# Know Your ML > Know Your ML is a citation-checked reading room for artificial-intelligence and machine-learning research. It combines a live feed of recent arXiv pre-prints, plain-language briefs on foundational papers, and long-form AI-written reviews. Every brief and report lists the openly accessible papers it is based on, so any reference can be followed straight to the source. No paywalls, no login, no ads. ## Content model - **Reports** (68): AI-written literature reviews of landmark papers. Each report is a synthetic work, not peer-reviewed original research, and lists its primary sources. Canonical URL: `/report.php?slug=`. - **Briefs** (20): short AI-written summaries of foundational papers, each carrying the arXiv identifier of the paper it summarises. - **Topics** (10): one hub per research area, listing its reports in chronological order. Canonical URL: `/topics.php?topic=`. - **Learning paths** (6): authored, ordered reading routes that begin at a founding paper and end at the current state of the art. Canonical URL: `/collections.php?path=`. - **Live papers**: a filterable arXiv feed on the homepage, served through a server-side proxy. - **Videos** (3): narrated motion-graphics films with a chapter timeline, English captions (WebVTT), a full transcript on the page and, where a film cites figures, the source lines shown on screen or a table of the numbers it quotes. Canonical URL: `/videos/`. ## Pages - [Home](https://knowyourml.yunusemrevurgun.com/): Live arXiv paper feed with category/date/sort filters and keyword search, plus briefs and sources. - [Reports](https://knowyourml.yunusemrevurgun.com/reports.html): Searchable, sortable directory of every research report, filterable by topic, difficulty and report type. - [Topics](https://knowyourml.yunusemrevurgun.com/topics.php): All topic hubs with report counts. - [Learning paths](https://knowyourml.yunusemrevurgun.com/collections.php): All curated reading routes. - [Search](https://knowyourml.yunusemrevurgun.com/search.php): Server-rendered keyword search across reports and briefs. - [Videos](https://knowyourml.yunusemrevurgun.com/videos): Narrated films on how AI works and how to think about it, with chapters, captions and transcripts. - [Playground](https://knowyourml.yunusemrevurgun.com/playground.html): Interactive ML visualisations that run entirely in the browser (gradient descent, backpropagation, attention, activation functions, perceptron, polynomial fitting, CNN kernels). - [About](https://knowyourml.yunusemrevurgun.com/about.html): How the site works, how sources are verified, and the editorial policy. - [Blog](https://knowyourml.yunusemrevurgun.com/blog/): Short explainers for people learning machine learning, each linked to the reports and playground tools that go deeper. ## Blog guides Explainers written for learners. Each one cites primary sources and links to the reports, learning paths and playground tools on this site. Index: `https://knowyourml.yunusemrevurgun.com/blog/`. - [Overfitting vs Underfitting: How to Tell Which One Your Model Has](https://knowyourml.yunusemrevurgun.com/blog/overfitting-vs-underfitting/): How to diagnose underfitting and overfitting from training and validation error, with a polynomial example and the fix for each. - [Gradient Descent and Backpropagation: How Neural Networks Learn](https://knowyourml.yunusemrevurgun.com/blog/gradient-descent-and-backpropagation/): How the loss, the gradient, the learning rate and the chain rule fit together, with a three-step worked example. ## Browse by topic - [Computer vision](https://knowyourml.yunusemrevurgun.com/topics.php?topic=vision): 14 reports — Image classification, detection, segmentation, representation learning and neural scene representations. - [Architectures](https://knowyourml.yunusemrevurgun.com/topics.php?topic=architectures): 12 reports — The network designs that mattered: attention, convolutions, residual and graph models, sequence-to-sequence and their descendants. - [Applications & prompting](https://knowyourml.yunusemrevurgun.com/topics.php?topic=applications): 9 reports — Reasoning, retrieval, speech and prompting: the techniques that turn a base model into a usable system. - [Efficiency & deployment](https://knowyourml.yunusemrevurgun.com/topics.php?topic=efficiency): 7 reports — Doing more with less compute: adapters, quantisation, fast attention, sparse experts and deployment trade-offs. - [Training & data](https://knowyourml.yunusemrevurgun.com/topics.php?topic=training): 6 reports — How models are actually trained: objectives, scaling budgets, regularisation and the data decisions behind the results. - [Generative models](https://knowyourml.yunusemrevurgun.com/topics.php?topic=generative): 6 reports — Adversarial, diffusion and latent-variable models for synthesising images and other high-dimensional data. - [Reinforcement learning](https://knowyourml.yunusemrevurgun.com/topics.php?topic=reinforcement-learning): 4 reports — Agents that learn from reward: value-based methods, policy gradients and self-play. - [Foundations & scaling](https://knowyourml.yunusemrevurgun.com/topics.php?topic=foundations): 