CS 445: Machine Learning
James Madison University, Fall 2026

Poster Projects

Most computer science conferences include a poster session where presenters stand beside large-format posters that summarize research results. Conference attendees then have the opportunity to sip beverages while circulating among the posters and talking to the presenters about their work.

The goal of our poster session is to give you a chance to dig into some area of machine learning that interests you by doing a small research project, and to present what you have learned. Every poster must be built around a project that you carry out by writing and running your own code: either an experimental question that you answer, or a system that you build and evaluate. Your reading of the research literature provides the context for your project: what is already known, what methods exist, and why your project is interesting.

Choosing a Project

I'll be flexible about topics. Projects can take two forms.

Answering a question. Some kinds of experimental projects that work well:

A good experimental question is narrow. "How do convolutional networks work?" is a topic, not a question. "How does the benefit of data augmentation for a small CNN on CIFAR-10 change as the training set shrinks from 50,000 to 1,000 images?" is a question. A useful template is: Does changing X affect Y on dataset D, as measured by M? Good questions change one thing at a time, and you shouldn't be able to predict the answer with confidence before running the experiment.

Building something. You may instead create a system or application that uses machine learning in an interesting way: a game-playing agent, a tool that sorts photos, a generator for music or text in some niche style. Projects like this must include an evaluation that examines how well your system works and where it fails. A live demo at the poster session is encouraged.

Either way, your project must involve working directly with machine learning models: training them, fine-tuning them, or otherwise working with the models themselves. A project that only sends requests to a hosted model through an API, such as a chatbot service, does not meet this requirement.

Keep your project feasible:

Your results need not be groundbreaking or publishable, and they need not be positive. A clearly established negative result ("this technique did not help, and here is the evidence") is a perfectly good outcome. Your results do need to be honest and to address your question or goal.

Partners

You may work alone or with a partner, and I encourage you to work with a partner. A team of two can take on a more ambitious project, and has a second person to catch mistakes in the experiments.

Using AI Tools

I strongly encourage you to use AI coding tools for this project. They make it possible to get much further in a few weeks than you could on your own. You do not need to understand every line of your code. Your job is to direct the work: decide which experiments to run, check that the code does what you intended, and interpret the results.

You do need to understand your experiments. At the poster session you should be able to explain what each experiment does, what data it uses, how that data was split, and why your results can be trusted. AI-generated code can run without errors while silently doing the wrong thing, such as evaluating on the training data. Build in sanity checks: compare against a simple baseline, look at actual model outputs, and be suspicious of results that seem too good. Your repository's README should briefly describe how you used AI tools.

The rules for the annotated bibliography are stricter. See below.

Deadlines

There will be five deadlines for this project. See the course schedule for dates.

Topic Proposal

For this deadline you should submit a short (1-2 paragraph) description of your topic and either the experimental question you plan to investigate or the system you plan to build, including the data you expect to use. You should also include at least three representative references. I suggest that you follow APA style in formatting your references. I won't be picky about formatting, but you should be consistent and you must provide complete bibliographic information.

Note that a URL is not sufficient. At the least, your references must include the name of the journal or conference, the names of the authors and the original publication date.

Annotated Bibliography, Project Plan and Repository

Annotated Bibliography

Your annotated bibliography must include at least five peer-reviewed conference or journal papers. At least three of these papers must contain a reference to some other paper in your bibliography. At least three of your papers must have been published within the last four years. For each paper, you must provide complete bibliographic information as well as a brief summary of the paper. The summary should describe the key results, note any references to other papers in your bibliography, and explain the connection to those referenced papers. The target audience for these summaries should be other students in CS 445.

Your bibliography should support your project. Your project should be motivated by at least one paper in your bibliography. Each summary should end with one or two sentences explaining how the paper bears on your project: for example, a method, baseline, dataset, or result that you are building on or testing.

You should make your best effort to understand the content of the papers, but your summary should be clear about elements of the paper that you did not understand.

I won't be picky about the overall formatting, but make sure to include your name, a title, etc. Here is an Overleaf template that you could use as a starting point: Annotated Bibliography Template.

You may use AI tools to help understand your papers, but the use of AI tools in generating the text of your summaries will be considered an honor-code violation. The summaries themselves are not the point. The point is for you to engage with these papers.

Zotero can be very helpful for storing and organizing references. It is also useful for citation formatting. JMU Libraries has a guide to using Zotero.

Project Plan

Along with your bibliography, submit a project plan of about one page. It should cover:

GitHub Repository

Submit a link to a GitHub repository for your project. By this deadline the repository should contain:

The repository may be public or private. If it is private, add me as a collaborator. Commit regularly.

Project Checkpoint

Submit a progress report of at most two pages describing what you have done so far. Include figures showing your preliminary results.

Poster

The poster should be landscape 36"x24" (3 feet wide and 2 feet tall). Here is a link to some slides with good advice on preparing a scientific poster:

ScientificPosterTips_FA12.pdf

Your poster should present your question or goal, enough background from the literature to understand it, your methods, and your results. It must clearly distinguish your own results from results reported in the papers you read. Include a link or QR code for your GitHub repository.

When planning your poster, imagine it as a sequence of slides for a 7-10 minute presentation on your topic. However, keep in mind that a poster differs from a PowerPoint presentation in terms of the amount of text that is appropriate:

Finding Papers

Here are some possible starting points for finding high quality papers.

Ask AI to suggest relevant papers. Be careful: AI tools sometimes invent references that look plausible but do not exist, or attach the wrong authors or venue to a real paper. Before a paper goes in your bibliography, find it yourself on the publisher's or conference's site, or through its DOI, and confirm the bibliographic details.

Use Google Scholar or Semantic Scholar to check citation counts. There are many low-quality or uninteresting papers out there. Citation counts provide a good mechanism for focusing attention on noteworthy papers. Once you find an interesting paper you can follow forward and backward citations to get a deeper understanding of the topic.

Conference publications are the main avenue for disseminating research results in computer science. I've highlighted a few of the top conferences in several AI areas below. Except where noted, the proceedings for these conferences should be available on-line.

First-tier Conferences:
First-tier Journals

A Note on arXiv.org

The arXiv.org web site provides a popular avenue for quickly disseminating research results that may not have undergone a formal process of peer review. Sometimes arXiv papers are under review for a journal or conference, sometimes they are pre-publication versions of papers that have since appeared in a peer-reviewed publication, sometimes they are more like informal white papers that are not intended for formal peer review.

The arXiv is great, but for the purposes of this project you may only use arXiv papers if they have actually appeared in a peer reviewed publication, or if they are clearly seminal papers (hundreds of citations). The reference information you provide must include the full publication information.

Reading Research Papers

Reading a research paper is not like reading a novel or even a textbook. Research papers are usually written under the assumption that the audience will be other researchers in the same field. In addition, papers are often written under strict page limits that restrict the amount of background information the authors can provide. The keys to making sense of research papers are patience and perseverance. I suggest the following steps.

By the time you finish, you should understand the key points that are being made in the paper. You may not understand every sentence and every equation, but you should know what you don't know, and be in a position to discuss it.

Grading

The grade for this project will be calculated as follows:

Proposal 10%
Bibliography, Project Plan and Repository 20%
Project Checkpoint 5%
Poster Draft 5%
Final Poster 40%
Poster Presentation 10%
Your Evaluations 10%

Your poster and presentation will be evaluated by me, as well as by other members of the class, and possibly other members of the department.