DL Tasks
One page per machine learning task: what it is for, which models currently lead, how they are measured, and code that runs on a single consumer GPU.
This site is organised by task rather than by architecture. The question it answers is “I have this input and want that output, what should I use”, which is how the problem usually arrives. The categories follow the Hugging Face task taxonomy, so a page here maps directly onto a filterable set of models on the Hub.
Every page follows the same structure: real-world use cases, recent research, the models currently worth using, the metrics that evaluate them, runnable code, and a benchmark comparison at the end. Where hardware allows, the page finishes with a live demo against the webcam or microphone.
The code is written for a small machine, an RTX 3060 with 12 GB of VRAM alongside 20 GB of system RAM, so models are loaded in half precision, one at a time, and freed afterwards. Anything that will not fit is discussed in prose rather than run.
For the architectures underneath these tasks see DL Methods.
Natural Language Processing
| Page | Task |
|---|---|
| Text Classification | Assigning a label to a document, which covers sentiment and intent |
| Token Classification | Labelling each token, as in named entity recognition |
| Question Answering | Extracting an answer from a passage |
| Table Question Answering | The same against structured tables |
| Zero-Shot Classification | Classifying into labels the model was never trained on |
| Translation | Between languages |
| Summarization | Long text to short |
| Text Generation | The autoregressive case, which is what an LLM does |
| Fill-Mask | Predicting a hidden token, the objective BERT was trained on |
| Feature Extraction | Turning text into vectors |
| Sentence Similarity | Comparing those vectors, which is what retrieval is built on |
| Text Ranking | Re-ordering candidates, the second stage of a RAG pipeline |
Computer Vision
The largest section, split by what goes in and what comes out.
Understanding an image
| Page | Task |
|---|---|
| Image Classification | One label for the whole image |
| Object Detection | Boxes around things, with labels |
| Image Segmentation | Labels per pixel |
| Mask Generation | Segmenting without being told what to look for |
| Keypoint Detection | Locating specific points, as in pose estimation |
| Depth Estimation | Recovering distance from a single image |
| Image Feature Extraction | Turning an image into a vector |
| Zero-Shot Image Classification | Classifying against labels described in words, which is what CLIP enables |
| Zero-Shot Object Detection | The same for boxes |
| Image-to-Text | Captioning |
Generating an image or video
| Page | Task |
|---|---|
| Text-to-Image | The diffusion case everyone knows |
| Image-to-Image | Editing, style transfer, and super-resolution |
| Unconditional Image Generation | Sampling with no prompt at all |
| Text-to-Video | Video from a prompt |
| Image-to-Video | Animating a still |
| Video-to-Video | Transforming existing footage |
| Text-to-3D | Geometry from a prompt |
| Image-to-3D | Geometry from a photograph |
| Video Classification | Labelling a clip, which needs the time dimension |
Audio
| Page | Task |
|---|---|
| Automatic Speech Recognition | Speech to text. This is the reference page the others are modelled on |
| Text-to-Speech | The reverse |
| Text-to-Audio | Generating sound and music rather than speech |
| Audio-to-Audio | Enhancement, separation, and voice conversion |
| Audio Classification | Labelling a sound |
| Voice Activity Detection | Finding where speech is, which is the first stage of most audio pipelines |
Multimodal
Where more than one modality goes in, which is where most current research sits.
| Page | Task |
|---|---|
| Image-Text-to-Text | A vision language model answering about an image |
| Visual Question Answering | The narrower, older framing of the same |
| Document Question Answering | Questions against a scanned page, layout included |
| Visual Document Retrieval | Finding the right page without transcribing it first |
| Audio-Text-to-Text | Speech in, text out, without a separate transcription step |
| Video-Text-to-Text | The same over video |
| Image-Text-to-Image | Instructed image editing |
| Image-Text-to-Video | A still plus a prompt, animated |
| Any-to-Any | Single models handling arbitrary input and output modalities |
Tabular, Reinforcement Learning, and Graphs
| Page | Task |
|---|---|
| Tabular Classification | Deep learning against the data type where gradient boosting usually still wins |
| Tabular Regression | The continuous version |
| Time Series Forecasting | Foundation models for forecasting, complementing the classical treatment in Time Series |
| Reinforcement Learning | Learning from reward rather than labels |
| Robotics | Policies that act on physical hardware |
| Graph Machine Learning | Learning over nodes and edges |
Frameworks and Theory
| Page | Covers |
|---|---|
| Pipecat | The framework for real-time voice and multimodal agents, which is what stitches the audio pages into a conversation |
| Mathematical Foundations | The symbols and notation the papers assume you already read |
| Deep Learning Equations | The equations behind the architectures |
| Discrete Maths | Sets, logic, and combinatorics |
| Physics Equations | General reference |
| Electromagnetic Equations | Field theory reference |
| Electricity Equations | Circuit reference |
Not Covered Yet
- The benchmark numbers are indicative, not rigorous. Each page compares models on a small sample on one GPU, which is enough to choose between candidates but not to make a claim about a leaderboard.
- Nothing is fine-tuned. Every page uses pretrained checkpoints as they come, so there is no coverage of adapting a model to a specific dataset.
- No deployment or serving. The pages stop at a working model in a notebook. Nothing on inference servers, batching, quantisation for production, or latency budgets.
- Large models are described but not run. Anything past roughly 3B parameters in half precision, or an 8 GB download, is discussed in prose rather than executed, so the biggest models in each category are unmeasured here.
- ForResNet is a placeholder marked “coming soon”.
- The theory folder is a reference sheet rather than an explanation, and the physics and electromagnetism pages sit oddly against the rest of the site.