fast.ai

fast.ai is a research lab, online course series, and open-source Python deep learning library founded by Jeremy Howard (former president and chief scientist of Kaggle, founder of FastMail and Optimal Decisions Group) and Rachel Thomas (UCSF / University of San Francisco data ics researcher), buil…

Semantic Classification

Content

Compositional Relationships (Components)

SubClassOf(ai:FastAI
  ObjectSomeValuesFrom(ai:hasPart ai:FastaiLibrary))
SubClassOf(ai:FastAI
  ObjectSomeValuesFrom(ai:hasPart ai:PracticalDeepLearningCourse))
SubClassOf(ai:FastAI
  ObjectSomeValuesFrom(ai:hasPart ai:Nbdev))
SubClassOf(ai:FastAI
  ObjectSomeValuesFrom(ai:hasPart ai:LearnerAbstraction))
SubClassOf(ai:FastAI
  ObjectSomeValuesFrom(ai:hasPart ai:DataBlockAPI))
SubClassOf(ai:FastAI
  ObjectSomeValuesFrom(ai:hasPart ai:CallbackSystem))
SubClassOf(ai:FastAI
  ObjectSomeValuesFrom(ai:hasPart ai:ULMFiT))
SubClassOf(ai:FastAI
  ObjectSomeValuesFrom(ai:hasPart ai:AnswerAI))

## Dependency Relationships
SubClassOf(ai:FastAI
  ObjectSomeValuesFrom(ai:requires ai:PyTorch))
SubClassOf(ai:FastAI
  ObjectSomeValuesFrom(ai:requires ai:Python))
SubClassOf(ai:FastAI
  ObjectSomeValuesFrom(ai:requires ai:JupyterNotebook))
SubClassOf(ai:FastAI
  ObjectSomeValuesFrom(ai:requires ai:CUDAGPU))
SubClassOf(ai:FastAI
  ObjectSomeValuesFrom(ai:requires ai:TransferLearning))
SubClassOf(ai:FastAI
  ObjectSomeValuesFrom(ai:dependsOn ai:NumPy))
SubClassOf(ai:FastAI
  ObjectSomeValuesFrom(ai:dependsOn ai:Pandas))
SubClassOf(ai:FastAI
  ObjectSomeValuesFrom(ai:dependsOn ai:Fastcore))
SubClassOf(ai:FastAI
  ObjectSomeValuesFrom(ai:dependsOn ai:Fastprogress))
SubClassOf(ai:FastAI
  ObjectSomeValuesFrom(ai:dependsOn ai:StochasticGradientDescent))

## Capability Relationships
SubClassOf(ai:FastAI
  ObjectSomeValuesFrom(ai:enables ai:RapidPrototyping))
SubClassOf(ai:FastAI
  ObjectSomeValuesFrom(ai:enables ai:TransferLearningWorkflows))
SubClassOf(ai:FastAI
  ObjectSomeValuesFrom(ai:enables ai:DemocratisedDeepLearning))
SubClassOf(ai:FastAI
  ObjectSomeValuesFrom(ai:enables ai:CodeFirstEducation))
SubClassOf(ai:FastAI
  ObjectSomeValuesFrom(ai:enables ai:ProductionReadyDefaults))
SubClassOf(ai:FastAI
  ObjectSomeValuesFrom(ai:supports ai:ComputerVision))
SubClassOf(ai:FastAI
  ObjectSomeValuesFrom(ai:supports ai:NaturalLanguageProcessing))
SubClassOf(ai:FastAI
  ObjectSomeValuesFrom(ai:supports ai:TabularLearning))
SubClassOf(ai:FastAI
  ObjectSomeValuesFrom(ai:supports ai:CollaborativeFiltering))
SubClassOf(ai:FastAI
  ObjectSomeValuesFrom(ai:supports ai:MedicalImaging))
SubClassOf(ai:FastAI
  ObjectSomeValuesFrom(ai:supports ai:StableDiffusionTraining))

## Implementation Relationships
SubClassOf(ai:FastAI
  ObjectSomeValuesFrom(ai:implements ai:OneCycleLearningRateSchedule))
SubClassOf(ai:FastAI
  ObjectSomeValuesFrom(ai:implements ai:DiscriminativeLearningRates))
SubClassOf(ai:FastAI
  ObjectSomeValuesFrom(ai:implements ai:FrozenUnfrozenFineTuning))
SubClassOf(ai:FastAI
  ObjectSomeValuesFrom(ai:implements ai:MixupAugmentation))
SubClassOf(ai:FastAI
  ObjectSomeValuesFrom(ai:implements ai:TestTimeAugmentation))
SubClassOf(ai:FastAI
  ObjectSomeValuesFrom(ai:implements ai:LabelSmoothing))
SubClassOf(ai:FastAI
  ObjectSomeValuesFrom(ai:implements ai:LayeredAPIDesign))
SubClassOf(ai:FastAI
  ObjectSomeValuesFrom(ai:uses ai:ConvolutionalNeuralNetworks))
SubClassOf(ai:FastAI
  ObjectSomeValuesFrom(ai:uses ai:TransformerArchitecture))
SubClassOf(ai:FastAI
  ObjectSomeValuesFrom(ai:uses ai:ResNet))
SubClassOf(ai:FastAI
  ObjectSomeValuesFrom(ai:uses ai:AdamOptimiser))
SubClassOf(ai:FastAI
  ObjectSomeValuesFrom(ai:uses ai:CosineAnnealing))

## Reduction Relationships
SubClassOf(ai:FastAI
  ObjectSomeValuesFrom(ai:reduces ai:DeepLearningLearningCurve))
SubClassOf(ai:FastAI
  ObjectSomeValuesFrom(ai:reduces ai:BoilerplateCode))
SubClassOf(ai:FastAI
  ObjectSomeValuesFrom(ai:reduces ai:TrainingTime))
SubClassOf(ai:FastAI
  ObjectSomeValuesFrom(ai:reduces ai:HyperparameterTuningEffort))
SubClassOf(ai:FastAI
  ObjectSomeValuesFrom(ai:reduces ai:DataPipelineComplexity))

