Similarities between insect brain connectome and machine learning architectures revealed

For the first time, an international research team has mapped out the entire connectome of the fruit fly larva (Drosophila melanogaster). Their study reveals structural similarities between the brain’s neural networks and the advanced machine learning (ML) architectures.

Imaging whole brains and mapping the neural connections has been a challenge for a long time. However, the latest technological advancements like CATMAID allowed the researchers to capture vast amounts of data at nanometer scale and produce detailed reconstructions of the brain.

Brain hemispheres of the fruit fly larva

The brain structure of the fruit fly larva consists of 3,013 neurons and 544,000 synapses, with nested recurrent loops and multi-layer shortcuts that bear similarity to resilient ML models.

The connectome of the insect exhibits a broadly distributed recurrent architecture, wherein approximately 41% of neurons participate in repetitive poly-synaptic loops. This structure may explain how shallow biological neural networks can achieve performance levels akin to significantly deeper artificial neural networks (ANNs).

The research revealed that the brain neurons follow longer and shorter paths that bypass layers, similar to the designs found in specific ANNs such as ResNets, DenseNets, and U-Net convolutional architectures.

ANNs utilize shortcuts to enhance their computational abilities, potentially allowing them to rival or surpass deep networks that lack these shortcuts. Such observations could imply a similar function in the brain. The layer-skipping process identified in the brain’s network may have the potential to increase its computational capacity.

Conclusion

The connectome analysis provides a roadmap for forthcoming research on the connection between artificial neural networks (ANNs) and the brain. Additionally, it could encourage the creation of new machine learning (ML) designs.

In the future, it may become possible for researchers to devise more effective treatments for neurological disorders and create improved AI systems that can learn and adapt in ways similar to the brain.

Learn more:

Other popular posts