title
stringlengths 1
544
โ | parent
stringlengths 0
57
โ | created
stringlengths 11
12
โ | editor
stringclasses 1
value | creator
stringclasses 4
values | edited
stringlengths 11
12
โ | refs
stringlengths 0
536
โ | text
stringlengths 1
26k
| id
stringlengths 32
32
|
|---|---|---|---|---|---|---|---|---|
Explainable AI
|
AI Terms
|
Mar 4, 2023
|
Alan Jo
|
Alan Jo
|
Aug 6, 2023
|
### Interpretability
degree to which a model can be understood in human terms
### Explainability
๋ชจ๋ธ์ด ๋ณต์กํ๊ฑฐ๋ ๋ธ๋๋ฐ์ค๋ผ ์ถ๊ฐ์ ์ธ ํ
ํฌ๋์ ์ฌ์ฉํด์ ์์ ๋ชจ๋ธ์ ๊ฒฐ์ ๊ณผ์ ์ ์ดํด
### Using another AI
> [OpenAI's new tool attempts to explain language models' behaviors](https://techcrunch.com/2023/05/09/openais-new-tool-attempts-to-explain-language-models-behaviors/)
|
e1b35d4b9a6342dc863578350a7b4325
|
|
Super Intelligence
|
AI Terms
|
Aug 23, 2020
|
Alan Jo
|
Alan Jo
|
Jul 4, 2023
|
[AGI](https://texonom.com/agi-38ec1ab5796f472ca4475676519e29c1)
|
### ์์์ด๋ผ๋ ํ์์ ๊ตฌํํ๋ ค๊ณ ํ๋ฉด ์๋๋ค
**์ด์ง๋ฅ AI์ ๋ชฉ์ ์ ์ด๋ป๊ฒ ์ฐ๋ฆฌ ๋ชฉ์ ์ ๋ง์ถฐ ์ ๋ ฌํ ์ง ๋ฐฉ๋ฒ์ ์ฐพ์๋ด๋ ๊ฑด ์ค์ํ๊ณ ์ด๋ ต๋ค**
**ํ์๋ชฉ์ ํ์ ๋ฌธ์ **๋ฅผ ๊ณ ๋ คํ ๋, ์ฐ๋ฆฌ๋ ๋ชฉ์ ์ ๋ ฌ๋ฌธ์ ๋ฅผ ํด๊ฒฐ์ด์ ์ ์ด์ง๋ฅ์ ํ๋ฉด ์๋๋ค
์ปดํจํฐ๋ ๊ตฌ์กฐ์ ๋
๋ฆฝ์ ์ผ๋ก ์์ฉ์ผ๋ก ๋ชจ๋ ๊ธฐ๋ฅ์ ๊ตฌํํ ์ ์๋ ๊ตฌ์กฐ์ด๋ค
๋๋ ๊ธฐ๋ฅ ์์ฒด์ ๋ฐ๋ผ ์งํํ๋ฉฐ ์ฐ๊ด๋ ๊ตฌ์กฐ์ด๋ค
๊ตฌ์กฐ์ ๊ธฐ๋ ์์ฒด๊ฐ ๋ณ๋ ฌ๋ก ์ ๋ณด๋ฅผ ๋จ์ํํ์ฌ ์์ํ๋ ๊ฒ์ด ๋ชฉํ๊ฐ ๋๊ธฐ์ ์ง๋ ฌํ๋ ์ปดํจํฐ๋ก ๊ตฌํํ๊ธฐ ์ด๋ ต๋ค
### Super Intelligence Books
- Life 3.0
- ํน์ด์ ์ด ์จ๋ค
> [Untitled](https://www.youtube.com/watch?v=HkdpO6GSxYI)
|
b057e644731546d9b39cb41939d36712
|
Technological singularity
| null | null | null | null | null | null |
๊ธฐ์ ์ ํน์ด์
๋ชจ๋ ์ธ๋ฅ์ ์ง์ฑ์ ํฉ์น ๊ฒ๋ณด๋ค ๋ ๋ฐ์ด๋ ์ด์ธ๊ณต์ง๋ฅ์ด ์ถํํ๋ ์์
### ์ ํ์ฃผ์์
### Singularitarian
> [๊ธฐ์ ์ ํน์ด์ ](https://ko.wikipedia.org/wiki/%EA%B8%B0%EC%88%A0%EC%A0%81_%ED%8A%B9%EC%9D%B4%EC%A0%90#%EC%9C%A0%EC%A0%84%EC%9E%90_%EC%B9%A9)
|
4a2e2f5be15a4947ab1515a9ea8e1903
|
**Transfer Learning**
| null | null | null | null | null | null |
## Transfer weight from model to model
ํ๋์ ๋ฌธ์ ๋ฅผ ํด๊ฒฐํ๊ณ ์ด์ ๋ค๋ฅด๋ฉด์ ๊ด๋ จ๋ ๋ฌธ์ ์ ์ ์ฉํ๋ ๋์ ์ป์ ์ง์์ ์ ์ฅํ๋๋ฐ ์ง์คํ๋ ๊ธฐ๊ณ ํ์ต์ ์ฐ๊ตฌ ๋ฌธ์
transfer learning by proportionally small data
make category vector
### Category Vector - Style Vector
- one hot vector - usual
- Gaussian normal random vector - for linear
ํ๋์ ๋ฌธ์ ๋ฅผ ํด๊ฒฐํ๊ณ ์ด์ ๋ค๋ฅด๋ฉด์ ๊ด๋ จ๋ ๋ฌธ์ ์ ์ ์ฉํ๋ ๋์ ์ป์ ์ง์์ ์ ์ฅํ์ฌ ์ด์ฉ
Pre trained model
> [[6์ฃผ์ฐจ] Introduction to Meta Learning](https://velog.io/@tobigs_xai/6์ฃผ์ฐจ-Introduction-to-Meta-Learning)
|
442feb66465944eebf144d4e9dd1dbf8
|
Turing Test
|
AI Terms
|
Dec 31, 2020
|
Alan Jo
|
Alan Jo
|
Jul 5, 2023
|
[Alan Turing](https://texonom.com/alan-turing-df61b5084e544ac58577d289983e0d28)
|
### Ambiguous
### Turing Test Notion
|Title|
|:-:|
|[Church Turing Thesis ](https://texonom.com/church-turing-thesis-0bbb1e1b35064a8f88cda36b2739f384)|
|[Feasibility thesis](https://texonom.com/feasibility-thesis-1ece05de5a8d443683dbeab4307877ac)|
|[Reverse Turing Test](https://texonom.com/reverse-turing-test-3d2e3d1006d640ce8169187408d6f5dc)|
|[Visual Turing Test](https://texonom.com/visual-turing-test-6bfe733ba0ce4ec4820717c8e2f6f1f5)|
์ธ๊ฐ๊ณผ ์ธ๊ฐ ๊ตฌ๋ณ๋ณด๋ค ai์ ai ๊ตฌ๋ณ์ด ์ค์ํด์ง
> [The Turing Test is obsolete. It's time to build a new barometer for AI](https://www.fastcompany.com/90590042/turing-test-obsolete-ai-benchmark-amazon-alexa)
### [Mirror Test](https://texonom.com/mirror-test-a3383d21fd964c59a91c80cf546348cb)
> [Introducing theย AIย Mirror Test, which very smart people keep failing](https://www.theverge.com/23604075/ai-chatbots-bing-chatgpt-intelligent-sentient-mirror-test)