## Association Relationships
SubClassOf(ai:FastAI
  ObjectSomeValuesFrom(ai:relatedTo ai:HuggingFaceAccelerate))
SubClassOf(ai:FastAI
  ObjectSomeValuesFrom(ai:relatedTo ai:BERT))
SubClassOf(ai:FastAI
  ObjectSomeValuesFrom(ai:relatedTo ai:FSDPQLoRA))
SubClassOf(ai:FastAI
  ObjectSomeValuesFrom(ai:relatedTo ai:StableDiffusion))
SubClassOf(ai:FastAI
  ObjectSomeValuesFrom(ai:contrastsWith ai:Keras))
SubClassOf(ai:FastAI
  ObjectSomeValuesFrom(ai:contrastsWith ai:PyTorchLightning))
SubClassOf(ai:FastAI
  ObjectSomeValuesFrom(ai:contrastsWith ai:HuggingFaceTransformers))
SubClassOf(ai:FastAI
  ObjectSomeValuesFrom(ai:contrastsWith ai:RawPyTorch))
SubClassOf(ai:FastAI
  ObjectSomeValuesFrom(ai:contrastsWith ai:TensorFlow))

## Data Properties (Characteristics)
DataPropertyAssertion(ai:hasIdentifier ai:FastAI "AI-1078"^^xsd:string)
DataPropertyAssertion(ai:authorityScore ai:FastAI "0.87"^^xsd:decimal)
DataPropertyAssertion(ai:foundationYear ai:FastAI "2016"^^xsd:integer)
DataPropertyAssertion(ai:githubStars ai:FastAI "26000"^^xsd:integer)
DataPropertyAssertion(ai:courseLearners ai:FastAI "500000"^^xsd:integer)
DataPropertyAssertion(ai:libraryMajorVersion ai:FastAI "2"^^xsd:integer)
DataPropertyAssertion(ai:nbdevAdoptingProjects ai:FastAI "10000"^^xsd:integer)
DataPropertyAssertion(ai:answerAIFoundationYear ai:FastAI "2024"^^xsd:integer)

## Property Constraints
SubClassOf(ai:FastAI
  DataMinCardinality(1 ai:hasBackend xsd:string))
SubClassOf(ai:FastAI
  DataAllValuesFrom(ai:isOpenSource xsd:boolean))
SubClassOf(ai:FastAI
  DataSomeValuesFrom(ai:pythonVersion xsd:string))
SubClassOf(ai:FastAI
  DataMinCardinality(1 ai:hasCourseEdition xsd:integer))

## Annotations
AnnotationAssertion(rdfs:label ai:FastAI "fast.ai"@en)
AnnotationAssertion(rdfs:comment ai:FastAI "Research lab, online course series, and open-source PyTorch-based Python deep learning library founded in 2016 by Jeremy Howard and Rachel Thomas with the mission of making deep learning practically accessible to working coders without requiring graduate-level mathematics. Comprises the fastai library (layered API with Learner abstraction, DataBlock API, Callback system, 1cycle/discriminative-LR/frozen-unfrozen training defaults), the Practical Deep Learning for Coders MOOC (500,000+ learners across v1-v4 editions 2016-2022), the Computational Linear Algebra and Practical Data Ethics courses, the From Deep Learning Foundations to Stable Diffusion course (2022), and the nbdev literate-programming framework. Originator of ULMFiT (Howard & Ruder 2018) directly preceding BERT/GPT transfer-learning era. Howard co-founded Answer.AI in March 2024, publishing FSDP-QLoRA enabling 70B-parameter LLM fine-tuning on consumer GPUs, plus Claudette and FastHTML. Influence visible in Hugging Face Accelerate, PyTorch Lightning, and modern PyTorch idioms."@en)
AnnotationAssertion(dcterms:identifier ai:FastAI "AI-1078"^^xsd:string)
AnnotationAssertion(dcterms:subject ai:FastAI "Deep Learning Education, PyTorch Ecosystem, Transfer Learning, Open Source ML, Applied AI Research"@en)

)

Property Characteristics

AsymmetricObjectProperty(ai:requires) AsymmetricObjectProperty(ai:enables) AsymmetricObjectProperty(ai:implements) AsymmetricObjectProperty(ai:contrastsWith) TransitiveObjectProperty(ai:dependsOn) FunctionalDataProperty(ai:foundationYear) FunctionalDataProperty(ai:answerAIFoundationYear)

About fast.ai

• fast.ai is simultaneously a research lab, an open-source Python deep learning library, an online course series, and a sustained pedagogical experiment in democratising machine learning. Founded in late 2016 by Jeremy Howard (former president and chief scientist of Kaggle, founder of email service FastMail and insurance startup Optimal Decisions Group) and Rachel Thomas (data ethics researcher at the University of California San Francisco and later founding director of the USF Center for Applied Data Ethics), fast.ai emerged from Howard’s conviction—formed during his Kaggle leadership and his earlier 25-year career applying machine learning to medicine, insurance and consulting—that the established educational pipeline for deep learning was unnecessarily mathematics-heavy and was excluding precisely the practitioners (medics, biologists, lawyers, journalists, school teachers, domain experts) who could most benefit from applying it to their fields.
• The fast.ai mission statement, articulated in dozens of Howard and Thomas talks, blog posts and the foreword to Deep Learning for Coders with fastai and PyTorch (Howard & Gugger, O’Reilly 2020), is to make neural networks uncool again—to remove the artificial gatekeeping around deep learning and treat it as another piece of practical software engineering accessible to anyone with high-school maths and a year of Python. This mission is implemented through three intertwined deliverables that are deliberately developed together: a software library, a free online course, and a research output stream feeding back into both.
• The most influential research artefact to come out of the project is the ULMFiT paper (Howard & Ruder, “Universal Language Model Fine-tuning for Text Classification”, ACL 2018), which demonstrated that transfer learning—the technique that had powered the Computer Vision revolution since AlexNet 2012 and ImageNet-pretrained ResNets—could be applied with equal force to Natural Language Processing. ULMFiT introduced a three-stage protocol (general-domain language model pretraining on Wikitext-103, target-domain LM fine-tuning, classifier head fine-tuning) using discriminative learning rates and slanted-triangular schedules, and achieved 18-24% error reduction on six text classification benchmarks. ULMFiT directly preceded and influenced the BERT (Devlin et al. 2018), GPT-1 (Radford et al. 2018) and subsequent transformer-based language model lineage that would dominate NLP from 2018 onwards. The paper has accumulated 4,000+ citations and is recognised as a foundational moment in NLP transfer learning.
• In March 2024, Howard co-founded Answer.AI with Eric Ries (author of The Lean Startup) as a fully-distributed end-to-end AI R&D lab structured as a public benefit corporation. Within months of launch Answer.AI published the FSDP-QLoRA technique enabling 70-billion-parameter LLM fine-tuning on consumer-grade dual-RTX-3090/4090 GPUs (~£3,500 hardware budget) by combining PyTorch FSDP (Fully Sharded Data Parallel) with bitsandbytes 4-bit QLoRA quantisation—work co-authored with Tim Dettmers and the Hugging Face team. Answer.AI’s subsequent releases include Claudette (an idiomatic Python interface to Anthropic Claude with literate-programming source notebooks), FastHTML (a Python-only web application framework eliminating JavaScript through HTMX and HyperScript), the cosette OpenAI client, and the fasthtml-example gallery.

