Participants interested in exploring the project (+potential role, e.g. driver, advisor, funder, etc)
Roman Bauer (computational modeling & validation based on the experimental data)
Judd Rosenblatt (funding [particularly focusing towards much larger amounts and moving more projects forward faster] and/or help with accelerating software implementation)
This talk summary discusses the concept of the Neural Pretty Printer, a technology that aims to convert artificial neural networks into biologically realistic representations that can be tested and reconstructed. The Neural Pretty Printer involves converting a neural network into a Hodgkin Huxley compartment model and then into a computer graphics representation of neurons with synapses. This graphics representation is further translated into a noisy picture to test and reconstruct biological features.
The advantages of the Neural Pretty Printer include the ability to simulate various types of live data and gain insights into biological mechanisms. It can also be modified to study complex biological phenomena. However, there are challenges in implementing the Neural Pretty Printer, such as generating non-overlapping neural networks and simulating their physical interactions realistically. Temporal synchronization and training with biologically plausible models are also obstacles that need to be addressed.
In terms of impact and future directions, the Neural Pretty Printer can be used to improve scanning and reconstruction methods for neural networks. It has the potential to become a comprehensive system for predicting outcomes with more physicality. Collecting real data to train and improve the system’s fidelity may lead to more accurate testing and emulation of biological structures.
The talk also highlights the importance of considering the functional effects of glial cells and the need to update the model as more is learned about these effects. Testing the model on different scenarios can help identify any missing elements or functions that need to be accounted for.
Two suggested metrics for success are masking part of the network and having the model accurately regenerate that part, as well as comparing the input-output of the original data with the model’s output to measure performance. These metrics can help evaluate the success of the Neural Pretty Printer in replicating neural network structure and function accurately.