[Advanced Machine Learning Course]
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:
- What is the problem? Why is it important for machine learning research or real-world applications?
- How do existing methods address this problem? What are their key assumptions and main limitations?
- What is your proposed technical idea? What is new compared with existing methods?
- Why might the idea work? What experiments, theoretical analyses, or case studies can be used to validate it?
- In what scenarios might the method fail? How could it be improved in future work?
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:
- Foundation models, large language models, multimodal models, and parameter-efficient fine-tuning;
- Self-supervised learning, weakly supervised learning, semi-supervised learning, and active learning;
- Out-of-distribution generalization, test-time adaptation, robust learning, and uncertainty estimation;
- Trustworthy machine learning, including interpretability, fairness, privacy protection, security, and adversarial robustness;
- Data-centric machine learning, data quality assessment, data selection, and synthetic data;
- Graph machine learning, causal machine learning, reinforcement learning, and decision intelligence;
- Federated learning, edge intelligence, low-resource learning, and model compression;
- Applications of machine learning in scientific discovery, healthcare, financial risk control, education, industry, and social computing.
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:
- Problem formulation: Clearly define the research object, inputs and outputs, task setting, application scenario, and evaluation goals;
- Related work: Discuss at least 8 related papers, with recent work from the past five years forming the main body of the review;
- Method analysis: Compare the core ideas, applicability, and limitations of existing methods;
- 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;
- Feasibility analysis: Explain why the proposed idea may be effective, and provide an experimental design, theoretical analysis, or prototype validation plan;
- 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:
- Proposing a new machine learning problem setting or evaluation protocol;
- Designing a new model architecture, loss function, training procedure, or inference mechanism;
- Combining existing techniques in a new scenario and explaining the necessity and technical challenges of the combination;
- Proposing a new method for data construction, data selection, data augmentation, or data quality control;
- Providing a new theoretical explanation, empirical observation, or failure-mode analysis;
- Designing a deployable machine learning system for a real-world application scenario.
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:
- Title, author information, and abstract;
- Introduction: research background, problem motivation, and main contributions;
- Related work: summary and comparison of representative methods;
- Problem formulation: task setting, notation, and evaluation goals;
- Innovative method: core idea, technical details, algorithmic procedure, or system design;
- Feasibility analysis: experimental plan, theoretical analysis, case study, or expected results;
- Discussion: limitations, risks, ethical issues, and future work;
- References.
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
- The report may be written in either Chinese or English;
- It is recommended to use common academic paper templates, such as NeurIPS, ICML, AAAI, ACM, or IEEE;
- The main text is recommended to be 6--10 pages, excluding references and appendices;
- The reference format should be consistent, and all citations must be properly cited in the text;
- Please indicate at the end of the report whether generative AI tools were used, and specify how they were used.
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.
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.
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.
Deadline: June 21, 2026, 23:59.
This assignment will count toward the Assignment component of the course grade.
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.