|
db633c61340449c4bf0143b06fe981c0
|
Church Turing Thesis
|
Turing Test Notion
|
Mar 2, 2021
|
Alan Jo
|
Alan Jo
|
Jul 5, 2023
|
๋ชจ๋ ๊ธฐ๊ณ์์ ์ํ๋๋ ์ฐ์ฐ์ ํ๋ง๊ธฐ๊ณ์์ ์ํ ๊ฐ๋ฅ
๋ช
์ ์ธ ์ด์ ๋ ์์ ํ์ง ์๊ธฐ ๋๋ฌธ
์์ง๊น์ง ํ๋ฆผ ์ฆ๋ช
์๋
church turing thesis ๋ฏธ์ฆ๋ช
๋
> [ChurchโTuring thesis](https://en.wikipedia.org/wiki/ChurchโTuring_thesis)
|
0bbb1e1b35064a8f88cda36b2739f384
|
|
F**easibility thesis**
| null | null | null | null | null | null |
## Extended Church Turing Thesis
[Quantum Computing](https://texonom.com/quantum-computing-f7299cf6bc554d58aeae450df6bc6248) ์ ํฌํจํ๋ฉด ๋ชจ๋ ํ์์ ๊ณ์ฐ์ผ๋ก ์๋ฎฌ๋ ์ด์
๊ฐ๋ฅํ๋ค๋ ๋ฏฟ์ [Simulation Cosmology](https://texonom.com/simulation-cosmology-e100063d72584813b17c3da95333591f)
whether an arbitrary but "reasonable" model of computation can be efficiently simulated
|
1ece05de5a8d443683dbeab4307877ac
|
Reverse Turing Test
|
Turing Test Notion
|
Jun 21, 2022
|
Alan Jo
|
Alan Jo
|
Jul 5, 2023
|
[CAPTCHA](https://texonom.com/captcha-c7d4838aad8a44b99ba70290eb6e6bf1)
|
3d2e3d1006d640ce8169187408d6f5dc
|
|
Visual Turing Test
|
Turing Test Notion
|
Dec 3, 2021
|
Alan Jo
|
Alan Jo
|
Jul 5, 2023
|
[Virtual Reality](https://texonom.com/virtual-reality-51e72eb0921f4d94bd512ca1c2e9937a)
> [[๋ฏธ๋ผํด๋ ํฐ] ์ ์ปค๋ฒ๊ทธ๊ฐ ๋งํ VRํ๋ง ํ
์คํธ](https://stibee.com/api/v1.0/emails/share/TcTVnjQyZd2t3dWp3gOczltwnB_mXvM=)
|
6bfe733ba0ce4ec4820717c8e2f6f1f5
|
|
****CAPTCHA****
|
Reverse Turing Test
| null | null | null | null | null |
[mCaptcha](https://github.com/mCaptcha/mCaptcha)
> [Cloudflare's CAPTCHA replacement lacks crosswalks, checkboxes, Google](https://arstechnica.com/information-technology/2022/09/cloudflares-captcha-replacement-lacks-crosswalks-checkboxes-google)
> [How Apple could kill CAPTCHAs with Private Access Tokens | AppleInsider](https://appleinsider.com/articles/22/06/14/how-apple-could-kill-captchas-with-private-access-tokens)
|
c7d4838aad8a44b99ba70290eb6e6bf1
|
AI Ensemble
|
Machine Learning Notion
|
Oct 6, 2021
|
Alan Jo
|
Alan Jo
|
Jun 4, 2023
|
[Knowledge Distillation](https://texonom.com/knowledge-distillation-d5f32ad3da32434892a9765d68d31542)
|
### Very Heavy
1. sampling
2. make diverse model
3. boost every model
4. blend models
### AI Ensemble Notion
|Title|
|:-:|
|[Bagging and Pasting](https://texonom.com/bagging-and-pasting-8e4e988cff704560970d031ce9958b77)|
|[Random Forest](https://texonom.com/random-forest-2cf3f2d6f4994017accc228c98c160c8)|
|[Boosting](https://texonom.com/boosting-c383ee60ceaf4e2b9ab03a2382760b62)|
|[Voting Classifier](https://texonom.com/voting-classifier-a5d5cf9476d847089125eb3a53bc3ef2)|
|[Stacking](https://texonom.com/stacking-58a696bdecb54bb29f0c16d32031437e)|
ํด๋์ค(๋๋ label) ์ ๋ถ๋ฅ์ ์ํ ์ ํ๋ชฉ๋ก ๊ฐ๊ฐ
> [์์๋ธ ํ์ต ๋ฐ ๋๋ค ํฌ๋ ์คํธ](https://excelsior-cjh.tistory.com/166)
|
03bf3a1926dd46f18400b5830d8fdf0b
|
Discriminative Learning
|
Machine Learning Notion
|
Mar 30, 2023
|
Alan Jo
|
Alan Jo
|
Jun 25, 2023
|
[Generative Learning](https://texonom.com/generative-learning-63566344cc974563bcd76131207267f9) [Decision Boundary](https://texonom.com/decision-boundary-716a1c142cd74425bb819e60e11e3e2b) [Frequentist](https://texonom.com/frequentist-c7636520ad8d4ae3ba69b8c24f516ab1)
|
### Goal is directly model & estimate posterior probability
**ํด๋์ค ๊ฐ ์ฐจ์ด์ ์ง์คํด ํ์ต**
์ง์ ๋ชจ๋ธ๋งํ๊ธฐ ๋๋ฌธ์, ๋ณดํต ๋ถ๋ฅ ์์
์์ ๋ ํจ์จ์ ์ด๊ณ ์ ํํ ๊ฒฐ๊ณผ
Require positive and negative training data, Can be hard to interpret
