What can artificial intelligence learn from biological brains? At this Wednesday’s lunch seminar series at CDS, Professor Partha Mitra from the Cold Spring Harbor Laboratory explained how he has been mapping biological brain connectivity in his Mouse Brain Architecture Project to discover how we can transfer biological brain architectures to machine brains.
AI has made major strides in the last decade but, as Mitra explained, they still have some drawbacks when compared to biological brains. Not only do they use more power (100,000 W compared to 10 W!), but machine brains also require a much larger data set to train on than biological brains, and have non-biological fragilities.
The largest challenge, however, is that machine brains are still largely ‘niche adapted’, meaning that they are unable to transfer their skills successfully to tasks outside of their main purpose.
While some, like CDS’ very own Kyunghyun Cho, have turned to exploring new encoder-decoder approaches in neural networks to overcome this problem, Mitra suggested that analyzing machine brains within the context of biology and neuroanatomy also has a lot to offer.
Drawing on the classic ‘nature versus nurture’ debate that faces human brains, Mirta pointed out that biological learning takes place on two different scales. First, it takes place at the level of the mesocircuit, a series-specific brain circuit responsible for consciousness (nature). Second, it takes place at the level of adapting to individual environments or experiences over time (nurture).
An ideal machine brain would comprise of similar components—on one hand, it would have a specifically engineered architecture but, on the other, its neural networks would also mimic biological brain plasticity that would enable it to adapt, change, and learn in real time. A primary goal behind Mitra’s investigation of mouse brains and their nervous system, then, is to discover how this plasticity works in biological brains, and see if it can be reproduced in machines.
Alongside this project, Mitra explained that he is also working on linking the field of statistical physics to machine learning and data science.
by Cherrie Kwok