Founding Timeline and Organisational Structure

Pre-history (2014-2016): Howard finishes his term as president and chief scientist of Kaggle (2011-2013), takes a research appointment at the Singularity University, and spends 2014-2016 founding and exiting the medical AI startup Enlitic. Thomas, a Duke mathematics PhD, transitions from Uber data science to academic research and writing on bias, surveillance and the social effects of AI. The two meet through Howard’s USF deep-learning seminar and find common ground on the pedagogical-democratisation problem.

2016 — Foundation and v1 Course: First Practical Deep Learning for Coders course launched as a USF Data Institute professional certificate, with seven three-hour video lectures recorded in front of a live cohort. Open-sourced to YouTube within months. Library v0.7 built on Keras + TensorFlow with custom training-loop wrappers. ~5,000 learners in the first year.

2017 — Cutting Edge Deep Learning: Second seven-lecture course covering RNNs, attention, Neural Machine Translation, segmentation, super-resolution, GANs, and meta-learning. Library migrated to PyTorch 0.3 reflecting Howard’s growing dissatisfaction with TensorFlow’s symbolic graphs.

2018 — ULMFiT and v3 Preparation: Howard & Ruder publish ULMFiT at ACL 2018. Library development paused for full architectural rewrite. fastai v1.0 released October 2018 coinciding with PyTorch 1.0 launch (the two were developed in coordination, with Howard giving the joint keynote at the PyTorch 1.0 conference).

2019 — v3 Course Global Phenomenon: Practical Deep Learning for Coders v3 released January 2019 across seven lectures covering vision, NLP, tabular, collaborative filtering, segmentation and the new fastai v1 API. The course was extensively translated (Spanish, Chinese, Portuguese, Korean) and adopted as the primary deep-learning curriculum by hundreds of university courses globally. 400,000+ learners by end of year.

2020 — fastai v2 and the O’Reilly Book: Howard & Gugger publish fastai: A Layered API for Deep Learning in the Information journal (August 2020) accompanied by the fastai v2 release, a complete library rewrite around the layered API design. The Deep Learning for Coders with fastai and PyTorch O’Reilly book (Howard & Gugger 2020) becomes a bestseller and the canonical fastai reference. Rachel Thomas launches Practical Data Ethics as a standalone course.

2022 — Ground-Up Rewrite and Stable Diffusion Course: Practical Deep Learning v4 released as a full ground-up rewrite using the v2 API. Months later the From Deep Learning Foundations to Stable Diffusion course launches, co-taught with Jonathan Whitaker and Tanishq Mathew Abraham, deriving diffusion models from first principles starting with basic neural-network training loops.

2024 — Answer.AI and FSDP-QLoRA: Howard + Eric Ries found Answer.AI in March 2024 as a fully-distributed public benefit corporation. FSDP-QLoRA released within months, followed by Claudette, FastHTML, and a continuous stream of Practical R&D technical blog posts.

2025-2026 — Ongoing: fastai library at v2.8.x (v2.8.7 as of February 2026) in active maintenance; Howard divides time between fast.ai education and Answer.AI research; Sylvain Gugger departed Hugging Face in August 2023 and is now a Machine Learning Engineer at Jane Street, though continuing open-source contributions to Transformers and Accelerate; Tanishq Abraham departed Stability AI in January 2025 and is now CEO and co-founder of Sophont, a medical AI startup building multimodal foundation models for medicine (seed funding: $9.2M).

Components / Architecture

The fastai Library: Layered API Design

The fastai library is structured according to the layered API principle articulated in the Howard-Gugger 2020 paper, exposing three distinct levels of abstraction simultaneously so that a beginner can train a state-of-the-art model in five lines whilst a researcher can drop down to direct tensor manipulation when needed.

High-Level API: The Learner Abstraction

At the highest level the user constructs a Learner object encapsulating model, data, loss function, optimiser, metrics and training callbacks. For the canonical five-line image classifier:

from fastai.vision.all import *
path = untar_data(URLs.PETS)
dls = ImageDataLoaders.from_name_func(path, get_image_files(path/'images'),
                                       valid_pct=0.2, seed=42,
                                       label_func=lambda f: f[0].isupper(),
                                       item_tfms=Resize(224))
learn = vision_learner(dls, resnet34, metrics=error_rate)
learn.fine_tune(1)

This five-line snippet trains a ResNet-34 on the Oxford Pets dataset to ~95% accuracy in under two minutes on a single GPU, using transfer learning with all opinionated defaults applied: 1cycle LR schedule, discriminative learning rates between backbone and head, freezing the pretrained weights for the first epoch and unfreezing for fine-tuning, sensible Mixup-free augmentation, label smoothing, and FP16 mixed-precision if a GPU supports it.

Mid-Level API: DataBlock, Callbacks, Optimisers, Metrics

Beneath the high-level wrappers, the DataBlock API provides a declarative pipeline-construction language separating the what (block types: ImageBlock, CategoryBlock, MultiCategoryBlock, RegressionBlock, MaskBlock, BBoxBlock, TextBlock, TabularBlock) from the how (get_items, splitter, get_x, get_y, item_tfms, batch_tfms). A single DataBlock definition can produce DataLoaders for any of vision, text, tabular, audio, collaborative filtering, point-cloud, or medical-imaging domains.