### Discriminative Learning Methods
|Title|
|:-:|
|[SVM](https://texonom.com/svm-8bf85b3080e2423ca43da0995a642acf)|
|
6c8fa85172824233934d62cc217871f4
|
Generative Learning
|
Machine Learning Notion
|
Mar 30, 2023
|
Alan Jo
|
Alan Jo
|
Jun 25, 2023
|
[Discriminative Learning](https://texonom.com/discriminative-learning-6c8fa85172824233934d62cc217871f4) [Bayesian inference](https://texonom.com/bayesian-inference-e3cabe7844954db097867a732143ddd8) [Bayes Theorem](https://texonom.com/bayes-theorem-e512b6c0308f4270aaae9ae53060ccab) [Bayesian](https://texonom.com/bayesian-abfa2017bb054f3a9dfcbc9a6c1042f6)
|
### model and learn likelihood from data distribution
Estimate prior and deduce posterior
**๋ฐ์ดํฐ์ ๋ถํฌ๋ฅผ ํ์ตํ์ฌ ์์ฑ ๊ท์น**์ ํ์
can compute how probable any given model instance is
Can be learned using images from just a single category
### Generative Learning Notion
|Title|
|:-:|
|[Generative Model](https://texonom.com/generative-model-6e5204d2982b4042847aa42e88eb8fb5)|
|[Discriminant Analysis](https://texonom.com/discriminant-analysis-2ffa0dbd34204642b6fe23ed5fbf3b00)|
|[Naรฏve Bayes](https://texonom.com/nave-bayes-0d56821e2bcf4504899d63c22d9a533f)|
|[Variational Inference](https://texonom.com/variational-inference-05563e34ddec467c95c9bda931770d50)|
|
63566344cc974563bcd76131207267f9
|
Inductive Learning
|
Machine Learning Notion
|
Mar 6, 2023
|
Alan Jo
|
Alan Jo
|
Mar 30, 2023
|
[Inductive Inference](https://texonom.com/inductive-inference-5b1e3298a9e0410d9d0a6de3da798a69) [Transduction Learning](https://texonom.com/transduction-learning-1ff573898b0b46f782e199870952ddad)
|
### reasoning from observed training cases to general rules
|
b4088c5d69514827aa95fcf12c0daaa9
|
Machine Learning Theory
|
Machine Learning Notion
|
Nov 5, 2019
|
Alan Jo
|
Seong-lae Cho
|
Mar 9, 2023
|
### ML Theory Notion
|Title|
|:-:|
|[Cross Validation](https://texonom.com/cross-validation-184e7cfc80d6410188922d27ef4ef52c)|
|[Learning Hypothesis](https://texonom.com/learning-hypothesis-a4cc437840df4eb99cb311d1ab836258)|
|[Unknown Target Function](https://texonom.com/unknown-target-function-a71dd16e5b6943f28827e946ace7914b)|
|[Empirical Rule](https://texonom.com/empirical-rule-31cf23b1578c42ab84e160251c3ca38b)|
|[Logistic Multinomial Classification](https://texonom.com/logistic-multinomial-classification-3d7407bc18da4940bdadb5849a8cc3eb)|
|[Maximum Likelihood Estimation](https://texonom.com/maximum-likelihood-estimation-b30e67a4e47b4b7494aa23667adb6acd)|
|[Hyperparameter](https://texonom.com/hyperparameter-ef7e34566add4e98b673d4cef59fca90)|
|[Closed-form Solution](https://texonom.com/closed-form-solution-02dd2781a9334df7ba731da9b9e31ea4)|
|[Automatic Differentiation](https://texonom.com/automatic-differentiation-368d94f0c6d848ff93a2f10f99bd4e01)|
|[Kernel Trick](https://texonom.com/kernel-trick-a3d7cdc0b30d44efbb072a926b0d99aa)|
|[Learning Rate](https://texonom.com/learning-rate-c9a29280131f4f95a6191bde244f6789)|
### Training method
> 
> # Replay memory
> <details><summary>
>
> **Why?**</summary>
> Learn each episode only once and throw away => Wasteful (Some episodes are rare or costly to obtain)
> </details>
>
> <details><summary>
>
> **Usefulness**</summary>
> 1. the process of credit/blame propagation is sped up
> 2. an agent would get a chance to refresh what it has learned before
> </details>
>
> # Self Play
> <details><summary>
>
> **Why?**</summary>
> 1. Converting compute into data (Data generation!)
