Week 2 Presentation Tamique de Brito What Ive

Week 2 Presentation Tamique de Brito

What I’ve learned this week ● ● ● How to write data loaders with Py. Torch Optical flow How to use the Newton cluster (SLURM, Linux Command Line) Problems in Medical Imaging Network compression

What I learned from Assignment #1 Experimented with a variety of parameters and architecture and built up more intuition about neural networks. For example: ● Tried different filter sizes and found that for a fixed training time, 3 x 3 produces best accuracy. ● Tried different dropout rates and found that there is an initial increase in accuracy, but sudden drop when rate goes beyond 0. 35. ● Found that the best accuracy came from a 3 -layer 3 x 3 convnet with many features and 1 hidden layer. ● Tried some other “creative” architectures such as deeper network with only 2 x 2 filters, a network that progressively “switched out” deep layers for convolutional layers, randomly perturbing weights in certain layers and retraining, but these didn’t work well.

Exploration into research projects ● Narrowed down list of interesting projects to 8, then to 3. ● Read a bit more for each of the 3 projects then talked with each proposer to get a better idea. ● The projects I have found most interesting: ○ ○ ○ Self-supervised video rep. with predictive coding (Yogesh Rawat) Beyond frame-based segmentation (Kevin Duarte) Supervised temporal-contrastive learning (Ishan Dave)