Dear Arunabha,

"I'm particularly interested in perspectives on how we can bridge the gap between AI's analytical strength and the nuanced, dynamic nature of the brain."

Most questions, most challenges one"s set himself to solve are the results of mis-conceptions and confusion. I am not saying, it is the case here but that it is generally the case. How one find out? By investigating the question closely and making sure it does not containt self-contradiction. The above question makes a parallel between a brain, one must assume a human brain, and AI. Very strange parallel. AI are human programing artefacts that do not even have a common denominator except being programs made by humans. Human brains are the part of the human nervous system located in the cranial cavity which came about throw the biological evolution of the last 500 million year of animal nervous systems which are only one aspect of their bodies and these bodies are just one aspect of the life of these animals. Since the question above involve making a parallel between human brains and AI, the first step would be to clarify what is parellel is about? I do not even see a parallel. Obviously, you seems to be able to see

such parallel. I don"nt. Maybe some clear statement of what this parallel is would be the first step in seeing if the question is the result of confusion or of insight would be appropriate. Assuming it is not confusionn and the two sides of the parallel are clarified, then one may begin to speculate about type of steps to bridge the gap. The hard part in clarifying the two sides is to clarify the nature of what we call brain since they are not our own artefacts, and they are only one aspect of our life and cannot be understood without understanding their role in our life. It seems to me that a huge job of clarification of the question is required before even thinking in terms of bridging a gap between two unknown.

Regards,

- Louis

One of the most significant roadblocks to building generalizable AI models such as those using Graph Neural Networks (GNNs) for understanding complex brain function is the high-dimensional and dynamic nature of brain connectivity. GNNs are well-suited for modeling brain networks, but challenges remain in acquiring consistent and large-scale graph-structured brain data ensuring biological interpretability of learned representations.

Arunabha Chowdhury Khadersab Adamsab Louis Brassard

The most significant roadblocks are:

1. Data scarcity and quality. We lack large‑scale, multimodal recordings of human brain activity at both high spatial and temporal resolution. Existing methods force trade‑offs (e.g., fMRI’s spatial detail vs. EEG’s millisecond timing), so AI models never see the brain’s full dynamic picture.

2. Inter‑subject variability. Human brains differ wildly in their wiring (connectomes), neurochemistry, and life‑history. A model trained on one cohort often fails to generalize to another because the underlying biological “fingerprint” changes from person to person. In a way, each human is like a star—distinct in structure, development, and energetic output. Just as no two stars share identical compositions or behaviors, no two brains function in precisely the same way. This individuality poses a major challenge for building broadly generalizable AI models.

3. Interpretability and mechanistic understanding. Deep learning excels at prediction but offers few insights into how networks of neurons give rise to cognition. Without interpretable architectures aligned with known neurobiology, AI remains a black box that’s hard to map onto real circuits.

4. Scale and complexity mismatch. Brains contain ~86 billion neurons with trillions of synapses, operating across molecular, cellular, circuit, and behavioral levels. Capturing that hierarchy in a single model is computationally and conceptually daunting.

5. Ethical and practical constraints. Invasive experiments (e.g., intracranial electrodes) provide the richest data but can’t be deployed widely in healthy humans. This bottleneck limits the ground truth we can use to train and validate AI.