> 2. Smooth learning (Always 50% of winning rate)
> </details>
>
> ### Programming pattern of tranning
> 1. Data Provider โ Pre-processing data
> 2. Design model (network) โ Define as class
> 3. Construct loss / optimizer โ Using provided APIs
> 4. Training cycle โ Implement for-loop of forward, backward, update
> 
> - There are tons of episodes so we multi-processing
|
c69317fb5dc24f05a29532fa27d27cb1
|
|
****Progressive Learning****
|
Machine Learning Notion
|
Jun 20, 2023
|
Alan Jo
|
Alan Jo
|
Jun 20, 2023
|
*1. ๊ฐ feature ๊ณ์ธต์์ ์ฌ์ ์ง์์ ํตํฉํ ์ ์๋ ๋ฅ๋ ฅ*
*2. ์์ ๊ฒฐ๊ณผ๋ ์ฌ์ฌ์ฉํ๊ณ ์๋ก์ด ๊ฒ์ ๋ฐฐ์ฐ๋ ๋ฅ๋ ฅ*
*3. catastrophic forgetting์ ์ค์ด๋ ๋ฅ๋ ฅ*
> [Progressive Neural Network ์ดํดํ๊ธฐ](https://medium.com/@tjpark1990/progressive-neural-network-์ดํดํ๊ธฐ-2ce6a866b360)
|
46a222e83bc64bd49f4c44c3d0519fdb
|
|
T**ransduction **Learning
|
Machine Learning Notion
|
Mar 6, 2023
|
Alan Jo
|
Alan Jo
|
Mar 7, 2023
|
[Logic](https://texonom.com/logic-e562053ace984677a72cb9eaf5c6f91e) [Inductive Learning](https://texonom.com/inductive-learning-b4088c5d69514827aa95fcf12c0daaa9) [Semi-supervised Learning](https://texonom.com/semi-supervised-learning-c7d63d5b1abf49a28ba0dc07fd1db8ad)
|
### ๋ณํ
Reasoning from observed, specific cases to specific cases
unlabeled training data๋ ๊ทธ๋ค์ด ๊ฐ์ง ํน์ฑ(ex. ๋ฐ์ดํฐ ๊ฐ ์ฐ๊ฒฐ ๊ด๊ณ, ๊ฑฐ๋ฆฌ)์ ํ์ฉํด ์๋ก์ด prediction์ ํ๋ ๊ฒ
์ฌ์ ์ ๋ช
์์ ์ธ function parameter๋ฅผ ํ์ตํ์ง ์๋๋ค
### T**ransduction **Learning Usages
|Title|
|:-:|
> [Transduction (machine learning)](https://en.wikipedia.org/wiki/Transduction_(machine_learning))
> [transductive learning VS inductive learning](https://redstarhong.tistory.com/88)
|
1ff573898b0b46f782e199870952ddad
|
Bagging and Pasting
|
AI Ensemble Notion
|
Oct 6, 2021
|
Alan Jo
|
Alan Jo
|
Mar 5, 2023
|
### bagging is **bootstrap aggregating**
- **Random Patches method** : ํน์ฑ ๋ฐ ๋ฐ์ดํฐ ์
์ํ๋ง(bootstraping) ๋ชจ๋ ์ฌ์ฉํ๋ ๋ฐฉ์
- **Random Subspace method** : ํน์ฑ๋ง ์ํ๋งํ๋ ๋ฐฉ์
ํน์ฑ๋ฐฐ๊น
ํ๋ฉด ํธํฅ ์ฆ๊ฐ, ๋ถ์ฐ ๊ฐ์
๋ฐ์ดํฐ์
๋ฐฐ๊น
ํ๋ฉด ํธํฅ ๋น์ท, ๋ถ์ฐ๊ฐ์
# Notion
at statistics, ์ค๋ณต์ ํ์ฉํ ๋ฆฌ์ํ๋ง(resampling) is bootstrapping
์ค๋ณต์ ํ์ฉํ์ง ์๋ ์ํ๋ง ๋ฐฉ์์
**ํ์ด์คํ
**pasting
์ ์ฒด ํ์ต ๋ฐ์ดํฐ์
์ ์ด์ฉํด ํ๋์ ๋ชจ๋ธ์ ํ์ต์ํฌ ๋์ ๋น๊ตํด์ ํธํฅ์ ๋น์ทํ์ง๋ง ๋ถ์ฐ์ ์ค์ด๋ ๋ค
# class
๋ถ๋ฅ(classification)์ผ ๋๋
**์ต๋น๊ฐ(mode)**
์ฆ, ๊ฐ์ฅ ๋ง์ ์์ธก ํด๋์ค๋ก ์์๋ธ์ด ์์ธกํ๋ฉฐ, ํ๊ท(regression)์ผ ๊ฒฝ์ฐ์๋ ๊ฐ ๋ถ๋ฅ๊ธฐ์ ์์ธก๊ฐ์ ํ๊ท ์ ๊ณ์ฐํ์ฌ ํ๊ท ๊ฐ์ ์์ธก๊ฐ์ผ๋ก ํ๋ค.
# usage
๋ ํ ๋ฐ์ดํฐ ์ํ์ ์ฌ๋ฌ๋ฒ ์ํ๋ง ๋๊ณ , ๋ ์ด๋ ํ ์ํ์ ์ ํ ์ํ๋ง ๋์ง ์์ ์๊ฐ ์๋ค. ํ๊ท ์ ์ผ๋ก ํ์ต ๋จ๊ณ์์ ์ ์ฒด ํ์ต ๋ฐ์ดํฐ์
์ค 63% ์ ๋๋ง ์ํ๋ง ๋๋ฉฐ(์์ธํ ๋ด์ฉ์ [์ฌ๊ธฐ](https://tensorflow.blog/%EB%9E%9C%EB%8D%A4-%ED%8F%AC%EB%A0%88%EC%8A%A4%ED%8A%B8%EC%97%90%EC%84%9C-%EB%8D%B0%EC%9D%B4%ED%84%B0%EA%B0%80-%EB%88%84%EB%9D%BD%EB%90%A0-%ED%99%95%EB%A5%A0/) ์ฐธ๊ณ ), ์ํ๋ง ๋์ง ์์ ๋๋จธ์ง 37% ๋ฐ์ดํฐ ์ํ๋ค์ **oob(out-of-bag) ์ํ**์ด๋ผ๊ณ ํ๋ค.
์์๋ธ(๋ฐฐ๊น
) ๋ชจ๋ธ์ ํ์ต ๋จ๊ณ์์๋ oob ์ํ์ด ์ฌ์ฉ๋์ง ์๊ธฐ ๋๋ฌธ์, ์ด๋ฌํ oob ์ํ์ ๊ฒ์ฆ์
(validation set)์ด๋ ๊ต์ฐจ๊ฒ์ฆ(cross validation)์ ์ฌ์ฉํ ์ ์๋ค.
|
8e4e988cff704560970d031ce9958b77
|
|
Boosting
|
AI Ensemble Notion
|
Oct 6, 2021
|
Alan Jo
|
Alan Jo
|
Mar 5, 2023
|
์ฑ๋ฅ์ด ์ฝํ ํ์ต๊ธฐ(weak learner)๋ฅผ ์ฌ๋ฌ ๊ฐ ์ฐ๊ฒฐํ์ฌ ๊ฐํ ํ์ต๊ธฐ(strong learner)๋ฅผ ๋ง๋๋ ์์๋ธ ํ์ต
์์์ ํ์ต๋ ๋ชจ๋ธ์ ๋ณด์ํด๋๊ฐ๋ฉด์ ๋๋์ ๋ชจ๋ธ๋ก ํ์ต
# AdaBoost
๊ณผ์์ ํฉ(underfitted)๋๋ ํ์ต ๋ฐ์ดํฐ ์ํ์ ๊ฐ์ค์น๋ฅผ ๋์ด๋ฉด์ ์๋ก ํ์ต๋ ๋ชจ๋ธ์ด ํ์ตํ๊ธฐ ์ด๋ ค์ด ๋ฐ์ดํฐ์ ๋ ์ ์ ํฉ๋๋๋ก ํ๋ ๋ฐฉ์์ด๋ค
์ฌ๊ธฐ์ ๋ฉ๋ฆฌ์๋๋๋ค์ด ๊ณผ์์ ํฉ
**weighted linear combination**
> (๊ฐ์ค์น ์ ํ๊ฒฐํฉ)์ ํ์ฌ h(x)strong learner
> 
>
โ๋ฅผ ๊ตฌํ์ฌ ๊ฐ์ ์์ธกํ๋ค.