The Callback system is the engineering core of the library. The training loop emits 14 events (before_fit, before_epoch, before_train, before_batch, after_pred, after_loss, before_backward, after_backward, before_step, after_step, after_batch, after_train, after_validate, after_epoch) and callbacks subscribe by overriding the corresponding methods. The 1cycle schedule, mixed precision, gradient accumulation, distributed training, hooks, EarlyStopping, model checkpointing, and TensorBoard/W&B logging are all implemented as callbacks composable in any order.

The Optimiser module decouples step computation from gradient processing through a stat-and-step decomposition, enabling Adam, AdamW, RAdam, LAMB, Ranger, RangerLars, QHAdam and Look-Ahead optimisers to share implementation rather than each maintaining its own training loop.

Low-Level API: Tensor Type Dispatch and fastcore

At the lowest level fastai uses Python type dispatch (implemented in the companion fastcore library) to extend torch.Tensor with semantic subclasses (TensorImage, TensorMask, TensorBBox, TensorCategory, TensorText, TensorAudio). Augmentations and transforms are dispatched on tensor type, so the same flip_lr transform correctly flips an image, mirrors a segmentation mask, and reflects bounding-box coordinates without user intervention. This unification eliminates the boilerplate that plagues raw-PyTorch vision pipelines.

Opinionated Training Defaults

fast.ai’s distinctive contribution to applied deep learning is the systematic codification of opinionated production defaults that consistently outperform raw PyTorch training loops in head-to-head benchmarks. These defaults are:

1cycle Learning Rate Schedule (Smith 2018, A disciplined approach to neural network hyper-parameters): A two-phase schedule that anneals from lr_max/div_factor to lr_max over the first 25-50% of training, then anneals from lr_max to lr_max/final_div_factor over the remainder, with optional momentum cycling inversely correlated to learning rate. Empirically converges 3-10× faster than constant LR with comparable or better final accuracy.

Discriminative Learning Rates: Different LRs applied to different layer groups, with deeper (closer-to-input) layers receiving 10-100× smaller LRs than the classification head. Reflects the empirical observation that pretrained early features are more universal and require less perturbation than task-specific later features.

Frozen-then-Unfrozen Fine-Tuning: Initial epoch trains only the freshly-initialised head with the backbone frozen, then learn.unfreeze() releases all layers for end-to-end training with discriminative rates. This staged approach prevents the head’s random gradients from destroying pretrained features.

Mixup Augmentation (Zhang et al. 2018): Convex combinations of input pairs with corresponding label interpolation, reducing memorisation and improving generalisation.

Test-Time Augmentation (TTA): At inference, average predictions over 5-10 augmented versions of each test image, typically improving accuracy by 0.5-2 percentage points.

Label Smoothing (Szegedy et al. 2016): Replace one-hot targets with smoothed distributions (e.g. 0.9 for true class, 0.1/(C-1) elsewhere), preventing overconfidence and improving calibration.

Progressive Resizing: Train initially at smaller resolutions (128² or 192²) then resume at larger (224² or 384²), reducing wall-clock training time by 30-50% with comparable final accuracy.

nbdev: Literate Programming for Python

nbdev (Howard 2019, with significant contributions from Hamel Husain) is the literate-programming framework underlying the fastai library itself. It inverts the traditional Python development model: instead of writing .py modules and then writing notebook tutorials demonstrating them, the Jupyter notebook is the source of truth and .py modules are generated by export. A single notebook simultaneously holds the implementation code, the tests (interleaved as notebook cells), the prose documentation (markdown cells), and the worked examples (output cells preserved in CI). nbdev tools generate documentation websites, run tests, manage versioning, and synchronise the notebook ↔ module bidirectionally.

nbdev has been adopted by approximately 10,000+ open-source projects beyond fastai itself, including the Answer.AI library suite (Claudette, FastHTML, cosette), several Hugging Face research projects, and university teaching toolkits.

Use Cases / Major Families

1. Educational Curriculum and Free MOOCs

fast.ai’s free Practical Deep Learning courses constitute the most widely-deployed deep-learning curriculum globally outside the major university platforms (Stanford CS231n, Andrew Ng’s Coursera Deep Learning Specialization, MIT 6.S191). Cumulative enrolment across v1-v4 editions exceeds 500,000 unique learners as of 2024. The course is taught in person or as supplementary curriculum at Stanford, MIT, USF, Queensland, University of Edinburgh, Imperial College London, and approximately 200+ other universities globally. Industrial onboarding programmes at DeepMind, OpenAI, Anthropic, Hugging Face, NVIDIA, Google Brain alumni and various FAANG ML teams cite the fast.ai courses as recommended pre-onboarding material.

The pedagogical method itself constitutes a major use case: educators rebuilding curricula around the top-down approach (start by training a working model; introduce theory progressively) report dramatically higher completion rates than bottom-up alternatives. Howard’s 2018 fast.ai’s pedagogy essay codified the approach into eight principles (immediate practical engagement; meaningful contexts; whole-game first; just-in-time theory; growth mindset framing; mistakes celebrated; deliberate emphasis on intuition; community-driven learning). These principles have informed curriculum design at the Allen Institute for AI (AI2 Tango), Hugging Face Course, Google Cloud Skills Boost, and various national digital-skills initiatives including UK Aire Skills, Bristol Robotics Lab outreach, and the AI-for-All initiative.

2. Rapid Production Prototyping

Enterprise ML teams use fastai as a rapid-iteration layer atop PyTorch for prototyping classifiers, fine-tuning vision and NLP models, and establishing baselines before moving to bespoke training infrastructure. Documented production deployments include NASA (satellite imagery classification for Earth observation programmes), Lyft (rider intent classification reducing support escalations), Goldman Sachs (document classification across financial filings and contracts), IBM Watson, Carbon Health (clinical notes triage), Splunk (security log anomaly detection), and dozens of medium-sized firms across radiology, agriculture, retail demand forecasting and content moderation. The typical prototyping workflow involves a 1-2 week fastai notebook establishing baseline performance against a target metric, after which the team either deploys the fastai pipeline directly (~40% of cases) or migrates to a bespoke PyTorch training infrastructure with the fastai-validated hyperparameters and training recipe (~60% of cases).

