Noise in the primate nervous system appears to follow a systematic progression over the course of development. Specifically, neuro-physiological evidence reveals that cortical neurons early in development exhibit less noise than when more mature. This empirical observation led us to ask whether the seemingly deliberate network noise during later stages of development may serve a greater purpose for generalization. To investigate this, we explored the impact of a bio-mimetic training progression on the ability of a deep convolutional neural network to accurately classify images with degradations, such as various noise distributions into the testing images. Our experiments reveal clear benefits of the bio-mimetic noise progression in terms of making the networks more resilient to input degradations. Furthermore, results with this training regimen are superior to other training regimens, and demonstrate the importance of the temporal order of the simulated phases of the developmental progression. These findings suggest that noise progression during biological development may indeed serve an important adaptive purpose. By extension, these regimens suggest useful training strategies for improving artificial systems’ generalization capabilities.

I am taking SuperUROP because I have performed research in the past and it has been one of the highlights of my MIT experience. I have taken more computer science and machine learning courses and am hoping to use the knowledge I have acquired to make a tangible difference with my research.