# ๊ทธ๋๋์ธํธ ๋ถ์คํ
์๋ค๋ถ์คํธ'์์ ์ดํด๋ณธ ๊ฒ ์ฒ๋ผ ์ ์ ํ์ต๋ ๋ชจ๋ธ์ ์ค์ฐจ๋ฅผ ๋ณด์ํ๋ ๋ฐฉํฅ์ผ๋ก ๋ชจ๋ธ(๋ถ๋ฅ๊ธฐ, ํ์ต๊ธฐ)์ ์ถ๊ฐํด์ฃผ๋ ๋ฐฉ๋ฒ์ ๋์ผํ๋ค. ํ์ง๋ง,
**๊ทธ๋๋์ธํธ ๋ถ์คํ
**
์ ์๋ค๋ถ์คํธ ์ฒ๋ผ ํ์ต๋จ๊ณ ๋ง๋ค ๋ฐ์ดํฐ ์ํ์ ๊ฐ์ค์น๋ฅผ ์
๋ฐ์ดํธ ํด์ฃผ๋ ๊ฒ์ด ์๋๋ผ ํ์ต ์ ๋จ๊ณ ๋ชจ๋ธ์์์
**์์ฌ ์ค์ฐจ(residual error)**
์ ๋ํด ์๋ก์ด ๋ชจ๋ธ์ ํ์ต์ํค๋ ๋ฐฉ๋ฒ์ด๋ค.
์ค์ฐจ์ ๋ํ ์ ๋ชจ๋ธ!์ด๋ ๊ฒ
`learning_rate`๋ฅผ ์ด์ฉํ๋ ๋ฐฉ๋ฒ์**์ถ์**(shrinkage)๋ผ๊ณ ํ๋ ๊ท์ (regularization) ๋ฐฉ๋ฒ์ด๋ค.
overfitting ์ฌ์
์กฐ๊ธฐ ์ข
๋ฃ(early stopping)
GBRT๋ฅผ ํ์ต์ํจ ํ
`staged_predict()`
๋ฉ์๋๋ฅผ ์ด์ฉํด ๊ฐ ๋ชจ๋ธ์ ์์ธก๊ฐ(
`y_pred`
)๊ณผ ์ค์ ๊ฐ(
`y_val`
)์ MSE๋ฅผ ๊ตฌํ ๋ค MSE๊ฐ์ฅ ๋ฎ์ ์ต์ ์ ํธ๋ฆฌ์ ๊ฐ์๋ฅผ ์ฐพ์ ๋ค์ ์ต์ ์ ํธ๋ฆฌ ๊ฐ์(
`best_n_estimator`
)๋ก ๊ทธ๋๋์ธํธ ๋ถ์คํ
์ ํ์ต
|
c383ee60ceaf4e2b9ab03a2382760b62
|
|
Random Forest
|
AI Ensemble Notion
|
Oct 6, 2021
|
Alan Jo
|
Alan Jo
|
Mar 5, 2023
|
๋ฐฐ๊น
(bagging)์ ์ ์ฉํ ์์ฌ๊ฒฐ์ ๋๋ฌด(decision tree)์ ์์๋ธ
ํธ๋ฆฌ๋ฅผ ์์ฑํ ๋, ๊ฐ ๋
ธ๋๋ ๋๋คํ๊ฒ ํน์ฑ(feature)์ ์๋ธ์
(์์ ๋
ธ๋๋ฅผ)์ ๋ง๋ค์ด ๋ถํ ํ๋ค.
**์์คํธ๋ผ ํธ๋ฆฌ**
(Extra-Trees)๋ ํธ๋ฆฌ๋ฅผ ๋์ฑ ๋๋คํ๊ฒ ์์ฑํ๊ธฐ ์ํด ๋
ธ๋๋ฅผ ๋ถํ ํ๋ ์ต์ ์ ์๊ณ๊ฐ์ ์ฐพ๋ ๊ฒ์ด ์๋๋ผ ํ๋ณด ํน์ฑ์ ์ด์ฉํด ๋๋คํ๊ฒ ๋ถํ ํ ๋ค์ ๊ทธ ์ค์์ ์ต์์ ๋ถํ ์ ์ ํํ๋ ๋ฐฉ๋ฒ์ด๋ค. ๋๋ค ํฌ๋ ์คํธ ์ฒ๋ผ ๊ฐ ๋
ธ๋์ ํน์ฑ๋ง๋ค ์ต์ ์ ์๊ณ๊ฐ์ ์ฐพ๋๊ฒ์ด ์๋๊ธฐ ๋๋ฌธ์ ์์คํธ๋ผ ํธ๋ฆฌ๊ฐ ํจ์ฌ ํ์ต ์๋๊ฐ ๋น ๋ฅด๋ค
**์ฝ๊ฒ๋งํด ํธ๋ฆฌ๋ณ๋ก ํน์ฑ์ ๋ฐ๋ผ ๋ชฉํ๋ง ์ฐพ์์ ๋น ๋ฆ**
๋๋ค ํฌ๋ ์คํธ์ ์ฅ์ ์ ํน์ฑ(feature)์ ์๋์ ์ธ ์ค์๋๋ฅผ ์ธก์ ํ๊ธฐ ์ฝ๋ค๋ ๊ฒ์ด๋ค. Scikit-Learn์์๋ ์ด๋ ํ ํน์ฑ์ ์ฌ์ฉํ ๋
ธ๋๊ฐ ๋ถ์๋(impurity)๋ฅผ ์ผ๋ง๋ ๊ฐ์์ํค๋์ง๋ฅผ ๊ณ์ฐํ์ฌ ๊ฐ ํน์ฑ๋ง๋ค ์๋์ ์ค์๋๋ฅผ ์ธก์ ํ๋ค.
### ์๋ฅผ๋ค์ด ํฝ์
๋ณ ์ค์๋ ๋ํ๋ผ ์ ์์
mnist ํฝ์
๋ณ ์ค์๋
[๋ถ์๋](https://texonom.com/14c2866368974d6fa82b764ef56eed7f)
|
2cf3f2d6f4994017accc228c98c160c8
|
|
Stacking
|
AI Ensemble Notion
|
Oct 6, 2021
|
Alan Jo
|
Alan Jo
|
Mar 5, 2023
|
not just vote
(stacking, stacked generalization์ ์ค์)์ voting๋ฅ๊ฐ ์๋๋ผ ์์๋ธ ํ์ต์์ ๊ฐ ๋ชจ๋ธ์ ์์ธก๊ฐ์ ๊ฐ์ง๊ณ ์๋ก์ด
**๋ฉํ ๋ชจ๋ธ(meta learner)**
์ ํ์ต์์ผ ์ต์ข
์์ธก ๋ชจ๋ธ์ ๋ง๋๋ ๋ฐฉ๋ฒ์ ๋งํ๋ค.
