Memory and Speed in Black and White Images

NOTE to Dr. Zhang and Dat:, We had to rework our project slightly. It is still the same thing as what we had before, but now we do not have the Unets making the dimensionality reduction. We were working on them last week and came to the conclusion that half or more of the project would be getthing that working, and, in that case, it would pretty much be a data creation project. Since we were not supposed to do a project like that, we decided to scratch the Unet idea and just easily make the dataset black and white. That way the majority of the project would be on what we already were going to do.

Abstract

Many problems with current networks are with memory consumption and speed. We will pre-process the images into black and white in order to decrease memory for new networks, by needing fewer scalars, and increase the networks speed, by needing fewer hidden layers to get a prediction faster. More specifically, we will be testing a common architecture, with normal pictures, against a simpler one, with pre-processing done, to see if the pre-processing of the pictures can yield similar accuracy. Then, we will test the memory usage and time for prediction per architecture.

Goals/Milestones

Goals

Determine if the pre-processing can allow a CNN to obtain the same accuracy as without.
See if the new, simplified CNN can use sufficiently less memory and be sufficiently faster than the common CNN.

Milestones

Progress milestones

Have the selected dataset ready and being used.
Have the common CNN obtained, and have the simplified CNN working.

Final report milestones

Collect and report the data
Compare the 2 CNNs.
Submit.

Dataset(s)

Fashion mnist dataset
Note: The Fashion mnist dataset is currently being listed as a backup in case CIFAR-10 is too difficult to work with.

CIFAR-10 dataset

Software

Libraries

Tensorflow
Numpy

Programs

We will first create a simple python program to turn the CIFAR-10 dataset black and white. Then we will create 2 other python programs to create the normal and simplified CNN architectures that will be used.