Industry surveys conducted by Anyscale, Weights & Biases and Linux Foundation AI in 2022-2024 consistently rank fastai amongst the top-three “high-level training libraries used in production” alongside Hugging Face Transformers and PyTorch Lightning, with particular strength in computer vision, tabular learning, and rapid prototyping segments.

3. Medical Imaging

The fastai vision API and its integration with the MONAI ecosystem make it a frequent choice for academic medical imaging research. Tanishq Abraham (formerly Stability AI Research Director 2023-2025, now CEO of Sophont) led several fastai-based medical imaging projects whilst at fast.ai. Documented deployments include diabetic retinopathy screening (Aravind Eye Hospitals India, 100,000+ patient screenings annually), dermatology classification (Stanford collaboration achieving dermatologist-level accuracy on HAM10000 benchmark), pathology slide analysis (multiple histopathology research groups), and the influential Covid-19 chest-X-ray classification work conducted in early 2020 during the pandemic onset that contributed to the COVID-Net open-source family of models.

The Aravind deployment exemplifies the fast.ai mission: a non-profit Indian eye hospital network used fastai-trained models to extend diabetic retinopathy screening to rural clinics that lack on-site ophthalmologists, with the screening models achieving sensitivity and specificity comparable to published Google Health benchmarks but trained on a fraction of the data and compute. The deployment was developed in collaboration with the fast.ai community and is documented in the Aravind Eye Care System / fast.ai joint case study.

Kheiron Medical Technologies (London), founded by individuals with fast.ai community engagement, deploys breast cancer screening AI across NHS trusts using a PyTorch training stack with fastai-derived methodology. Faculty AI (London) and Cogstack (NHS digital) similarly draw on fast.ai-trained engineers for rapid medical AI prototyping.

4. Tabular and Collaborative Filtering

The fastai tabular_learner and collab_learner provide ergonomic wrappers over PyTorch tabular embeddings (categorical → learned-embedding lookup → MLP) and dot-product collaborative filtering with bias terms and embedding regularisation. These have been used in production at several recommender-systems shops, in the Practical Deep Learning Lesson 4 (Tabular) example that informed UpStart’s early credit-scoring pipeline, and as the conceptual basis for the embedding-based tabular approaches now widespread at fintech firms.

The fastai tabular treatment had broader impact than commonly recognised: by demonstrating that learned embeddings beat one-hot encoding for categorical features in tabular deep learning (a result formalised in Cheng et al. 2016 Wide & Deep and Guo & Berkhahn 2016 Entity Embeddings, both extensively taught in fastai courses), fast.ai accelerated the adoption of neural tabular models at firms previously reliant on gradient-boosted trees (XGBoost, LightGBM). As of 2026 hybrid GBT+embedding models are common at Stripe, Wise, Monzo, Revolut, Plaid and several London/Manchester fintechs, with fast.ai course materials cited as the canonical introduction.

5. Stable Diffusion and Generative Models

The 2022 From Deep Learning Foundations to Stable Diffusion course generated significant follow-on community work in diffusion-model research, with course graduates contributing to Hugging Face Diffusers, Stability AI training pipelines, and several open-source text-to-image projects. Tanishq Abraham’s course chapters covering CLIP, U-Nets and the variational lower bound serve as a widely-cited pedagogical reference, with the course’s notebook material being one of the most-bookmarked diffusion learning resources on GitHub (3,500+ stars on the fastai/diffusion-nbs repository).

Jonathan Whitaker (Hugging Face) and Tanishq Abraham continued co-teaching follow-on workshops at NeurIPS, MLSS Africa, and the EleutherAI study group throughout 2023-2025, with the course’s “build a diffusion model from scratch” methodology adopted by multiple subsequent educators (Andrej Karpathy’s Zero to Hero, Jay Alammar’s Illustrated Diffusion, the Hugging Face Diffusion Models Class).

6. ULMFiT and the Transfer-Learning-for-NLP Lineage

The ULMFiT protocol is no longer state-of-the-art (superseded by BERT/GPT-style pretraining at much larger scale) but its three-stage transfer learning recipe (general LM → target LM → classifier) remains the standard template for adapting modern large language models to downstream tasks, with HuggingFace Trainer-based fine-tuning following the same conceptual structure. The ULMFiT hyperparameters (slanted triangular LR schedule, discriminative learning rates per layer group, gradual unfreezing) translate directly to modern LoRA and full fine-tuning workflows.

Sebastian Ruder (DeepMind, formerly aylien, co-author of ULMFiT) maintains the NLP News newsletter and is among the most influential public voices on transfer-learning history; Ruder’s blog posts repeatedly cite ULMFiT as the moment NLP transfer learning became practical at scale, preceding the BERT/GPT explosion of late 2018-2019 by a critical six-month window.

7. Agriculture, Climate, and Earth Observation

fast.ai’s accessible vision API has been adopted by climate-tech and agriculture startups for satellite imagery and drone-photography analysis. Documented use cases include crop disease classification (PlantVillage and successor projects), forest carbon stock estimation, wildlife population surveys from drone imagery, and the Climate Change AI initiative’s tutorial materials which use fastai for hands-on labs. UK-based climate-tech firms (Cervest, Mitiga Solutions, ClimateScore) reportedly use fastai for prototype work though typically migrate to bespoke pipelines for production.

8. Reinforcement Learning and Robotics Adjacency

fast.ai itself does not provide a deep reinforcement learning API (Howard has consistently stated RL is outside the scope of Practical Deep Learning), but the underlying fastai pipeline (Learner, callbacks, mixed precision) is used as a perception backbone within RL projects at several research labs. Wayve (London autonomous driving) has cited fast.ai materials as a learning resource for new hires working on end-to-end driving models. Sense Robotics, Five AI (acquired by Bosch), and Oxbotica (Oxford) have similar associations.

Academic Context

fast.ai occupies an unusual position in the academic ecosystem: Howard holds the Distinguished Research Scientist title at USF (formerly Singularity University), has published in ACL, Information, and several deep-learning workshops, but does not maintain a traditional academic appointment or publish at the volume of full-time university researchers. Thomas published primarily at the Practical Data Ethics / digital-civil-liberties intersection until stepping back from public-facing research in 2023.

