly from fewer experiences in human brain
- Improving the interpretability and measurability of ANNs by experience accumulated from human brain research
- Inspired by the hierarchical structure of the cerebral cortex, higher-level features may guide lower levels to perform targeted feature extraction and learning
- A neuroscientist also left a comment of "nothing" .. :)
Human neuroscience benefit from ANN
- Backpropagation in ANNs
- ANNs might help us better understand and treat neurocognitive disorders (e.g. semantic disorder), by constructing ANN models that simulate the semantic processing functions of the human brain
- Current neuroscience research mostly focuses on deconstructing the structure and mechanisms of the human brain, while ANNs focus more on mimicking the external behaviour of the brain. An interesting hypothesis was raised: what if we treat the human brain as a black box function, without concerning ourselves with its internal working mechanisms, and only focus on its input-output relationships?
Challenges and opportunities
- Lack of ground truth for human brain signals. Currently, we are unable to accurately measure the activity of neurons in the human brain
- ANNs perform poorly on unseen data, while the human brain excels at generalisation.
- Some tasks that are difficult for ANNs are easy for the human brain.
- Difficulty in precise measurement of human brain activity.
- The predictive power of ANNs offers oppotunities of generating brain data for neuroscience research.
- How to keep up to date of the AI literature? (>_<)
We also had an interesting discussion around a list posted by an anonymous participant, about a perspective given by AI (ChatGPT)
- Agreements and disagreements
- Synergising ANNs and human brain...