Ongoing learning program

AI Algorithms Learning Laboratory

Started
2026-07
Last revised

A structured learning program that moves from mathematical derivation to implementation, tiny verification cases, and visual explanation while preparing for the 2026 Vietnamese Student AI Olympiad.

  • Python
  • TypeScript
  • NumPy
  • Dynamic programming
  • Linear algebra
RQ

Research question

Which small implementations and visual tests reveal the assumptions of an AI algorithm more clearly than using a high-level library alone?

01

From input to inspectable output

System architecture

Learning loop · one auditable artifact
  1. 01Deriveobjective + recurrence
  2. 02Implementclear reference code
  3. 03Interrogatevisible intermediate state
  4. 04Verifytiny cases + invariants
A failed check returns to the derivation, not just the code.
Fig. 03Every module ends with an explanation that can be checked against the implementation.
  1. 01

    Derive

    Write the objective, state variables, recurrence, and complexity bounds.

    in
    algorithm
    out
    mathematical note
  2. 02

    Implement

    Translate the derivation into a minimal dependency-light reference implementation.

    in
    equations
    out
    executable code
  3. 03

    Interrogate

    Expose intermediate tables, vectors, or decisions through a focused visual explanation.

    in
    execution trace
    out
    inspectable states
  4. 04

    Verify

    Use tiny hand-computable cases, invariants, and comparisons with trusted implementations.

    in
    reference output
    out
    test ledger
02

What the design must respect

Operating constraints

  1. 01

    The scope must stay narrow enough for each algorithm to be derived and tested from first principles.

  2. 02

    Reference implementations should favor clarity before vectorized performance.

  3. 03

    Visual explanations must correspond to actual intermediate states, not decorative animations.

03

Evaluation ledger

Current results

Placeholder results — no benchmark claim is being made.

Values will be added only with the evaluation set, protocol, ontology version, and relevant runtime conditions.

R01Modules completed
Not reportedPlaceholder

The public log will count only modules with derivation, implementation, and verification.

R02Verification coverage
Not reportedPlaceholder

A rubric is being defined before reporting a percentage.

04

Negative evidence

Failed or insufficient approaches

F01

Collecting disconnected notebooks

Code accumulates, but assumptions and failure cases are hard to recover later.

Design response

Require the same derive–implement–interrogate–verify structure for every module.

F02

Starting with optimized vectorized code

Compact implementations can conceal the state transitions being learned.

Design response

Keep a slow reference version beside any optimized variant.

05

What changed in the model of the problem

Lessons

  • A hand-computable example is often the fastest way to find an indexing error.
  • Visualization is most useful when it exposes the exact state used by the implementation.
  • Complexity analysis belongs before optimization.
06

Queue, not promises

Next experiments

  1. 01Complete the Viterbi module with brute-force agreement tests.
  2. 02Add stable softmax, log-sum-exp, and numerical gradient-checking notes.
  3. 03Build a small retrieval module comparing sparse and dense similarity.
07
  1. Understanding Viterbi Decoding in Linear-Chain CRFs

    A complete derivation and compact reference implementation.