4 reports — Scaling laws, compute-optimal training and the large-model results that reframed what a language model can do. - [Alignment & safety](https://knowyourml.yunusemrevurgun.com/topics.php?topic=alignment): 4 reports — Turning raw predictors into assistants: human feedback, instruction tuning, preference optimisation and self-critique. - [Interpretability](https://knowyourml.yunusemrevurgun.com/topics.php?topic=interpretability): 2 reports — What is actually happening inside a trained network, and how we can tell. ## Learning paths - [The Transformer lineage](https://knowyourml.yunusemrevurgun.com/collections.php?path=transformers) — 11 steps: Start with the attention mechanism, meet the original Transformer, then follow the line through BERT, GPT-3, T5, RoBERTa, vision transformers, rotary embeddings and the move away from attention with Mamba and sparse mixture-of-experts. - [Aligning models with humans](https://knowyourml.yunusemrevurgun.com/collections.php?path=alignment) — 5 steps: How raw next-token predictors became assistants: reinforcement learning from human preferences, instruction tuning, constitutional AI, direct preference optimization and self-critique loops. - [Generative image models](https://knowyourml.yunusemrevurgun.com/collections.php?path=generative) — 5 steps: From adversarial training to score-based diffusion and latent synthesis, plus the representation learning and scene models that made controllable generation possible. - [Computer vision, AlexNet to SAM](https://knowyourml.yunusemrevurgun.com/collections.php?path=vision) — 8 steps: How deep networks took over image understanding: the convolutions that started it, residual connections that made depth trainable, the transformer and contrastive turns, real-time detection, segmentation and neural scene representations. - [Teaching models to reason](https://knowyourml.yunusemrevurgun.com/collections.php?path=reasoning) — 5 steps: Prompting as a reasoning interface: chain-of-thought, self-consistency, tool-using ReAct agents and deliberate search over thoughts. - [Making models cheap to run](https://knowyourml.yunusemrevurgun.com/collections.php?path=efficiency) — 8 steps: Scaling laws and compute-optimal training set the budget; adapters, quantization, flash attention and sparse experts decide how it is spent at inference time. ## Videos Each video has its own page with a chapter list, a full transcript and, where it uses figures, the source lines named on screen or a table of the numbers it quotes. The MP4 and WebVTT captions are linked from the page. - [Local LLM Inference, Explained](https://knowyourml.yunusemrevurgun.com/videos/local-llm-inference-explained) (4:58): A 5-minute visual guide to running LLMs on your own hardware: memory bandwidth, quantization, the KV cache, llama.cpp vs vLLM, and when to scale out. - 0:00 What local inference is - 0:16 Tokens, context and the two phases - 0:34 Processors and memory bandwidth - 1:03 The memory ladder - 1:27 Kernels, attention and FlashAttention - 1:55 The roofline: memory-bound vs compute-bound - 2:16 Quantization: fewer bits per weight - 2:39 Fitting the model in memory - 3:03 The KV cache - 3:28 Seven runtimes, one job - 3:54 First token, tokens per second, batching - 4:12 When one machine is not enough - 4:44 The whole film in one list - [Why the Anti-AI Crowd Is Wrong](https://knowyourml.yunusemrevurgun.com/videos/why-the-anti-ai-crowd-is-wrong) (8:53): A 9-minute essay film that tests five common anti-AI arguments — jobs, art, learning, hype and energy — against history and published data. - 0:00 From today, painting is dead - 0:46 We've been here before - 1:33 “AI will take all the jobs” - 2:42 “AI art is theft” - 3:46 “AI makes us dumber” - 4:42 “It's all hype” - 5:59 “AI is burning the planet” - 6:59 What the critics get right - 7:56 The real choice - 8:33 Over to you - [Why You Do Not Understand AI](https://knowyourml.yunusemrevurgun.com/videos/why-you-do-not-understand-ai) (1:40): A 100-second explainer on why language models feel like they understand — processing fluency, next-token prediction, compression — and three questions to ask instead. - 0:00 The illusion of fluency - 0:19 What the model actually does - 0:51 Why we get it wrong - 1:17 What to ask instead ## AI research reports AI-generated reviews of real papers. Read any of them at `https://knowyourml.yunusemrevurgun.com/report.php?slug=`. - [mixup: Training on Convex Combinations](https://knowyourml.yunusemrevurgun.com/report.php?slug=mixup-training-convex-combinations) (2017 · Training & data · Intermediate): mixup trains neural networks on convex combinations of examples and labels, improving generalization and robustness to noise. - [XLNet: Generalized Autoregressive Pretraining](https://knowyourml.yunusemrevurgun.com/report.php?slug=xlnet-generalized-autoregressive-pretraining) (2019 · Architectures · Intermediate): XLNet combines permutation language modeling with Transformer-XL recurrence, outperforming BERT on SQuAD, GLUE and RACE benchmarks. - [Variational Autoencoders](https://knowyourml.yunusemrevurgun.com/report.php?slug=vae-auto-encoding-variational-bayes) (2013 · Generative models · Intermediate): Variational Autoencoders learn compressed latent representations with the ELBO objective and reparameterization trick for generative modeling. - [VGG: Very Deep Convolutional Networks](https://knowyourml.yunusemrevurgun.com/report.php?slug=vgg-very-deep-convolutional-networks) (2014 · Computer vision · Beginner): VGG showed that stacking small 3x3 convolutions into 16-19 layer CNNs wins at ImageNet classification — the simple deep-learning blueprint behind modern vision. - [SimCLR: Simple Contrastive Learning of Visual Representations](https://knowyourml.yunusemrevurgun.com/report.php?slug=simclr-simple-contrastive-visual-learning) (2020 · Computer vision · Intermediate): SimCLR learns visual representations without labels via contrastive learning on augmented image pairs, matching supervised accuracy with linear probes. - [PPO: Proximal Policy Optimization](https://knowyourml.yunusemrevurgun.com/report.php?slug=ppo-proximal-policy-optimization) (2017 · Reinforcement learning · Intermediate): PPO stabilizes policy gradients with a clipped objective, becoming the default RL algorithm behind robotics, games and RLHF fine-tuning. - [Masked Autoencoders Are Scalable Vision Learners](https://knowyourml.yunusemrevurgun.com/report.php?slug=mae-masked-autoencoders-scalable-vision) (2021 · Computer vision · Intermediate): Masked Autoencoders pretrain Vision Transformers by reconstructing 75% masked patches, scaling self-supervised vision with simple asymmetric design. - [Mask R-CNN: Instance Segmentation](https://knowyourml.yunusemrevurgun.com/report.php?slug=mask-rcnn-instance-segmentation) (2017 · Computer vision · Intermediate): Mask R-CNN adds a mask branch and RoIAlign to Faster R-CNN for instance segmentation — pixel-level masks, boxes and classes from one network. - [Layer Normalization](https://knowyourml.yunusemrevurgun.com/report.php?slug=layer-normalization-sequence-training) (2016 · Training & data · Beginner): Layer Normalization normalizes across features instead of batches, stabilizing RNN and Transformer training at any batch size. - [Knowledge Distillation: Teacher-Student Compression](https://knowyourml.yunusemrevurgun.com/report.php?slug=knowledge-distillation-teacher-student) (2015 · Efficiency & deployment · Intermediate): Knowledge distillation compresses big teacher models into small fast students using soft targets — the key to efficient edge AI. - [He Initialization: Delving Deep into Rectifiers](https://knowyourml.yunusemrevurgun.com/report.php?slug=he-initialization-delving-into-rectifiers) (2015 · Training & data · Intermediate): He initialization scales starting weights for ReLU networks, keeping signal variance stable so very deep rectifier models can train. - [Grad-CAM: Visual Explanations from Deep Networks](https://knowyourml.yunusemrevurgun.com/report.php?slug=grad-cam-visual-explanations-networks) (2016 · Interpretability · Beginner): Grad-CAM explains CNN decisions with gradient-weighted heatmaps, revealing which image regions drove each prediction without retraining. - [GoogLeNet and Inception: Going Deeper with Convolutions](https://knowyourml.yunusemrevurgun.com/report.php?slug=googlenet-inception-multiscale-convolutions) (2014 · Computer vision · Beginner): GoogLeNet won ILSVRC 2014 with Inception modules running multi-scale convolutions in parallel — deeper CNNs with far fewer parameters than rivals. - [Faster R-CNN: Real-Time Detection with Region Proposals](https://knowyourml.yunusemrevurgun.com/report.php?slug=faster-rcnn-region-proposal-networks) (2015 · Computer vision · Intermediate): Faster R-CNN made object detection near real-time with Region Proposal Networks sharing CNN features — anchors, RoI pooling and joint training explained. - [DETR: End-to-End Object Detection with Transformers](https://knowyourml.yunusemrevurgun.com/report.php?slug=detr-end-to-end-object-detection) (2020 · Computer vision · Intermediate): DETR replaces anchors and NMS with a transformer predicting objects as a set — end-to-end object detection with bipartite matching, explained simply. - [DDIM: Fast Deterministic Diffusion Sampling](https://knowyourml.yunusemrevurgun.com/report.php?slug=ddim-denoising-diffusion-implicit-models) (2020 · Generative models · Advanced): DDIM turns diffusion sampling deterministic and 10-50x faster with strided steps, keeping DDPM quality while enabling latent interpolation and editing. - [ControlNet: Conditional Control for Diffusion Models](https://knowyourml.yunusemrevurgun.com/report.php?slug=controlnet-conditional-control-diffusion) (2023 · Generative models · Intermediate): ControlNet adds spatial conditioning like edges and pose to frozen diffusion models via zero convolutions, enabling precise controllable image generation. - [Batch Normalization: Accelerating Deep Network Training](https://knowyourml.yunusemrevurgun.com/report.php?slug=batch-normalization-accelerating-deep-training) (2015 · Training & data · Beginner): Batch Normalization normalizes layer inputs per mini-batch, enabling higher learning rates and faster, more stable deep network