Project Details

ACTL3143 & ACTL5111 Deep Learning for Actuaries

Author

Patrick Laub

A complete deep learning project

Individual project over the term. You will:

  1. specify a supervised learning problem,
  2. collect and clean the data,
  3. perform an exploratory data analysis (EDA),
  4. create a simple (non-deep learning) benchmark model,
  5. fit two different deep learning architectures,
  6. perform hyperparameter tuning,
  7. write a discussion of the results.

Project components

The deliverables for the project will include:

  1. draft due at noon on Friday in Week 5 (10%),
  2. recorded presentation due at noon on Friday in Week 8 (15%),
  3. final report due at noon on Monday of Week 10 (15%).

Project draft (10%)

Draft should show that you have:

  1. specified your supervised learning problem,
  2. collected and cleaned the data,
  3. performed a basic exploratory data analysis,
  4. create a simple (non-deep learning) benchmark model.

Upload to Moodle by noon on Friday in Week 5, no late submissions.

Presentation

Create a 3–5 minute recording covering:

  1. the problem you are investigating,
  2. the source of the data,
  3. the deep learning approaches you are using, and
  4. preliminary results you have (table of metrics).

Deliverable: YouTube link (public or unlisted) to a special StoryWall page. Presentations will be “public” to the class.

Suggestions: aim to be fully public and give peer feedback.

Presentation marking scheme

  • Content (6%): did you cover the four points on previous slide?
  • Style (6%): are your slides/figures professional and do they enhance the presentation?
  • Delivery (3%): is the presentation interesting and within the time limit?
Tip

It is a critical skill to be able to condense a complicated project into a short pitch. The project report is where you will give us all the details.

Presentation tips

  • Each project is different, you decide which parts to focus on.
  • Not necessary to film yourself.
  • Nice to briefly show the data (look at my lecture slides for example).
  • Don’t go overboard on EDA. Mention the most important 1–2 facts (if any!) about the data. E.g. imbalanced classes for classification.
  • You can avoid adding ‘UNSW’ & the course code.

Report requirements

You are asked to cover the four requirements in the draft, and also:

  • fit two different deep learning architectures,
  • perform hyperparameter tuning,
  • write a discussion of the results and any potential ethical concerns.

Deliverable: Report (PDF file), Jupyter Notebook, and dataset (e.g. CSV or ZIP file). Submission not public, probably to Moodle.

Report marking criteria

  • Content (8%): did you cover the seven points in the ML workflow?
  • Style (5%): does your report look professional, are your plots/tables useful and unpixelated, do you have spelling or grammar errors, are you within the page limit, and is the text easy to read?
  • Code (2%): is your code clean and well-commented, have useless cells been pruned, does it give errors when the “Run All” button is pressed?

Avoid screenshots & code in the report.

Some comments on the report

  • Focus on deep learning: I’m most interested in seeing your ability to use and explain your neural networks. For example, your mastery of the Lee–Carter model is less important to demonstrate.
  • Hyperparameter tuning: The tuning is one significant change from the weekly StoryWall tasks. Add a table (for each neural network) showing (at least) two hyperparameters that you tuned.
  • Use appendices: If you run out of space, use appendices which are not counted in the page limit. E.g., the less urgent parts of your EDA can go in here.