Assignment 2: Innovative Technical Report on Machine Learning

[Advanced Machine Learning Course]


Research Background:

An Innovative Technical Report on Frontier Problems in Machine Learning

Machine learning is moving from single-task modeling toward more complex open environments. In recent years, foundation models, multimodal learning, data-centric machine learning, trustworthy machine learning, causal learning, reinforcement learning, graph learning, federated learning, and AI for science have developed rapidly. These directions not only improve model capabilities, but also introduce new technical challenges in generalization, interpretability, robustness, fairness, privacy protection, computational efficiency, and real-world deployment.

This assignment requires students to write an innovative technical report on a specific frontier problem in machine learning. The report should not be a simple survey of existing work. Instead, based on a solid understanding of related studies, it should propose a clear, concrete, and technically grounded idea, and analyze its feasibility, potential advantages, limitations, and possible validation methods.

A good innovative technical report should usually answer the following questions:

Requirement:

Overall Goal

Choose a specific research direction or application scenario related to machine learning and write a complete innovative technical report. The report should demonstrate your understanding of relevant literature, your analysis of the key problem, and an original idea with clear technical content. This assignment emphasizes problem formulation, technical insight, logical reasoning, and academic writing.


I. Topic Scope

The topic must belong to machine learning or a closely related interdisciplinary area. Possible directions include, but are not limited to:

The topic should be as specific as possible. For example, "large models" is too broad; "parameter-efficient adaptation for low-resource medical text classification" or "self-supervised graph representation learning for dynamic graph anomaly detection" would be more suitable report topics.


II. Technical Content Requirements

The report should include at least the following components:

  1. Problem formulation: Clearly define the research object, inputs and outputs, task setting, application scenario, and evaluation goals;
  2. Related work: Discuss at least 8 related papers, with recent work from the past five years forming the main body of the review;
  3. Method analysis: Compare the core ideas, applicability, and limitations of existing methods;
  4. Innovative proposal: Propose a concrete technical solution, such as a new model architecture, training objective, data mechanism, optimization strategy, evaluation protocol, theoretical perspective, or application framework;
  5. Feasibility analysis: Explain why the proposed idea may be effective, and provide an experimental design, theoretical analysis, or prototype validation plan;
  6. Limitations and risks: Analyze potential failure cases, implementation difficulties, data requirements, computational costs, and ethical risks.

If your report includes experiments or a prototype system, clearly describe the datasets, baseline methods, evaluation metrics, experimental settings, and major observations. A report without experiments must still provide sufficient technical reasoning and an executable validation plan.


III. Innovation Requirements

The report must contain clearly identifiable original content. The innovation may appear in any of the following forms:

Simply listing papers, restating existing methods, translating English materials, or presenting an idea without technical details does not satisfy the innovation requirement of this assignment.


IV. Suggested Report Structure

You are encouraged to organize the report in the form of an academic paper or technical white paper, including but not limited to:

The report should be logically organized, complete in structure, and carefully formatted. Figures, flowcharts, algorithm boxes, comparison tables, and necessary mathematical expressions are encouraged.


V. Format Requirements

Useful literature search portals include Google Scholar, DBLP, arXiv, OpenReview, PMLR, ACM Digital Library, and IEEE Xplore. Please prioritize original papers rather than relying only on blogs, news articles, or secondary interpretations.

How to submit?

Please submit your technical report in PDF format. If your work includes experimental code, data processing scripts, supplementary figures, or appendix materials, you may also submit them in a compressed package. The file or compressed package should be named "StudentID_Name_assignment2".

Submissions will be collected via Nanjing University Box. The submission link is: https://box.nju.edu.cn/u/d/724db663b35b40209991/ The password will be announced in the course QQ group.

Evaluation of the Report

Problem formulation and motivation (20%): The problem is clear, the motivation is sufficient, and the research or application value is well explained.

Understanding of related work (20%): The selected literature is appropriate, the summary is accurate, and the key differences and limitations of existing methods are clearly identified.

Innovation and technical depth (30%): The innovation is explicit, the technical solution is concrete, and the reasoning demonstrates sufficient machine learning expertise.

Feasibility and analysis (15%): The experimental design, theoretical analysis, or system validation plan is reasonable, and potential failure cases are discussed.

Writing and formatting (15%): The writing is concise and precise, the structure is clear, figures and tables are properly formatted, and citations are complete and consistent.

About the Deadline and Score

Deadline: June 21, 2026, 23:59.

This assignment will count toward the Assignment component of the course grade.

Additional Issues

Please follow academic integrity rules. Any text, figures, formulas, or ideas derived from existing papers, webpages, or generated by external tools must be properly cited or explicitly acknowledged. If you have any questions, please contact us via email: yuky@lamda.nju.edu.cn.