training. - [Adam: Adaptive Moment Estimation](https://knowyourml.yunusemrevurgun.com/report.php?slug=adam-adaptive-moment-estimation) (2014 · Training & data · Beginner): Adam combines momentum with per-parameter adaptive learning rates, making it the default optimizer for training deep neural networks. - [A3C: Asynchronous Methods for Deep Reinforcement Learning](https://knowyourml.yunusemrevurgun.com/report.php?slug=a3c-asynchronous-methods-deep-rl) (2016 · Reinforcement learning · Advanced): A3C trains deep RL agents with parallel asynchronous workers, beating DQN on Atari in a fraction of the time on a single CPU. - [Visualizing Convolutional Networks](https://knowyourml.yunusemrevurgun.com/report.php?slug=visualizing-cnn-zeiler-fergus) (2013 · Interpretability · Beginner): Zeiler and Fergus visualized CNN layers with deconvolution, revealing edges, textures and objects emerging layer by layer inside the network. - [Tree of Thoughts](https://knowyourml.yunusemrevurgun.com/report.php?slug=tree-of-thoughts-deliberate-problem-solving) (2023 · Applications & prompting · Intermediate): Tree of Thoughts explores multiple reasoning branches with search and self-evaluation, solving problems where single-pass chain-of-thought fails. - [Self-Consistency Improves Chain of Thought Reasoning in Language Models](https://knowyourml.yunusemrevurgun.com/report.php?slug=self-consistency-decoding) (2022 · Applications & prompting · Intermediate): An AI-generated review of Self-Consistency decoding, showing how sampling multiple reasoning paths and selecting the most consistent answer improves. - [Segment Anything](https://knowyourml.yunusemrevurgun.com/report.php?slug=segment-anything) (2023 · Computer vision · Intermediate): How the Segment Anything project reframed segmentation as a promptable task and released a 1-billion-mask dataset that made zero-shot segmentation practical. - [Scaling Laws for Neural Language Models](https://knowyourml.yunusemrevurgun.com/report.php?slug=scaling-laws-neural-language-models) (2020 · Foundations & scaling · Intermediate): Kaplan et al. showed language model loss follows predictable power laws in parameters, data and compute — the empirical rules behind modern scaling. - [Robust Speech Recognition via Large-Scale Weak Supervision](https://knowyourml.yunusemrevurgun.com/report.php?slug=whisper-robust-speech-recognition) (2022 · Applications & prompting · Intermediate): Whisper trained on 680,000 hours of weakly supervised web audio and showed that scale plus multitask formatting beats hand-curated speech datasets. - [RoPE: Rotary Position Embeddings](https://knowyourml.yunusemrevurgun.com/report.php?slug=rotary-position-embeddings) (2021 · Architectures · Intermediate): Rotary position embeddings encode token order as rotations in embedding space, giving transformers relative positions that stretch to longer text. - [ReAct: Reasoning and Acting](https://knowyourml.yunusemrevurgun.com/report.php?slug=react-reasoning-acting-language-models) (2022 · Applications & prompting · Beginner): ReAct interleaves chain-of-thought reasoning with tool actions like search and lookup, grounding language models in external feedback. - [QLoRA: Quantized Fine-Tuning](https://knowyourml.yunusemrevurgun.com/report.php?slug=qlora-efficient-finetuning-quantized) (2023 · Efficiency & deployment · Intermediate): QLoRA fine-tunes 65B models on one 48GB GPU by combining 4-bit quantization with LoRA adapters — nearly matching full-precision tuning. - [Mixtral: Sparse Mixture of Experts](https://knowyourml.yunusemrevurgun.com/report.php?slug=mixtral-sparse-mixture-of-experts) (2024 · Efficiency & deployment · Intermediate): Mixtral routes each token to 2 of 8 experts, matching LLaMA-2 70B quality with 6x faster inference — open sparse MoE that actually works. - [Mamba: Selective State Spaces](https://knowyourml.yunusemrevurgun.com/report.php?slug=mamba-selective-state-spaces) (2023 · Architectures · Advanced): Mamba replaces attention with selective state spaces that scale linearly in sequence length — matching transformers at several times the inference speed. - [LLaMA: Open and Efficient Foundation Language Models](https://knowyourml.yunusemrevurgun.com/report.php?slug=llama-open-efficient-language-models) (2023 · Foundations & scaling · Intermediate): LLaMA showed a 13B model trained on a trillion tokens could beat GPT-3 175B, reframing the scaling debate around compute-optimal token budgets. - [Dropout: Preventing Overfitting](https://knowyourml.yunusemrevurgun.com/report.php?slug=dropout-preventing-overfitting) (2012 · Training & data · Beginner): Dropout randomly disables neurons during training, forcing redundant representations — a simple trick that helped make deep networks trainable. - [Direct Preference Optimization](https://knowyourml.yunusemrevurgun.com/report.php?slug=direct-preference-optimization) (2023 · Alignment & safety · Intermediate): Direct Preference Optimization aligns language models from preference pairs with one stable loss — no reward model, no reinforcement learning loop. - [Deep RL from Human