How to Read a Machine Learning Research Paper

To stay current in AI — and to eventually contribute — you must read research papers. Beginners often find them intimidating: dense notation, unfamiliar terms, assumed knowledge. The good news is that reading papers is a skill, and it improves quickly with the right method. This lecture gives you a practical approach, based on S. Keshav's widely used "How to Read a Paper" method, adapted for machine learning.

Anatomy of an ML paper#

• Title and abstract — the claim in brief.
• Introduction — the problem, why it matters, the gap in prior work, the contributions (often a bulleted list — read it carefully).
• Related work — context and positioning.
• Method — the proposed approach: model, objective, algorithm.
• Experiments — datasets, baselines, metrics, main results, ablations (removing components to show which matter).
• Discussion / limitations — where it fails; often the most honest section.
• Conclusion — summary and future work.
• Appendix — hyperparameters, proofs, extra results; essential for reproduction.

The three-pass method#

Pass 1 — the bird's-eye view (5–10 minutes)#

Read the title, abstract, introduction, section headings, figures and tables (especially Figure 1 and the main results table), and the conclusion. Skip the maths.

Answer the five Cs: Category (what type of paper?), Context (which prior work does it build on?), Correctness (do the assumptions look valid?), Contributions (what is new?), Clarity (is it well written?). Decide whether to continue.

Pass 2 — understanding the content (about an hour)#

Read the whole paper carefully but skip detailed proofs. Study figures and tables closely: axes, error bars, which baselines are compared. Mark unfamiliar terms and references to read later. At the end, you should be able to summarise the main idea and evidence to a classmate.

Pass 3 — deep understanding (several hours)#

Re-derive the key equations; mentally (or actually) re-implement the method; question every assumption; identify what you would do differently. For papers central to your work, reproduce a result with the authors' code — the best way to truly understand it.

Reading the method section#

• Identify inputs, outputs and the objective (what is being optimised).
• Map notation to concepts; write your own notation table.
• Draw the architecture or pipeline.
• Ask: what is the one key idea? Most good papers have one.

Critical reading: questions to ask#

1. Baselines: are they strong, recent and equally tuned? Weak baselines inflate improvements.
2. Statistical rigour: multiple seeds? confidence intervals or standard deviations? Is the improvement larger than the noise?
3. Datasets: are they appropriate and diverse? Could there be leakage or contamination?
4. Ablations: do they show that each component matters?
5. Compute: how much was used? Would baselines improve with the same budget?
6. Generalisation: does it work beyond the benchmark? Out-of-distribution?
7. Claims vs evidence: do the words in the abstract match the numbers in the tables?
8. Limitations and societal impact: are they discussed honestly?

A reading template#

# Paper notes: <title> (<authors>, <venue/year>)

**One-sentence summary:**
**Problem & motivation:**
**Key idea:**
**Method (my words + diagram):**
**Evidence:** datasets / baselines / main numbers (with variance?)
**Ablations — what matters most:**
**Strengths:**
**Weaknesses / questions / red flags:**
**Relevance to my work:**
**Follow-up papers to read:**

Keeping notes in a consistent format builds a personal knowledge base you will use for years.

Finding and keeping up with papers#

• arXiv (cs.LG, cs.CL, cs.CV, cs.AI) — preprints, not yet peer-reviewed; read with appropriate caution.
• Major venues: NeurIPS, ICML, ICLR (general ML); ACL, EMNLP, NAACL (NLP); CVPR, ICCV, ECCV (vision); AAAI, IJCAI; plus domain venues (e.g. health, humanitarian technology, ACM COMPASS for computing and sustainable societies).
• Semantic Scholar and Google Scholar — search, citation graphs, alerts.
• Papers with Code-style resources and GitHub repositories for implementations.
• Survey papers — the best entry point to a new area.
• Reading groups — discussing a paper weekly with peers accelerates learning enormously.