RIT CIS Internship: Machine Learning

I started today debugging my code for my neural network. After a healthy amount of googling and some perseverance I was able to roughly train and evaluate my model. The performance was subpar but hopefully I will be able to improve that next week. When that was over I went to the free pizza lunch provided for all of the REU summer research students and I then attended a lecture on visual perception. Although this was not very related to my project I still found it to be a valuable use of my time. It was extremely mind-blowing to realize that our eyes can only see details in a very small field of vision. Our peripheral vision is a lot worse than I previously thought. One of the coolest visual demos was these two tables shown below. Believe it or not the tabletops are exactly the same size and shape. Human visual perception is fascinating and I plan to continue attending the lectures throughout the summer. Source:  http://www.optical-illusionist.com/illusions/table-size-illusio

I spent the majority of this morning working on my feed forward neural network. I had some trouble inputting the feature vector because it was one dimensional. After struggling to find a solution I finally found a model that would work with the feature vector. Later in the day Frank, a fifth year CE student gave a presentation on what he was doing this summer in k-lab. His project involved using google street view images to train a neural network to identify invasive species. I found his presentation very interesting. It opened my eyes to the vast number of uses of neural networks. Tomorrow I start to learn about VQA- Visual Question Answering.

Most of today was spent working on my program. Despite some initial struggles with importing data and feeding that data to the network I was able to extract a feature vector by the end of the day. A feature vector is a list of values given by the pre-trained neural network that can be used to classify images. Using this feature vector I was able to classify 10 images with 100% accuracy. Although with a larger data set that accuracy may have fallen I am still extremely impressed with the application of a pre-trained neural network. Source:  https://brilliant.org/wiki/feature-vector/

Today I began to experiment with using a pre-trained neural network in order to have a higher accuracy in testing. I used resnet-18 which is a neural net that has been trained on a data set with 14 million images. This allows resnet-18 to identify specific features like edges or corners. The detection of these edges is called feature identification and it is a large advantage of using a pre-trained neural network. Using resnet-18 also reduces training time because the network already has a good idea on how to tell apart images. I plan to test my pre-trained neural net on the caltech-101 data set later this week. Source: http://web.eecs.umich.edu/~honglak/cacm2011-researchHighlights-convDBN.pdf

Today I read some academic journals in order to gain a better understanding of lifelong learning. The problem lifelong learning hopes to solve is something called catastrophic forgetting. Catastrophic forgetting is a term used to describe the inability of a neural network to to remember data it is given after training. For a network that needs to continually update itself catastrophic forgetting is a large problem. One approach lifelong learning algorithms take is to continual train on the new and old data in order to maintain its effectiveness on its training data. This approach is not very memory efficient. As the week goes on I plan to learn more about lifelong learning.