Everything a working AI researcher reaches for, in one place: templates to copy, a notation cheat sheet, a checklist to run before every experiment, and the classic papers worth your three passes.
The paper template
A complete, minimal LaTeX paper with an equation you can reference, a citation and a publication-quality table. Paste it into Overleaf and start writing.
LaTeX
\documentclass{article}
\usepackage{amsmath, amssymb, amsthm} % mathematics
\usepackage{graphicx} % figures
\usepackage{booktabs} % professional tables
\usepackage{hyperref} % clickable references
\title{Your Title: The Claim in a Few Words}
\author{Your Name \\ Ariadne Technology}
\date{}
\begin{document}
\maketitle
\begin{abstract}
Context. Problem. Approach. Results (with numbers). Implications.
\end{abstract}
\section{Introduction}
The problem, why it matters, and why it is unsolved.
Our contributions are:
\begin{itemize}
\item a method that \dots;
\item experiments showing \dots;
\item an analysis of \dots.
\end{itemize}
\section{Method}
We minimise
\begin{equation}
\mathcal{L}(\theta) = \frac{1}{n}\sum_{i=1}^{n} \ell\big(f_\theta(x_i), y_i\big).
\label{eq:loss}
\end{equation}
\section{Experiments}
Table~\ref{tab:main} compares against baselines; Equation~\eqref{eq:loss} defines the objective.
\begin{table}[t]
\centering
\begin{tabular}{lcc}
\toprule
Method & Accuracy (\%) & Seeds \\
\midrule
Baseline & $80.1 \pm 0.4$ & 5 \\
Ours & $81.3 \pm 0.3$ & 5 \\
\bottomrule
\end{tabular}
\caption{Mean $\pm$ standard deviation over 5 seeds.}
\label{tab:main}
\end{table}
\section{Limitations}
What the results do not show.
\bibliographystyle{plain}
\bibliography{references}
\end{document}
Notation cheat sheet
The conventions you'll meet in almost every machine-learning paper.
| Symbol | LaTeX | Usually means |
|---|---|---|
\mathcal{D} | the dataset of examples | |
\theta | all learnable parameters | |
f_\theta(x) | the model's output for input | |
\hat{y} | a prediction | |
\mathcal{L}(\theta) | the loss over the dataset | |
\ell | the loss for one example | |
\nabla_\theta \mathcal{L} | the gradient of the loss | |
\eta | the learning rate | |
| , | \mathbf{w}, W | a weight vector, a weight matrix |
| , | \sigma, \phi | an activation function |
\odot | elementwise product | |
\mathbb{E}[X] | the expected value of | |
\theta^\star | the optimal parameters | |
\operatorname*{arg\,min} | the argument that minimises |
For every symbol the academy teaches, with how to say it aloud and search by name, meaning or LaTeX, open the Notation Glossary. To practise turning notation into code, try the Problem Archive.
Experiment checklist
Run through this before, during and after every experiment. It saves to your browser, so treat it as a working checklist, not a poster.
0/12
Before you run anything
While running
When reporting
Reading log template
Keep one entry per paper. The last line is where research ideas come from.
Markdown
## <Paper title> (<first author>, <year>)
**Pass reached:** 1 / 2 / 3
**Five Cs:** Category · Context · Correctness · Contributions · Clarity
**In three sentences:** what they did, what they found, why it matters.
**Key evidence:** Table/Figure … shows …
**What I doubt:** …
**One question, one idea:** …
Classic papers
Sixteen papers that shaped the field, in chronological order. Start with Keshav's two pages, then use the three-pass method on the rest.
- 1958The perceptron: A probabilistic model for information storage and organization in the brain F. RosenblattThe first neural network that learned from data.
- 1986Learning representations by back-propagating errors Rumelhart, Hinton & WilliamsBackpropagation, and hidden layers that learn features.
- 1998Gradient-based learning applied to document recognition LeCun, Bottou, Bengio & HaffnerConvolutional networks reading handwritten digits.
- 2007How to read a paper S. KeshavThe three-pass method. Read this one first.
- 2012ImageNet classification with deep convolutional neural networks Krizhevsky, Sutskever & HintonAlexNet: the result that launched the deep-learning boom.
- 2014Adam: A method for stochastic optimization Kingma & BaMomentum plus adaptive step sizes, now the default optimiser.
- 2014Dropout: A simple way to prevent neural networks from overfitting Srivastava, Hinton, Krizhevsky, Sutskever & SalakhutdinovRegularisation by randomly silencing units.
- 2015Batch normalization Ioffe & SzegedyNormalising activations to make deep networks train faster.
- 2015Deep residual learning for image recognition He, Zhang, Ren & SunSkip connections that let networks go hundreds of layers deep.
- 2016Understanding deep learning requires rethinking generalization Zhang, Bengio, Hardt, Recht & VinyalsNetworks can memorise random labels, so why do they generalise?
- 2017Attention is all you need Vaswani et al.The Transformer, the architecture behind modern language models.
- 2018Deep reinforcement learning that matters Henderson et al.How seeds and implementation details change results.
- 2018BERT: Pre-training of deep bidirectional transformers Devlin, Chang, Lee & ToutanovaPretrain once, fine-tune everywhere.
- 2020Scaling laws for neural language models Kaplan et al.Loss falls predictably as a power law in data, compute and parameters.
- 2020Language models are few-shot learners Brown et al.GPT-3 and in-context learning.
- 2020Denoising diffusion probabilistic models Ho, Jain & AbbeelThe foundation of modern image generation.
Tools of the trade
arXiv Where most ML papers appear firstGoogle Scholar Search papers, follow citations, export BibTeXSemantic Scholar AI-assisted paper search and citation graphsZotero Free reference manager for your reading logOverleaf Collaborative LaTeX in the browserPyTorch The most common framework in ML researchHugging Face Models, datasets and papers in one hubMLflow Open-source experiment tracking