To explore deep learning via hands on exercise. To try out deep learning on a new dataset.
- you will pose a deep learning solution for another problem
- you will import or generate data for that problem
- you will try 3 different deep learning architectures on your problem
- you will write a short report about your endeavour.
- you will commit both the report and the source code to a git repository
- you will host that repository on github
- you will share with the instructor the URI of that github repo e.g. https://github.com/abramhindle/theanets-tutorial
- 1 Github Repo URI that contains within it
- 1 derivative program that imports your data set
and sets up a deep learning network to
train, validate, and test on your dataset.
- your program will explore 3 different architectures
- each architecture should differ in the number of layers and in the number of neurons
- your program will explore 3 different architectures
- 1 sample dataset if it is less than 100mb
- or the script to generate it
- 1 PDF or Markdown Report
- 1 description of your data set
- include source code if you generate it
- instructions how to get the dataset or generate it
- 1 description of your problem
- 1 description of you condition your inputs
- 1 description of your interpret your outputs
- 1 description of performance of each architecture on the deep learning task. Describe accuracy and TP, TN, FP, FN.
- 1 discussion whereby you suggest possible ways of improving the result
- 1 summary
- 1 description of your data set
- 1 derivative program that imports your data set
and sets up a deep learning network to
train, validate, and test on your dataset.
10 marks:
- 20% Dataset generation / import
- 50% Project Report
- 30% Program
Each component is rated as Unsatisfactory, Poor, Satisfactory, Good, Excellent.
-
If you have more than 4 pages then you're probably writing way too much.
-
I expect about 1-2 pages.
-
You can modify my data-generators. But it has to be distinct and different.
-
You are free to use other source code as long as the licenses are compatible and you have permission and you can cite it.
-
The deliverable will be in a github repo
-
The only text you submit will be the url to your github repository
-
You github repo will contain your source code and report
-
You may fork the theanets-tutorial from https://github.com/abramhindle/theanets-tutorial
-
You will cite all sources -- for code and for writing.