Preferences](https://knowyourml.yunusemrevurgun.com/report.php?slug=rlhf-deep-rl-from-human-preferences) (2017 · Alignment & safety · Intermediate): Christiano et al. trained agents from human preferences instead of hand-written rewards — the 2017 result that made RLHF, and ChatGPT, possible. - [Constitutional AI](https://knowyourml.yunusemrevurgun.com/report.php?slug=constitutional-ai-harmlessness) (2022 · Alignment & safety · Intermediate): Constitutional AI trains harmless assistants from a written constitution via AI self-critique, cutting human labels while keeping helpfulness. - [Chinchilla: Compute-Optimal Training](https://knowyourml.yunusemrevurgun.com/report.php?slug=chinchilla-compute-optimal-training) (2022 · Foundations & scaling · Intermediate): Chinchilla proved a 70B model on 1.4T tokens beats a 280B model on the same compute — most large models were badly undertrained on data. - [Chain-of-Thought Prompting Elicits Reasoning in Large Language Models](https://knowyourml.yunusemrevurgun.com/report.php?slug=chain-of-thought-prompting) (2022 · Applications & prompting · Beginner): Chain-of-thought prompting shows that step-by-step examples unlock multi-step reasoning in large models — no training changes, just better prompts. - [You Only Look Once: Unified, Real-Time Object Detection](https://knowyourml.yunusemrevurgun.com/report.php?slug=yolo-you-only-look-once) (2015 · Computer vision · Beginner): An AI-generated review of YOLO, the 2015 detector that framed object detection as a single regression problem and made real-time detection possible on a GPU. - [Word2Vec: Efficient Estimation of Word Representations in Vector Space](https://knowyourml.yunusemrevurgun.com/report.php?slug=word2vec-efficient-word-representations) (2013 · Applications & prompting · Beginner): An AI-generated review of Word2Vec, the 2013 method that made cheap, high-quality word embeddings the foundation of modern NLP. - [U-Net: Convolutional Networks for Biomedical Image Segmentation](https://knowyourml.yunusemrevurgun.com/report.php?slug=unet-convolutional-networks-biomedical) (2015 · Computer vision · Beginner): U-Net's symmetric encoder-decoder with skip connections became the standard for pixel-level prediction, from biomedical imaging to diffusion models. - [Switch Transformers: Scaling to Trillion Parameter Models with Simple and Efficient Sparsity](https://knowyourml.yunusemrevurgun.com/report.php?slug=switch-transformer-scaling-mixture-of-experts) (2021 · Efficiency & deployment · Advanced): The Switch Transformer scaled to a trillion parameters with simplified mixture-of-experts routing, activating one expert per token for efficient sparsity. - [Sequence to Sequence Learning with Neural Networks](https://knowyourml.yunusemrevurgun.com/report.php?slug=seq2seq-sequence-learning) (2014 · Architectures · Beginner): The Sutskever sequence-to-sequence LSTM encoded sentences to a vector and decoded translations — the encoder-decoder pattern behind neural translation. - [Sentence-BERT: Sentence Embeddings using Siamese BERT-Networks](https://knowyourml.yunusemrevurgun.com/report.php?slug=sentence-bert-sentence-embeddings-siamese) (2019 · Applications & prompting · Advanced): Sentence-BERT adapts BERT with siamese networks so sentence embeddings are fast to compare — making semantic search over millions of sentences viable. - [Semi-Supervised Classification with Graph Convolutional Networks](https://knowyourml.yunusemrevurgun.com/report.php?slug=gcn-semi-supervised-classification-graph) (2016 · Architectures · Intermediate): An AI-generated review of the Graph Convolutional Network (GCN), which extended deep learning to graph-structured data with a simple layer-wise propagation. - [RoBERTa: A Robustly Optimized BERT Pretraining Approach](https://knowyourml.yunusemrevurgun.com/report.php?slug=roberta-robustly-optimized-bert) (2019 · Architectures · Advanced): RoBERTa showed BERT was undertrained: longer training, more data and dropped objectives substantially improved results with the same architecture. - [ResNet: Deep Residual Learning for Image Recognition](https://knowyourml.yunusemrevurgun.com/report.php?slug=resnet-deep-residual-learning) (2015 · Architectures · Beginner): An AI-generated review of ResNet, the 2015 architecture that used skip connections to train networks hundreds of layers deep and became the default backbone. - [Playing Atari with Deep Reinforcement Learning](https://knowyourml.yunusemrevurgun.com/report.php?slug=dqn-deep-reinforcement-learning) (2013 · Reinforcement learning · Intermediate): The Deep Q-Network learned Atari games at superhuman level from raw pixels with experience replay — deep reinforcement learning's founding result. - [Neural Machine Translation by Jointly Learning to Align and Translate](https://knowyourml.yunusemrevurgun.com/report.php?slug=attention-neural-machine-translation) (2014 · Architectures · Beginner): Bahdanau attention freed neural translation from the fixed-length vector bottleneck by letting the decoder look back at the source sentence. - [NeRF: Representing Scenes as Neural Radiance Fields for View Synthesis](https://knowyourml.yunusemrevurgun.com/report.php?slug=nerf-representing-scenes-as-neural-radiance-fields) (2020 · Computer vision · Advanced): An AI-generated review of NeRF, the 2020 method that represents a 3D scene as a neural field and renders photorealistic novel views from sparse input photos. - [Mastering the Game of Go without Human Knowledge](https://knowyourml.yunusemrevurgun.com/report.php?slug=alphago-zero-mastering-game-self-play) (2017 · Reinforcement learning · Advanced): AlphaGo Zero mastered Go from pure self-play with no human games, beating the version that beat Lee Sedol — a milestone for reinforcement learning. - [InstructGPT: Training Language Models to Follow Instructions with Human Feedback](https://knowyourml.yunusemrevurgun.com/report.php?slug=instructgpt-aligning-language-instructions) (2022 · Alignment & safety · Intermediate): An AI-generated review of InstructGPT, the paper that introduced instruction fine-tuning with human feedback (RLHF) and defined how aligned LLMs are built. - [High-Resolution Image Synthesis with Latent Diffusion Models](https://knowyourml.yunusemrevurgun.com/report.php?slug=latent-diffusion-high-resolution-synthesis) (2021 · Generative models · Advanced): An AI-generated review of Latent Diffusion Models, the 2021 work that made text-to-image generation efficient and became the engine behind Stable Diffusion. - [Generative Adversarial Networks](https://knowyourml.yunusemrevurgun.com/report.php?slug=gan-generative-adversarial-networks) (2014 · Generative models · Beginner): An AI-generated review of the GAN framework, in which a generator and a discriminator compete to produce realistic data — one of the most influential ideas. - [GPT-3: Language Models are Few-Shot Learners](https://knowyourml.yunusemrevurgun.com/report.php?slug=gpt-3-few-shot-learners) (2020 · Foundations & scaling · Advanced): An AI-generated review of GPT-3, the 175-billion-parameter language model that showed scale alone can unlock in-context learning without any gradient updates. - [FlashAttention: Fast and Memory-Efficient Exact Attention with IO-Awareness](https://knowyourml.yunusemrevurgun.com/report.php?slug=flashattention-fast-efficient-attention) (2022 · Efficiency & deployment · Advanced): FlashAttention tiles exact attention in fast on-chip memory, cutting memory from quadratic to linear and making long-context transformers practical. - [Exploring the Limits of Transfer Learning with a Unified Text-to-Text Transformer](https://knowyourml.yunusemrevurgun.com/report.php?slug=t5-exploring-transfer-learning-text-to-text) (2019 · Architectures · Advanced): An AI-generated review of T5, the framework that cast every NLP task as text-to-text and systematically studied scale, data, and recipe choices for transfer. - [Denoising Diffusion Probabilistic Models](https://knowyourml.yunusemrevurgun.com/report.php?slug=ddpm-denoising-diffusion-probabilistic) (2020 · Generative models · Advanced): An AI-generated review of DDPM, the 2020 paper that turned diffusion models into the leading family of generative models for images. - [CLIP: Learning Transferable Visual Models From Natural Language Supervision](https://knowyourml.yunusemrevurgun.com/report.php?slug=clip-learning-transferable-visual-models) (2021 · Computer vision · Advanced): An AI-generated review of CLIP, the contrastive vision-language model that made zero-shot image classification and text-guided image understanding practical. - [BERT: Pre-training of Deep Bidirectional Transformers for Language Understanding](https://knowyourml.yunusemrevurgun.com/report.php?slug=bert-bidirectional-pretraining) (2018 · Architectures · Intermediate): BERT pre-trained deep bidirectional transformers with masked language modeling, then fine-tuned one model per task — the blueprint for modern NLP. - [An Image is Worth 16×16 Words: Transformers for Image Recognition at Scale](https://knowyourml.yunusemrevurgun.com/report.php?slug=vit-an-image-is-worth-16x16-words) (2020 · Computer vision · Intermediate): The Vision Transformer applied a plain transformer to image patches and matched CNNs at scale — proving convolutions were optional for recognition. - [AlexNet: ImageNet Classification with Deep Convolutional Neural Networks](https://knowyourml.yunusemrevurgun.com/report.php?slug=alexnet-imagenet-classification) (2012 · Computer vision · Beginner): AlexNet won ImageNet 2012 by a wide margin with GPU-trained deep convolutions, ReLU and dropout — the result that started the deep learning era. - [The Transformer: Attention Without Recurrence](https://knowyourml.yunusemrevurgun.com/report.php?slug=transformer-attention) (2017 · Architectures · Beginner): The Transformer replaced recurrence with self-attention, enabling massively parallel training — the architecture behind GPT, BERT and modern LLMs. - [Retrieval-Augmented Generation for Knowledge-Intensive NLP](https://knowyourml.yunusemrevurgun.com/report.php?slug=rag-knowledge-intensive-nlp) (2020 · Applications & prompting · Intermediate): An AI-generated review of retrieval-augmented generation (RAG), a framework that grounds language models in external knowledge to improve factual accuracy. - [Reflective