DeepLearn

Machine learning Experimentation Pedagogy

Learn. Play. Progress.

Courses adapted to your level, models to experiment with. Learn to play better, and grow your AI.

Discover the project

Learn AI. Take action.

Learning AI shouldn't stop at theory. DeepLearn connects courses tailored to your level, models to experiment with, and challenges to put your knowledge to the test.

Who is DeepLearn for?

The platform supports students, developers and those curious about research, from the first networks to advanced architectures. Everyone advances at their own level, with concrete terrain to practice on.

LEARN TO BECOME BETTER

  1. 01Learn

    Courses adapted to your knowledge

  2. 02Train

    Transforming an idea into an agent

  3. 03Play

    Putting your policy to the test

  4. 04Understand

    Come back with a new question

Every challenge gives a reason to return to learn

The game provides a goal. The course provides the means.

Understanding a network to create a better fighter: the game gives a reason to continue. You learn, train your agent and observe his limits in the arena. Each duel brings you back to class with a concrete question.

Two progressions that complement each other

Courses, XP and activity series follow your learning. Elo reflects the results of your fighter. Better understanding helps you improve your strategy; the duels show what remains to be worked on.

First, let's get to know each other.

When creating the account, you indicate your level, your objective, your interests and the time available. These first answers give direction to your learning.

The context before the measurement

The six-step questionnaire constitutes your profile. The diagnosis that follows then confronts this declared level with concrete questions, to adapt the courses to your actual knowledge.

FROM PERCEPTION TO OBSERVATION

01“I think I know”Self-assessment
02“I put to the test”Diagnosis corrected
PROFILE BY CONCEPTMastery × trust

Confirmed acquired skills · prerequisites to be repeated

What you think you know. What you master.

After a self-assessment in mathematics, programming and concepts in the field, an adaptive MCQ specifies your knowledge. It highlights your strengths, fragile prerequisites and gaps between theory and practice.

Mastery and confidence, concept by concept

The diagnosis crosses 30 declarative items with weighted answers to the MCQ. Additional questions remove uncertainties. The more the evidence accumulates, the more it weighs in the profile: c = n / (n + 2).

PROGRESS FROM PROFILE

  1. Your achievementsWhat the diagnosis confirms
  2. Useful prerequisitesWhat is missing for your goal
  3. The adapted courseExplanations, code and exercises
  4. Putting it into practiceSandbox, training and play

Automatic route: integration currently being finalized.

Automatic route · currently being finalized

Know where to start from. And what to learn next.

Based on your level and your objective, DeepLearn will create an automatic course: the useful prerequisites, the concepts to be studied in more depth, then the practice. You'll know what to learn next, taking into account what you already master.

From profile to an explainable itinerary

The profile by concept, the prerequisite graph and the customization of courses already exist. Their orchestration remains to be finalized to automatically connect the diagnosis, the order of courses and the practice steps.

Fundamentals to advanced architectures.

Explore MLP, CNN, LSTM, ResNet, Transformer and GAN by topic or difficulty. Each architecture is discovered through its intuitions, its mathematics and its code. Your progress is saved so you can pick up where you left off.

Several entry points in the same concept

Lessons combine explanations, diagrams, and code examples. Personalized courses add exercises and multiple choice questions to check understanding, before manipulating the model in the sandbox.

PLAN BEFORE WRITING

ENTRANCEProfile + objective
CONSTRAINTSCoursePolicy & prerequisites
BLUEPRINTA section by section plan

Concepts · remediation · exercises

Plan refinement → validation → deterministic fallback if necessary

The same subject. A different course for you.

Understanding DQN or learning how to implement it does not require the same course. You choose an intention; DeepLearn crosses it with your knowledge to adjust the reminders, depth and support of the code.

Research at the heart of planning

A CoursePolicy transforms profile and intention into educational rules. Seven signals prioritize gaps in the prerequisite graph. A deterministic plan protects the essential notions and frames its refinement by the language model.

FIND USEFUL PASSAGES

THE WORDSBM25
THE MEANINGVectors
RRF MERGER · RECLASSIFICATIONSources for each section
WritingValidationMDX course

From sources to explanations. Then put it into practice.

The plan becomes a documented course, with explanations, code and exercises tailored to your goal. Each section draws on targeted sources to connect ideas to practice.

A RAG driven by an educational contract

The search combines BM25 and vectors, merges their results by RRF, then reclassifies useful passages. The model writes section by section; validators check the MDX structure and the presence of the expected educational elements.

Go back to ideas who changed AI.

To go further, explore the work that has evolved AI. The reading sheets shed light on the intuition behind an architecture or method, then invite you to consult the original article.

Publications made searchable

Publications are organized by authors, year and theme. The search and filters allow you to find references related to the concepts studied in the courses.

Build. Test. See what changes.

Assemble the layers of a network, adjust its parameters and start training. Curves allow you to test your hypotheses on images, a pendulum or a fighter, then adjust the model.

From visual graph to executable model

The React Flow editor produces a JSON configuration of layers and hyperparameters. Metrics track training; TensorFlow.js then allows you to test the model in the browser, on the MNIST, Fashion-MNIST, CIFAR-10, CartPole or Fighter tasks.

THREE WAYS TO MAKE AN AGENT LEARN

01 / IMITATIONObserve → reproduce

A state of the game, an action shown by a bot: the network learns this association.

02 / REINFORCEAct → receive a reward

The returns from the fighting orient the gradient of policy.

03 / SELF-PLAYPlay → face your copies

A frozen copy renews the challenge and changes the situations encountered.

Evaluate victories, damage, falls and diversity of actions.

Your network learns to make decisions.

Your fighter learns a strategy: first by imitating bots, then through rewards, or by facing his own copies. You choose the training mode and observe the behaviours that emerge.

Imitation, REINFORCE, self-play

After imitation priming, REINFORCE adjusts the policy according to the combat rewards. Self-play adds a frozen opponent, renewed periodically. Victories, damage, falls and entropy reveal the strengths and limits of the strategy.

The fighter plays. You understand his choices.

Test your agent at the dojo, challenge an opponent or start a ranked match. His movements, his attacks and his mistakes show you what you need to improve in the next training session.

A game engine, a learned policy

The engine provides 30 observation values and 10 possible actions. TensorFlow.js chooses the action by argmax from the already trained network. With identical agents and seeds, the duel is reproducible; the weights remain fixed during the fight.

SITUATE THE MODEL, THEN PROGRESS

Evolution of the rankingΔElo = K × (S − E)

S: match result · E: expected score
K: sensitivity according to the experience of the fighter

  1. Bronze
  2. Money
  3. Gold
  4. Platinum
  5. Diamond
  6. Master

Close opponent in Elo → duel → analysis → new training.

An opponent within your reach. A new milestone to cross.

Elo positions your fighter against others: an unexpected victory counts more. His rank and his history give an objective for the next training session, whether it is to defend better or to vary his strategy.

Adaptive Elo and weighted matchmaking

Matchmaking favors nearby Elos and broadens the search if necessary. The ranking evolves according to the gap between result and expected score, with a K factor adapted to the experience of the fighter.

Learn. Play. Progress.

A course to understand, a sandbox to experiment, an arena to surpass yourself. Discover DeepLearn and give your next ideas a playing field.

Discover DeepLearn