Despite this non-traditional posture, fast.ai has been deeply influential on mainstream academic deep learning:

• The ULMFiT paper (Howard & Ruder 2018) is one of the most-cited NLP transfer-learning papers, directly preceding the BERT/GPT lineage and cited as inspiration by both Devlin et al. (BERT) and Peters et al. (ELMo).

• The fastai v2 paper (Howard & Gugger 2020) in the Information journal articulates the layered-API design philosophy now adopted across the modern ML framework ecosystem.

• Course materials are used as teaching reference at hundreds of universities globally, with the fastai library installed in millions of academic Colab notebooks.

• The Computational Linear Algebra course (Rachel Thomas 2017) using PyTorch and notebooks rather than textbook proofs has been adopted as supplementary curriculum by numerical-methods courses at several universities.

• Sylvain Gugger (fast.ai 2018-2021, Hugging Face 2021-2023) was instrumental in building Hugging Face Accelerate as technical lead; he departed Hugging Face in August 2023 and is now a Machine Learning Engineer at Jane Street, while continuing open-source contributions to Accelerate and Transformers — the de facto industry-standard libraries for distributed PyTorch training that carry forward the fastai callback-system DNA.

  Howard’s most cited published work alongside ULMFiT is the Smith 2018 Cyclical Learning Rates / 1cycle policy (Leslie Smith’s papers were heavily promoted and validated empirically through the fastai courses), the Mixup paper (Zhang, Cisse, Dauphin, Lopez-Paz 2018, picked up and popularised by fastai), and various course technical notes published as Practical Deep Learning lecture supplements.

  In 2021 the fast.ai approach attracted academic critique from researchers preferring more theoretically-grounded curricula; the response from Howard and the community has consistently been to evaluate pedagogy on outcomes (graduates’ ability to deliver working systems) rather than coverage of underlying mathematics, a position validated by graduate-employment statistics and the demonstrated career trajectories of well-known alumni (Radek Osmulski → NVIDIA, Hamel Husain → GitHub → independent consultant, Tanishq Abraham → Stability AI Research Director → Sophont CEO, Jonathan Whitaker → Hugging Face).

Current Landscape (2026)

As of mid-2026 the fast.ai project comprises three concurrent activities in active operation:

1. The fastai Library v2.8.x: maintained primarily by Jeremy Howard with community contributions (v2.8.7 released February 2026). Pace of releases has slowed from the v2.0-v2.3 rush (2020-2022) into a maintenance posture as the library has stabilised. PyTorch 2.0+ compatibility achieved early 2023. The library is in production use across an estimated ~50,000-100,000 organisations globally based on PyPI download statistics and GitHub dependents (3,500+; 27,600+ stars as of 2026).

2. The Course Programme: Practical Deep Learning for Coders 2022 (v4) and From Deep Learning Foundations to Stable Diffusion (2022) remain freely available. No major new course release as of mid-2026, although Howard has indicated a 2026 refresh covering modern LLM fine-tuning techniques is in preparation. The 2022 cohort surveys show ~75,000 active enrolments year-on-year.

3. Answer.AI (March 2024-present): A fully-distributed AI R&D public benefit corporation co-founded by Howard and Eric Ries with seed-stage investors. Key 2024-2026 publications and software releases:

• FSDP-QLoRA (March 2024) — 70B-parameter LLM fine-tuning on consumer GPUs

• Claudette (June 2024) — idiomatic Python interface to Anthropic Claude

• FastHTML (August 2024) — Python-only web framework via HTMX/HyperScript

• cosette (2024) — OpenAI client mirror of Claudette

• fastlite (2024) — lightweight SQLite ORM

• monsterui (2025) — UI component library for FastHTML

• Solveit (October 2025) — new course teaching practical problem-solving using AI-assisted iteration for coding, writing, sysadmin, and research tasks

• Continuing Practical R&D blog series with technical write-ups on agentic coding, model fine-tuning, retrieval-augmented generation, and small-team ML R&D practice

  The broader fast.ai community runs ~60 active study groups globally (London, Manchester, Bristol, Edinburgh, Cambridge, San Francisco, New York, Toronto, Bangalore, Singapore, Sydney, Buenos Aires, Lagos and dozens more), supported by the fast.ai Discord (~12,000 members) and forum (60,000 registered users).

  Influence on the wider ecosystem (2025-2026):

• Hugging Face Accelerate — Sylvain Gugger (who departed Hugging Face in August 2023 and is now at Jane Street) built Accelerate as technical lead, carrying forward the fastai callback DNA; Accelerate remains the de facto distributed-training standard across Hugging Face Transformers, Diffusers, TRL and SFT-Trainer.

• PyTorch Lightning explicitly cites the fastai Callback design as inspiration in its v2.0 documentation.

• The Hugging Face Trainer API mirrors the Learner abstraction one-to-one.

• Modern PyTorch idioms (mixed precision, gradient accumulation, distributed training as composable callbacks rather than rewritten training loops) trace back to fastai’s 2018-2020 library work.

UK Context

fast.ai’s UK community is concentrated around the academic and industrial ML hubs:

Academic Institutions:

• Imperial College London: Several DSI (Data Science Institute) and Department of Computing teaching modules reference fast.ai courses as supplementary reading. ICL’s medical imaging group has published several fastai-based works on chest X-ray classification.

• University of Cambridge: Department of Engineering deep learning module recommends fast.ai for hands-on supplement. Several MPhil ACS dissertations are built on fastai pipelines.

• UCL (University College London): Centre for Artificial Intelligence postgraduate teaching incorporates fastai for student project work; Gatsby Computational Neuroscience Unit alumni have contributed to fastai forums.

• University of Edinburgh: ANC (Adaptive and Neural Computation) and School of Informatics use fastai as a teaching reference; the Edinburgh ML meetup has hosted multiple fast.ai-themed events.

• University of Manchester: Computer Science Department machine learning courses recommend fastai courses; Manchester is a noted hub for fastai practitioners with regular meetups.

• University of Oxford: OxCSML (Oxford Computational Statistical Machine Learning) and Oxford Robotics Institute have produced student work on fastai pipelines, particularly in medical imaging and agriculture.

  UK Industrial and Civic Use:

• NHS Trusts: Several NHS digital teams have used fastai for prototype medical imaging classifiers, including diabetic retinopathy screening pilots and dermatology triage tools, often in collaboration with academic medical imaging groups.