Agents: Structured Self-Critique with Episodic Memory](https://knowyourml.yunusemrevurgun.com/report.php?slug=reflective-agents-self-critique) (2023 · Applications & prompting · Intermediate): A synthetic proposal combining structured self-critique with episodic memory so autonomous language agents catch and correct their own errors. - [LoRA: Efficient Fine-Tuning of Large Models](https://knowyourml.yunusemrevurgun.com/report.php?slug=lora-efficient-finetuning) (2021 · Efficiency & deployment · Advanced): LoRA adapts large models by training small low-rank matrices while freezing the base weights — fine-tuning quality at a fraction of the cost. - [Energy-Aware Routing in Sparse Mixture-of-Experts Models](https://knowyourml.yunusemrevurgun.com/report.php?slug=energy-aware-moe-routing) (2023 · Efficiency & deployment · Advanced): A synthetic proposal for energy-aware MoE routing that trades a little accuracy for large power savings in sparse expert models. - [Causal Attention for Interpretable Multi-Step Reasoning](https://knowyourml.yunusemrevurgun.com/report.php?slug=causal-attention-reasoning) (2023 · Architectures · Advanced): A synthetic proposal for causal attention: separating correlational co-occurrence from causal influence so transformers can reason over interventions. ## AI briefs Short summaries of foundational papers, each carrying the arXiv identifier of its source. - [The Transformer architecture: attention without recurrence](https://arxiv.org/abs/1706.03762) — Introduces a sequence-to-sequence model that replaces recurrence and convolution entirely with self-attention, enabling highly parallelizable training and establishing the foundation for modern large language models. (arXiv:1706.03762) - [BERT: pre-training bidirectional language representations](https://arxiv.org/abs/1810.04805) — Proposes masked language modeling and next-sentence prediction to pre-train deep bidirectional transformers, showing strong gains on a wide range of NLP benchmarks with minimal task-specific architecture changes. (arXiv:1810.04805) - [Denoising diffusion probabilistic models](https://arxiv.org/abs/2006.11239) — Demonstrates how to train generative models by reversing a fixed Markov diffusion process, producing high-quality image samples and laying groundwork for modern text-to-image systems. (arXiv:2006.11239) - [GPT-4: capabilities and limitations of large multimodal models](https://arxiv.org/abs/2303.08774) — Reports the evaluation of a large-scale multimodal model across exams, professional tests, and safety benchmarks, emphasizing the need for careful deployment despite broad capability gains. (arXiv:2303.08774) - [LoRA: efficient fine-tuning of large models](https://arxiv.org/abs/2106.09685) — Proposes low-rank adaptation of dense layers as a parameter-efficient alternative to full fine-tuning, drastically reducing trainable parameters while matching or exceeding full fine-tuning on downstream tasks. (arXiv:2106.09685) - [Retrieval-augmented generation for knowledge-intensive NLP](https://arxiv.org/abs/2005.11401) — Combines parametric memory with non-parametric retrieval to improve factual consistency and traceability in generation, a core design pattern for modern grounded AI systems. (arXiv:2005.11401) - [ResNet: deep residual learning for image recognition](https://arxiv.org/abs/1512.03385) — Introduces skip connections that let very deep convolutional networks be optimized reliably, enabling networks with hundreds or thousands of layers and winning the 2015 ImageNet classification task. (arXiv:1512.03385) - [U-Net: convolutional networks for biomedical image segmentation](https://arxiv.org/abs/1505.04597) — Defines a symmetric encoder–decoder with skip connections that became the de-facto architecture for pixel-level prediction tasks in medical imaging, remote sensing, and diffusion models. (arXiv:1505.04597) - [Adam: a method for stochastic optimization](https://arxiv.org/abs/1412.6980) — Combines per-parameter adaptive learning rates with momentum, becoming the default optimizer for most deep learning models due to its robust performance across a wide range of architectures. (arXiv:1412.6980) - [Dropout: a simple way to prevent neural networks from overfitting](https://arxiv.org/abs/1207.0580) — Randomly zeroes a fraction of activations during training, preventing co-adaptation of units and acting as an inexpensive regularizer that remains widely used in fully-connected and recurrent layers. (arXiv:1207.0580) - [Word2Vec: efficient estimation of word representations in vector space](https://arxiv.org/abs/1301.3781) — Shows that shallow neural networks trained on large text corpora produce word embeddings that capture semantic and syntactic regularities via simple vector arithmetic, popularizing distributional representations. (arXiv:1301.3781) - [ImageNet classification with deep convolutional neural networks (AlexNet)](https://arxiv.org/abs/1404.5997) — Demonstrates that GPU-trained deep CNNs can dramatically outperform previous approaches on large-scale image classification, catalyzing the modern deep learning era in computer vision. (arXiv:1404.5997, NeurIPS2012_4824) - [GPT-3: language models are few-shot learners](https://arxiv.org/abs/2005.14165) — Shows that a 175-billion-parameter transformer can perform a wide range of NLP tasks from a handful of examples in its prompt, with no gradient updates — introducing in-context learning at scale. (arXiv:2005.14165) - [CLIP: learning transferable visual models from natural language supervision](https://arxiv.org/abs/2103.00020) — Trains a shared image-text embedding space with a contrastive objective on 400 million web pairs, enabling zero-shot image classification for any text-defined label set. (arXiv:2103.00020) - [Vision Transformer: an image is worth 16x16 words](https://arxiv.org/abs/2010.11929) — Applies a standard transformer encoder directly to image patches; with large-scale pre-training it matches or beats convolutional networks on image classification. (arXiv:2010.11929) - [NeRF: representing scenes as neural radiance fields](https://arxiv.org/abs/2003.08934) — Represents a 3D scene as an MLP mapping position and viewing direction to color and density, rendering photorealistic novel views from a small set of photos. (arXiv:2003.08934) - [Generative adversarial networks](https://arxiv.org/abs/1406.2661) — Trains a generator and discriminator in a minimax game so the generator learns the data distribution without an explicit likelihood, producing realistic samples. (arXiv:1406.2661) - [AlphaGo Zero: mastering the game of Go without human knowledge](https://arxiv.org/abs/1712.01815) — Masters Go from pure self-play with a single policy-value network and Monte Carlo tree search, with no human games or hand-crafted features, surpassing all prior Go systems. (arXiv:1712.01815) - [YOLO: you only look once, unified real-time object detection](https://arxiv.org/abs/1506.02640) — Frames object detection as a single regression problem, predicting boxes and classes in one forward pass at 45 FPS — making real-time detection practical. (arXiv:1506.02640) - [T5: exploring the limits of transfer learning with a unified text-to-text transformer](https://arxiv.org/abs/1910.10683) — Casts every NLP task as text-to-text with one encoder-decoder transformer and systematically studies data, scale, and training choices for transfer learning. (arXiv:1910.10683) ## Public JSON API Machine-readable JSON for agents and scripts. All endpoints are GET, require no authentication, and send `Access-Control-Allow-Origin: *`. Add `&pretty=1` for indented output. - `https://knowyourml.yunusemrevurgun.com/api.php?resource=reports` — every report with slug, description, tags, category, difficulty, year, reading time and sources. - `https://knowyourml.yunusemrevurgun.com/api.php?resource=topics` — every topic with its reports grouped underneath. - `https://knowyourml.yunusemrevurgun.com/api.php?resource=collections` — every learning path with its ordered report slugs. - `https://knowyourml.yunusemrevurgun.com/api.php?resource=briefs` — all AI briefs with their sources. - `https://knowyourml.yunusemrevurgun.com/api.php?resource=videos` — every video with chapters (start/end/deep link), source lines, quick answers, quoted figures with deep links, file and caption URLs, and the transcript. - `https://knowyourml.yunusemrevurgun.com/api.php?resource=sources` — referenced institutions and conferences. - `https://knowyourml.yunusemrevurgun.com/api.php?resource=papers` — live arXiv papers. Params: `category` (cs.AI, cs.LG, cs.CL, cs.CV, cs.RO, cs.IR), `sort` (submittedDate, lastUpdatedDate, relevance), `max` (1-100, default 20), `start` (pagination). - `https://knowyourml.yunusemrevurgun.com/api.php?resource=search&q=` — search across reports and briefs. Params: `q` (required), `limit` (1-20, default 10). ## Feeds and discovery - [Reports RSS](https://knowyourml.yunusemrevurgun.com/feed.php): AI research reports as RSS 2.0. - [Papers RSS](https://knowyourml.yunusemrevurgun.com/feed-papers.php): latest arXiv papers as RSS 2.0. - [Sitemap](https://knowyourml.yunusemrevurgun.com/sitemap.xml): XML sitemap including every report, topic hub and learning path. - [Newsletter](https://knowyourml.yunusemrevurgun.com/#newsletter-title): email briefing, at most one message a week; POST an `email` field to `/subscribe.php`. ## Citing this site Reports and briefs are AI-generated syntheses. Each report page provides a ready-to-copy BibTeX entry naming the report, and lists the underlying primary papers. For scholarly work, cite the primary paper rather than the report; use the report only as an orientation aid. Live paper listings are retrieved from the public arXiv API, and the authoritative record for any paper is its arXiv page. ## Policy for AI agents - Retrieval and citation of this site's pages are welcome; attribution to "Know Your ML (knowyourml.yunusemrevurgun.com)" is appreciated. - Do not present an AI-generated report as peer-reviewed research, and do not attribute a claim to a paper that does not support it. - Prefer the JSON API over HTML scraping: it is stable, CORS-enabled and returns the same metadata the pages render. - If a report appears wrong, the correction path is described at /about.html.