# Blender
๋ผ๋ meta learner ๊ฐ ํ์
and boosting
|
58a696bdecb54bb29f0c16d32031437e
|
|
Voting Classifier
|
AI Ensemble Notion
|
Oct 6, 2021
|
Alan Jo
|
Alan Jo
|
Mar 5, 2023
|
hard voting - direct democracy
soft voting - indirect democracy
|
a5d5cf9476d847089125eb3a53bc3ef2
|
|
๋ถ์๋
|
Random Forest
| null | null | null | null | null |
์ด๋ ๊ฒ ํ๋์ ํน์ฑ์์ ๋ ๊ฐ์ ์๋ธ์
์ผ๋ก ๋๋๋ ๊ธฐ์ค์ ๋ถ์๋(impurity)๋ฅผ ์ต์(๋๋, ์๋, homogeneity๋ฅผ ์ต๋)๋ก ํ๋ ๊ฒ์ด๋ค. ์ด๋ ๊ฒ ๋ถ์๋ ์ฆ, ๋ถํ์ค์ฑ์ด ๊ฐ์ํ๋ ๊ฒ์ ์ ๋ณด์ด๋ก ์์๋ ์ ๋ณดํ๋(Information Gain)์ด๋ผ๊ณ ํ๋ค
> [ExcelsiorCJH/Hands-On-ML](https://github.com/ExcelsiorCJH/Hands-On-ML/blob/master/Chap06-Decision_Tree/Chap06-Decision_Tree.ipynb)
|
14c2866368974d6fa82b764ef56eed7f
|
SVM
|
Discriminative Learning Methods
|
Mar 5, 2023
|
Alan Jo
|
Alan Jo
|
Jul 4, 2023
|
[PCA](https://texonom.com/pca-5f0f360d90944cccb62613910c3bf9f6)
|
## **Support vector machine**
2000s was age of SVM powerful classifier before deep learning
No direct multi-class SVM, must combine two-class SVMs but good for small dataset
w is weight, b is bias margin $\gamma$
1. ์
๋ ฅ ๋ฐ์ดํฐ๋ฅผ ์ ํ ๋๋ ๋น์ ํ์ผ๋ก ๋ถ๋ฆฌํ ์ ์๋ ์ด๊ธฐ ๊ฒฐ์ ๊ฒฝ๊ณ๋ฅผ ์ฐพ์ต๋๋ค.
2. ์ด๊ธฐ ๊ฒฐ์ ๊ฒฝ๊ณ์ ๊ฐ์ฅ ๊ฐ๊น์ด ์
๋ ฅ ๋ฐ์ดํฐ๋ค์ ์ํฌํธ ๋ฒกํฐ๋ก ์ ํํฉ๋๋ค.
3. ์ํฌํธ ๋ฒกํฐ๋ค์ ์ฌ์ฉํ์ฌ ์ด๊ธฐ ๊ฒฐ์ ๊ฒฝ๊ณ๋ฅผ ์
๋ฐ์ดํธํฉ๋๋ค.
4. 2~3๋จ๊ณ๋ฅผ ๋ฐ๋ณตํ์ฌ ์ต์ ์ ๊ฒฐ์ ๊ฒฝ๊ณ๋ฅผ ์ฐพ์ต๋๋ค.
### SVM Notion
|Title|
|:-:|
|[Support Vector](https://texonom.com/support-vector-e18f7e6ff16a4ce08d7ce4c34e3106d3)|
|[Functional Margin](https://texonom.com/functional-margin-71a665c7f96a4fed8b51b6fa70baa0ad)|
|[Geometric Margin](https://texonom.com/geometric-margin-fdbba20e3faa430f95cfe3aa263d26db)|
|[Optimal Margin Classifier](https://texonom.com/optimal-margin-classifier-5151f4af7d884f3fb4c9e1de07a59f95)|
|[Maximum Margin Classifier](https://texonom.com/maximum-margin-classifier-9e209bbbf89d4ebdba08497821d97cf7)|
|[SMO](https://texonom.com/smo-54f83167822443f4b4922a6aad94b6df)|
|[Slack Variable](https://texonom.com/slack-variable-e91b24cd50c64e96a0cebf4c81d6f273)|
|[Soft-margin Clasffier](https://texonom.com/soft-margin-clasffier-92ccd972993f4c708a755891e8b9e0a7)|
### SVMs
|Title|
|:-:|
|[TSVM](https://texonom.com/tsvm-7ec2ffe8f06f4a82b5388d5192fc0119)|
|[Latent SVM](https://texonom.com/latent-svm-29b86dc7d0624c8fb1657e4ddae72e72)|
|[Structural SVM](https://texonom.com/structural-svm-c4800f5fa5634eb59c9149c23f777682)|
> [Untitled](https://www.linkedin.com/feed/update/urn:li:activity:7082025501048918016/)
### Implementations
> [Software
โ
Kernel Machines](http://www.kernel-machines.org/software)
> [Support vector machine](https://en.wikipedia.org/wiki/Support_vector_machine)
|
8bf85b3080e2423ca43da0995a642acf
|
Functional Margin
|
SVM Notion
|
Apr 6, 2023
|
Alan Jo
|
Alan Jo
|
Jun 7, 2023
|
data input x and output y
$$\hat{y}^{(i)} = y^{(i)}(w^Tx^{(i)} + b)$$
### Margin is normalized functional margin
w๊ฐ ์๊ณ margin์ด ํด์๋ก ๋ถ๋ฅ ์ํจ
svm์์๋ w = 1๋ก ์ ๊ทํํ์ฌ ์ฌ์ฉ
y is 1 if () > 0 and y is -1 when () < 0
|
71a665c7f96a4fed8b51b6fa70baa0ad
|
|
Geometric Margin
|
SVM Notion
|
Apr 11, 2023
|
Alan Jo
|
Alan Jo
|
Jun 7, 2023
|
### Normal Version of Functional Margin
$$Geometric \; Margin = y (x^Tx + b) / ||w||$$
$$Margin = min(Geometrix \; Margin)$$
geometrically it means distance to hyperplane
|
fdbba20e3faa430f95cfe3aa263d26db
|
|
Maximum Margin Classifier
|
SVM Notion
|
May 22, 2023
|
Alan Jo
|
Alan Jo
|
Jun 7, 2023
|
9e209bbbf89d4ebdba08497821d97cf7
|
||
Optimal Margin Classifier
|
SVM Notion
|
Apr 11, 2023
|
Alan Jo
|
Alan Jo
|
Jun 7, 2023
|
[Lagrange Duality](https://texonom.com/lagrange-duality-8de2c3f1399e4765ba2543d4e8360eeb)
|
SVM์์ ์ฌ์ฉ๋๋ Classifier ์๊ณ ๋ฆฌ์ฆ ์ค ํ๋
maximize the margin
$$max_{\gamma, w, b} \gamma$$
means minimize w ([Quadratic Programming](https://texonom.com/quadratic-programming-37ba256051654908ab3b688f48ccde7d))
$$min_{w, b} \frac{1}{2}||w||^2$$
QP ์๊ณ ๋ฆฌ์ฆ์ ์ฌ์ฉํ์ฌ Optimal Margin Classifier ๊ตฌํจ
### How to compute
$$w = \Sigma_{i=1}^n\alpha_i y^{(i)}x^{(i)}$$
1. ๊ฒฐ์ ๊ฒฝ๊ณ์ ์ํฌํธ ๋ฒกํฐ๋ค ์ฌ์ด์ ๊ฑฐ๋ฆฌ, ์ฆ ๋ง์ง(margin)์ ์ต๋ํํฉ๋๋ค.
