I’m documenting my AI learning journey as a personal notebook of concise lessons, aha moments, and practical mental models — not a rewrite of course material. Expect short, actionable takeaways, intuition-first explanations, links to original resources, and checkpoints that save me from re-watching lectures. This log is for beginners following a structured path who want quick reminders and decision points.
Learning path
Learning path
- Intro to AI
- Math for AI/ML, intro to ML
- Machine learning
- Deep learning:
- NLP
- Frontier AI
- System
- Welcome to PyTorch Tutorials — PyTorch Tutorials 2.12.0+cu130 documentation
- Additional guidance for the training pipeline | Machine Learning | Google for Developers
- Additional guidance for the training pipeline | Machine Learning | Google for Developers
- farnazjmo/deep-learning-training-pipeline: Modular deep learning pipeline built with PyTorch
You can use roadmap.sh for shortcut, use it as a cheat sheet, but you need deeper material to master the topic.
AI, by definition
AI, by definition
Key takeaways:
- AI term is umbrella – too wide
- ML is subset of AI
- Example of ML task
- Classification
- Clustering
- Regression
- Deep Learning (DL) is subset of ML. example DL: use artificial neural network (ANN), on each layer, it will extract a feature, then pass it next layer. hence, model is dumb.
Definition:
Algorithm enabled by constraint exposed by representation that support model targeted at thinking, perception, and action.
Adam teaching principle
Adam teaching principle
And He taught Adam the names – all of them…
QS 2:31
Re-introduced as Rumpelstiltskin Principle, is a principle that if we know true name, we get the power over it, meaning:
- Naming change something unfamiliar into familiar thing that decrease anxiety.
- Label create taxonomy and abstraction:
- Taxonomy speed up tracking and tracing, and easier for modification.
- Abstraction make easier decision making since it simplify the details and put emphasize on significant pattern.
- Since abstraction simplify the details:
- Abstraction become an enabler to apply rule over it.
- Its provide identity.
Fun fact:
- Right representation → almost done
- Simple ≠ trivial
- Simple: easy to do
- Trivial: not only easy, but also insignificant/unimportant.
- Naming become mandatory
Without naming, system cannot represent knowledge.
How AI solve problem
How AI solve problem
Modeling problem solving
Generate and test
Problem reduction
Safe transformer
Problem reduction is an approach, to convert the problem into another problem that we know it proven to have a known solution. hence, this transformation is SAFE. Here illustration diagram:
This mean, bring an unknown problem into known problem by apply all safe transformer:
Example, in integral calculus, we have following proven solution for particular integral calculus:
Let’s apply to real integral calculus problem:
We apply several safe transformer, become:
Now we stuck, we need less credible transformer, we call it HEURISTIC transformer. Heuristic = not always work (hence less safe), but most likely work. example:
Apply the heuristic transformer, we will get:
Problem reduction tree
If you look at safe transformer for integral calculus rule number 3, that integral of sum function is equal to sum of integral function, our diagram of applied transformer become:
We have AND node in the middle, and several outgoing paths represent the integral functions, that later we will apply summation for all the result. example:
If we draw applied transformations so far:
If we continue to apply heuristic transform to last leave, it become:
And last leave above can have AND node to 3 transforms:
Now you can see, all leaves now safe transform, we show that problem now SOLVED.
Now we can update our diagram of generate and test approach:
For closing remark, we show statistic below:
- 54 over 56 integral calculus problem solved. you may guess that its because lack of memory (34kb at that time), but actually we lack of 2 transformers.
- Max tree depth: 7
- Average tree depth: 3
- Average useless tree branch: 1
- Size of safe transformer: about 12
- Size of heuristic transformer: about 12
From statistic above, we conclude that it’s actually easy to solve integral calculus.
Is this statistic applies to other domain: yes, mo re or less!
Reflections
Another Example how AI solve problem
Another example how AI solve problem : block world problem
Block world problem is a toy app implementation to show how AI can solve request to arrange blocks.
You can also ask explanation for the answer:
To solve this problem we can use problem reduction approach that we already exercise before. Here the diagram of the problem reduction:
Leave a Reply