• BBC R&D: Uses fastai for prototype content classification and recommendation work.

• Faculty AI (London): Reportedly uses fastai for rapid client-facing prototyping.

• Synthesia (London): Engineering blog posts have referenced fastai for early R&D iteration.

• DeepMind (London): Recommended as pre-onboarding curriculum for non-research engineering hires.

• Kheiron Medical Technologies (London): Founded by fast.ai community members, deploys breast cancer screening AI built on PyTorch with fastai-derived training methodology.

• BenevolentAI (London/Cambridge): Drug discovery startup with fastai-trained engineers.

• Wayve (London): Autonomous-driving startup with multiple fast.ai course graduates on staff.

  Northern English Industrial Hubs:

• Manchester: Manchester ML meetup is one of the largest active fastai-aligned communities in the UK; the Manchester-based AI consultancy Filament Consulting uses fastai for rapid client prototyping.

• Leeds: Leeds Beckett University and University of Leeds run hands-on ML courses using fastai materials; LIDA (Leeds Institute for Data Analytics) hosts fastai study groups.

• Sheffield: University of Sheffield NLP group has produced ULMFiT-derived work; Sheffield Hallam ML modules reference fastai.

• Newcastle: Newcastle University DEI (Digital Economy Institute) hosts data science programmes incorporating fastai content.

• Liverpool: Liverpool John Moores University machine learning teaching uses fastai for applied modules.

• Bristol: University of Bristol (notably its Visual Information Lab and ML group) has produced multiple fastai-based vision research outputs; Bristol is also home to one of the more active UK fastai meetups.

  UK Regulatory and Ethics Connection: Rachel Thomas’s Practical Data Ethics course is referenced in UK AI Council and Ada Lovelace Institute discussions of responsible-AI curriculum. The course was cited in the 2023 Centre for Data Ethics and Innovation review of AI education standards.

Future Directions (2026-2030)

1. fastai Library Evolution

The fastai library is expected to converge with the Answer.AI release stream over 2026-2028. Howard has indicated in public posts that a possible fastai v3 would unify the fastai training abstractions with the Answer.AI distributed training tools (FSDP-QLoRA, Claudette tool-calling primitives) into a single library positioned for the LLM-fine-tuning era. Compatibility with PyTorch 3.0 (expected late 2026 or 2027) will likely drive a maintenance release.

2. Course Programme Refresh

The 2022 Practical Deep Learning v4 course remains current as of mid-2026 but is increasingly dated relative to the LLM era. A 2026-2027 v5 release is widely anticipated covering: modern transformer fine-tuning (LoRA, QLoRA, FSDP), retrieval-augmented generation, agentic systems, multimodal models, and small-model deployment. Howard has stated the new course will use Answer.AI’s tools as primary infrastructure.

3. Answer.AI Research Trajectory

Answer.AI is positioning around four overlapping research streams:

• Consumer-GPU LLM training: extending FSDP-QLoRA to larger models, multi-node consumer-hardware clusters
• Python-only web stacks: FastHTML expansion eliminating the JavaScript dependency for ML application deployment
• Agentic systems: Claudette and tool-calling primitives for AI assistants
• Applied R&D on small teams: methodology for high-leverage two-to-five-person AI R&D groups Funded operations are expected to continue through 2028+ with publication-driven impact metrics rather than commercial-product launches.

4. Pedagogical Method Diffusion

The top-down code-first method pioneered by fast.ai has been adopted by several follow-on educators (Sebastian Raschka’s Build a Large Language Model from Scratch course and book, Andrej Karpathy’s Zero to Hero video series, Jay Alammar’s Illustrated Transformer tutorials). Expect continued diffusion of the fast.ai pedagogical style across the broader ML educational ecosystem, particularly in the bootcamp and continuing-professional-education segments.

5. Influence on PyTorch Ecosystem

The fastai callback DNA already runs through Hugging Face Accelerate, PyTorch Lightning, and the Hugging Face Trainer API. Expect this influence to consolidate into a de facto standard PyTorch training-loop abstraction by 2027-2028, possibly upstreamed into core PyTorch through the torch.distributed.elastic and torch.compile-aware training utilities.

6. UK and Global ML Education

As LLMs replace bespoke per-task training for many workflows, the fast.ai mission shifts from “train your own deep learning model” toward “deploy, fine-tune and reason about LLMs as a working coder.” The 2026-2027 course refresh is expected to lean heavily into this transition. UK academic adoption is likely to deepen as universities respond to industry pressure for applied-ML curricula.

Research & Literature

Foundational Works:

1. Howard, J., & Ruder, S. (2018). Universal Language Model Fine-tuning for Text Classification. Proceedings of the 56th Annual Meeting of the Association for Computational Linguistics (ACL 2018), 328-339. arXiv:1801.06146 [ULMFiT, 4,000+ citations, precursor to BERT/GPT transfer-learning era]
2. Howard, J., & Gugger, S. (2020). fastai: A Layered API for Deep Learning. Information, 11(2), 108. DOI: 10.3390/info11020108 [The canonical fastai v2 paper articulating the layered API design philosophy]
3. Howard, J., & Gugger, S. (2020). Deep Learning for Coders with fastai and PyTorch: AI Applications Without a PhD. O’Reilly Media. ISBN: 978-1492045526 [The canonical fastai book, used as university curriculum globally]

Closely Related Methodology Papers: 4. Smith, L.N. (2018). A disciplined approach to neural network hyper-parameters: Part 1 — learning rate, batch size, momentum, and weight decay. arXiv:1803.09820 [1cycle policy, validated through fastai courses] 5. Smith, L.N. (2017). Cyclical Learning Rates for Training Neural Networks. IEEE Winter Conference on Applications of Computer Vision (WACV 2017), 464-472. arXiv:1506.01186 [Cyclical LR foundation] 6. Zhang, H., Cisse, M., Dauphin, Y., & Lopez-Paz, D. (2018). mixup: Beyond Empirical Risk Minimization. International Conference on Learning Representations (ICLR 2018). arXiv:1710.09412 [Mixup, popularised through fastai defaults] 7. Szegedy, C., Vanhoucke, V., Ioffe, S., Shlens, J., & Wojna, Z. (2016). Rethinking the Inception Architecture for Computer Vision. IEEE Conference on Computer Vision and Pattern Recognition (CVPR 2016), 2818-2826. arXiv:1512.00567 [Label smoothing]