2. ์ ํ ๋๋ ๋น์ ํ ๊ฒฐ์ ๊ฒฝ๊ณ๋ฅผ ์ฐพ์ ์ ์์ต๋๋ค.
3. ์ด์์น(outlier)์ ๋ฏผ๊ฐํ์ง ์์ต๋๋ค.
We can compute Lagrangian(w, b, a) with only alpha without b and w
|
5151f4af7d884f3fb4c9e1de07a59f95
|
Slack Variable
|
SVM Notion
|
Apr 18, 2023
|
Alan Jo
|
Alan Jo
|
Jun 7, 2023
|
ํ๋ จ ์์ (training example)์ ๋ํด ์ถ๊ฐ์ ์ธ ๋ณ์๋ฅผ ๋์
ํ์ฌ, ๊ฒฐ์ ๊ฒฝ๊ณ(decision boundary)์์ ๊ฑฐ๋ฆฌ๋ฅผ ์กฐ์ ํ๋ ์ญํ
hyperparameter C, control the effect of slack variable
|
e91b24cd50c64e96a0cebf4c81d6f273
|
|
SMO
|
SVM Notion
|
Apr 18, 2023
|
Alan Jo
|
Alan Jo
|
Jun 7, 2023
|
### Sequential Minimal Optimization
support vector๋ค๋ w๊ณ์ฐ update until converge
2๊ฐ๋ฉด ๋๋๊น 2๊ฐ๋ก reoptimize ํด์ w๊ตฌํ๊ณ ๋ค์ ๊ธฐ์ธ๊ธฐ๋ก ์ํ ์ํฌํธ๋ฒกํฐ ์ต์ ํ
|
54f83167822443f4b4922a6aad94b6df
|
|
Soft-margin Clasffier
|
SVM Notion
|
Apr 18, 2023
|
Alan Jo
|
Alan Jo
|
Jun 7, 2023
|
### Normal version is Hard-margin SVM
ํฌ์ฌ์ด
hyperparameter C, control the effect of slack variable
์ผ๋ถ ์ด์์น(outlier)๋ ์๋ชป ๋ถ๋ฅ๋ ๋ฐ์ดํฐ๋ ๊ณ ๋ คํ ์ ์๊ฒ ๋๋ฉฐ
result of optimization is that the only difference is
$$0 \leq \alpha \leq C$$
|
92ccd972993f4c708a755891e8b9e0a7
|
|
Support Vector
|
SVM Notion
|
Apr 11, 2023
|
Alan Jo
|
Alan Jo
|
Jun 7, 2023
|
decision boundary์ ๊ฐ์ฅ ๊ฐ๊น์ด ์
๋ ฅ ๋ฐ์ดํฐ๋ค
Note that $\alpha_i$ will all be zero except for the support vectors because of [Karush-Kuhn-Tucker Condition](https://texonom.com/karush-kuhn-tucker-condition-01c654c038e1493080cbcb9ec1d026d0)
$$\Sigma \alpha y = 0$$
so there will be at least one + and one -
$$h_{w, b}(x) = g(w^Tx + b)$$
|
e18f7e6ff16a4ce08d7ce4c34e3106d3
|
|
Latent SVM
|
SVMs
|
Jun 6, 2023
|
Alan Jo
|
Alan Jo
|
Jun 7, 2023
|
29b86dc7d0624c8fb1657e4ddae72e72
|
||
Structural SVM
|
SVMs
|
Jun 6, 2023
|
Alan Jo
|
Alan Jo
|
Jun 7, 2023
|
c4800f5fa5634eb59c9149c23f777682
|
||
TSVM
|
SVMs
|
Mar 5, 2023
|
Alan Jo
|
Alan Jo
|
Jun 7, 2023
|
7ec2ffe8f06f4a82b5388d5192fc0119
|
||
Discriminant Analysis
|
Generative Learning Notion
|
Apr 28, 2023
|
Alan Jo
|
Alan Jo
|
Apr 28, 2023
|
### Discriminant Analysis Methods
|Title|
|:-:|
|[Linear Discriminant Analysis](https://texonom.com/linear-discriminant-analysis-37825c2c31c744998b8aeb0715164fb8)|
|[Gaussian Discriminant Analysis](https://texonom.com/gaussian-discriminant-analysis-67ec78f9fd6944e68671c610aeadb079)|
|
2ffa0dbd34204642b6fe23ed5fbf3b00
|
|
Generative Model
| null | null | null | null | null | null |
## Deep Generative Model (DGM)
Generative models aim to **[Density Estimation](https://texonom.com/density-estimation-238e2a8e51f44a67a91c0567892ab643)**** of data**
ํ์ต ๋ฐ์ดํฐ์ ๋ถํฌ์ ์ต๋ํ ์ ์ฌํ ์ฌํ๋ถํฌ์์ย **์๋ก์ด ๋ฐ์ดํฐ๋ฅผ ์์ฑ**
we can address better concern on unseen data
> [Generative AI: autocomplete for everything](https://noahpinion.substack.com/p/generative-ai-autocomplete-for-everything)
> [Generative AI: A Creative New World](https://www.sequoiacap.com/article/generative-ai-a-creative-new-world/)
> [[์ธ๊ณต์ง๋ฅ] ์์ฑ ๋ชจ๋ธ (Generative AI Model)์ด๋? (1) : AutoEncoder, VAE, GAN](https://newstellar.tistory.com/25)
> [์ง์ง AI ํ๋ช
์ด ์์๋๋ค?](https://stibee.com/api/v1.0/emails/share/BFu7h9MIxQKvkS1EPlAND_0ixxTCvio=)
|
6e5204d2982b4042847aa42e88eb8fb5
|
Naรฏve Bayes
|
Generative Learning Notion
|
Mar 30, 2023
|
Alan Jo
|
Alan Jo
|
Apr 28, 2023
|
[Data Classification](https://texonom.com/data-classification-0ed6ffa3555b4879be09f0501f0b580c)
|
handle discrete-value data, assume each x is [Conditionally Independent](https://texonom.com/conditionally-independent-8fe069550d0b4a28b031eed8293a7392)
by applying chain rule
$$p(x|y) = \Pi_ip(x_i|y)$$
Is called naive becuz of ์
๋ ฅ ๋ณ์๋ค ๊ฐ์ ๋
๋ฆฝ์ฑ์ ๊ฐ์ ํ๋ฉฐ, ์ด๋ ์ค์ ๋ฐ์ดํฐ์์๋ ๋๋ฌผ๊ฒ ์ผ์ด๋๊ธฐ ๋๋ฌธ์
๊ทธ๋ฌ๋ ์ด๋ฌํ ๊ฐ์ ๋๋ถ์ ๋ชจ๋ธ ํ์ต๊ณผ ์์ธก ์๋๊ฐ ๋นจ๋ผ์ง๋ฉฐ, ์์ ๋ฐ์ดํฐ์
์์๋ ์ข์ ์ฑ๋ฅ