Transfer-Learning Lineage Following ULMFiT: 8. Peters, M.E., Neumann, M., Iyyer, M., Gardner, M., Clark, C., Lee, K., & Zettlemoyer, L. (2018). Deep contextualized word representations (ELMo). NAACL-HLT 2018. arXiv:1802.05365 [Concurrent NLP transfer learning] 9. Devlin, J., Chang, M.W., Lee, K., & Toutanova, K. (2019). BERT: Pre-training of Deep Bidirectional Transformers for Language Understanding. NAACL-HLT 2019. arXiv:1810.04805 [BERT cited ULMFiT as inspiration] 10. Radford, A., Narasimhan, K., Salimans, T., & Sutskever, I. (2018). Improving Language Understanding by Generative Pre-Training (GPT-1). OpenAI Technical Report [Generative pretraining lineage] 11. Brown, T., et al. (2020). Language Models are Few-Shot Learners (GPT-3). Advances in Neural Information Processing Systems 33 (NeurIPS 2020), 1877-1901. arXiv:2005.14165 [Modern LLM era]

PyTorch and Deep Learning Frameworks: 12. Paszke, A., et al. (2019). PyTorch: An Imperative Style, High-Performance Deep Learning Library. Advances in Neural Information Processing Systems 32 (NeurIPS 2019), 8024-8035. [PyTorch foundational paper, fastai’s underlying framework] 13. Falcon, W., et al. (2019). PyTorch Lightning. GitHub: PyTorchLightning/pytorch-lightning [Concurrent training-loop abstraction] 14. Wolf, T., et al. (2020). Transformers: State-of-the-art Natural Language Processing. EMNLP 2020 System Demonstrations, 38-45. arXiv:1910.03771 [Hugging Face Transformers]

Answer.AI Era (2024+): 15. Answer.AI Team (Dettmers, T., Belkada, Y., Howard, J., et al.) (2024). You can now train a 70b language model at home — Answer.AI. Answer.AI Technical Blog, March 6, 2024. https://www.answer.ai/posts/2024-03-06-fsdp-qlora.html [FSDP-QLoRA] 16. Howard, J. (2024). Introducing Claudette, a new friend that makes Claude 3.5 Sonnet even nicer. Answer.AI Technical Blog, June 21, 2024. https://www.answer.ai/posts/2024-06-21-claudette.html [Claudette] 17. Howard, J. (2024). About FastHTML. Answer.AI. https://about.fastht.ml/ [FastHTML announcement] 18. Dettmers, T., Pagnoni, A., Holtzman, A., & Zettlemoyer, L. (2023). QLoRA: Efficient Finetuning of Quantized LLMs. NeurIPS 2023. arXiv:2305.14314 [QLoRA underlying FSDP-QLoRA] 19. Rajbhandari, S., Rasley, J., Ruwase, O., & He, Y. (2020). ZeRO: Memory Optimizations Toward Training Trillion Parameter Models. SC20: Supercomputing. arXiv:1910.02054 [FSDP/ZeRO foundation]

Pedagogy and Data Ethics: 20. Thomas, R. (2017). Computational Linear Algebra for Coders. fast.ai Course Materials and GitHub: fastai/numerical-linear-algebra [Computational Linear Algebra] 21. Thomas, R. (2020). Practical Data Ethics. USF Center for Applied Data Ethics, fast.ai [Data ethics course] 22. Thomas, R. (2019). Bias in word embeddings and beyond. fast.ai blog [Algorithmic bias commentary]

Surveys and Influential Reviews: 23. Ruder, S. (2019). Neural Transfer Learning for Natural Language Processing. PhD Thesis, National University of Ireland Galway [ULMFiT context] 24. Howard, J. (2018). Introducing state of the art text classification with universal language models. fast.ai blog [ULMFiT announcement] 25. Pratt, L.Y. (1993). Discriminability-Based Transfer between Neural Networks. NIPS 1992, 204-211. [Foundational transfer learning]

Software Documentation and Open-Source Resources: 26. fastai documentation. https://docs.fast.ai (v2.8.x, 2024-2026) 27. nbdev documentation. https://nbdev.fast.ai (2019-2026) 28. fastcore documentation. https://fastcore.fast.ai (2019-2026) 29. Answer.AI website. https://www.answer.ai (2024-2026) 30. Practical Deep Learning for Coders 2022. https://course.fast.ai (2022 edition, current as of 2026)

Metadata

• Last Updated: 2026-05-16
• Review Status: Comprehensive editorial review during Phase 6 enrichment sprint
• Verification: Founding facts and publication details verified against fast.ai official documentation, Howard-Gugger 2020 Information journal paper (DOI: 10.3390/info11020108), ULMFiT ACL 2018 proceedings, Answer.AI blog posts (March-August 2024), the Deep Learning for Coders with fastai and PyTorch O’Reilly book, and PyPI download statistics. UK community deployment claims cross-referenced against London/Manchester/Edinburgh meetup public records and university course syllabi.
• Regional Context: UK academic institutions (Imperial College London, University of Cambridge, UCL, University of Edinburgh, University of Manchester, University of Oxford, Bristol), UK industry deployments (DeepMind, Synthesia, Wayve, BenevolentAI, Kheiron Medical, Faculty AI, BBC R&D, NHS Trusts), Northern English innovation hubs (Manchester, Leeds, Sheffield, Newcastle, Liverpool) detailed with concrete community and deployment evidence
• Preferred Term Correction: Original stub used preferred-term:: Fast AI (incorrect casing); corrected to canonical lowercase dotted form fast.ai matching the project’s own brand usage. Added alternative-terms for discoverability.
• Production-Ready: Complete OWL formal semantics, comprehensive content coverage (history, library architecture, training defaults, ULMFiT and the transfer-learning lineage, Answer.AI 2024 work, UK context, future directions), 30 academic and technical references spanning 2014-2026
• Authority Score: 0.87 (founding work for modern PyTorch transfer-learning pedagogy, ULMFiT precursor to BERT/GPT lineage, 500K+ course learners, 27.6K+ GitHub stars, Answer.AI 2024 expansion, mature production ecosystem, ongoing research relevance through Answer.AI publications)

Provenance