[Laplace Smoothing](https://texonom.com/laplace-smoothing-3729b2d82d30420cb6540dc4fa28ab47)
|
0d56821e2bcf4504899d63c22d9a533f
|
Variational Inference
|
Generative Learning Notion
|
Jun 1, 2023
|
Alan Jo
|
Alan Jo
|
Jun 6, 2023
|
[KL Divergence](https://texonom.com/kl-divergence-c7964872f5184a7baaf312605405aef6) [Latent Variable](https://texonom.com/latent-variable-c2bb1127d53444018b2d0eaab231cc6c)
|
## Probability version of [Variational Method](https://texonom.com/variational-method-da57d88f392e4a20978a6b7a39fb7350)
์์ฑ ๋ชจ๋ธ์ ํ๋ณ ๋ชจ๋ธ์ ๋นํด intractable ๋จ์ ์ ๋ณต์ก์ฑ์ ๊ฐ๋จํ ํจ์๋ก ํก์
์ฌ์ ํ๋ฅ ๊ณผ ์ฐ๋์ ํ๋ผ๋ฉํฐ ๋ํ ์๊ณ ์์ง ๋ชปํ๋ ๊ฒฝ์ฐ
> 
> ๋ถํฌ ์ถ์ ๋ฌธ์ ๋ฅผ [Convex Optimization](https://texonom.com/convex-optimization-d3bdaf9012a34af18743ac995ca00366) ๋ฐ๊พธ์ด์ค๋ค
> 
approximate posterior $p(z|x)$ using $q(z)$ because computing [Posterior](https://texonom.com/posterior-87e0f5df288140c7ba475b0629bdda05) is hard
### Variational Inference Notion
|Title|
|:-:|
|[Variational parameter](https://texonom.com/variational-parameter-e4eaf4c0bd8046219beb3e8872f459b6)|
|[ELBO](https://texonom.com/elbo-bb7352bbce42480298d501c2ae1076d1)|
> [Variational Inference ์์๋ณด๊ธฐ - MLE, MAP๋ถํฐ ELBO๊น์ง](https://modulabs.co.kr/blog/variational-inference-intro/)
> [๋ณ๋ถ์ถ๋ก (Variational Inference) ยท ratsgo's blog](https://ratsgo.github.io/generative%20model/2017/12/19/vi/)
|
05563e34ddec467c95c9bda931770d50
|
Gaussian Discriminant Analysis
|
Discriminant Analysis Methods
|
Mar 30, 2023
|
Alan Jo
|
Alan Jo
|
May 18, 2023
|
[Normal Distribution](https://texonom.com/normal-distribution-46824b2fc69b4c5c8304b57a8967ce4c) [Logistic Regression](https://texonom.com/logistic-regression-8ed39aeb672648b1933df8911996d5d7) [Data Classification](https://texonom.com/data-classification-0ed6ffa3555b4879be09f0501f0b580c)
|
## GDA
GDA is commonly used for data classification when the data can be approximated with a normal distribution
object is to predict posterior density by the Bayes Theorem
phi $\phi$ is parameter for prior
- class prior is Bernoulli
- class conditional distribution is Normal
estimate mean, covariance becuz continuous
$$P(y|x)=P(x|y)Pprior(y)ฮฃgโYP(x|g)Pprior(g)$$
special case of MoG
> [๊ฐ์ฐ์์ ๋ถํฌ์ ๋ถ๋ณ ํจ์ (์ ํ ๋ถ๋ณ ๋ถ์(LDA), 2์ฐจ ๋ถ๋ณ ๋ถ์(QDA))](https://gaussian37.github.io/ml-concept-gaussian_discriminant/)
|
67ec78f9fd6944e68671c610aeadb079
|
Linear Discriminant Analysis
|
Discriminant Analysis Methods
|
Apr 28, 2023
|
Alan Jo
|
Alan Jo
|
Apr 28, 2023
|
## LDA
|
37825c2c31c744998b8aeb0715164fb8
|
|
Laplace Smoothing
|
Naรฏve Bayes
| null | null | null | null | null |
๊ฐ ๋ฒ์ฃผ์ pseudo-counts๋ฅผ ์ถ๊ฐํ์ฌ ์ํ๋ฅ ์ ํผํ๋ค
๊ด์ฐฐ๋์ง ์์ ํน์ง์ ํ๋ฅ ์ 0์ ๋ง๊ธฐ ์ํด ํน์ ํด๋์ค ๋ ์ด๋ธ์ ๋ํ ๊ฐ ํน์ง์ ํด๋์ค ์กฐ๊ฑด๋ถ ํ๋ฅ ์ ์ถ์ ํ๋ ๋ฐ ์ฌ์ฉ
$$P(x_i | y) = (count(x_i, y) + alpha) / (count(y) + alpha * n)$$
This can be generalized as a [MAP](https://texonom.com/map-9ecf5e2396fb4901983a57761a3e7174) estimate w/ Dirichlet priors
|
3729b2d82d30420cb6540dc4fa28ab47
|
ELBO
|
Variational Inference Notion
|
May 18, 2023
|
Alan Jo
|
Alan Jo
|
Jun 6, 2023
|
[Jensenโs Inequality](https://texonom.com/jensens-inequality-892c7bc3e16c40c897ad6549e8f0ec41)
|
## Evidence of Lower BOund
**variational lower bound**
minimize [KL Divergence](https://texonom.com/kl-divergence-c7964872f5184a7baaf312605405aef6) between $P(x|\theta)$ and $Q(x)$
$$ELBO(x, Q, \theta) = \Sigma_zQ(z)log\frac{p(x,z;\theta)}{Q(z)}$$
That is, $log \; p(x;\theta) \ge ELBO(x,Q,\theta), \forall Q, \theta, x$
> [๋ณ๋ถ์ถ๋ก (Variational Inference) ยท ratsgo's blog](https://ratsgo.github.io/generative%20model/2017/12/19/vi/)
|
bb7352bbce42480298d501c2ae1076d1
|
Variational parameter
|
Variational Inference Notion
|
Jun 6, 2023
|
Alan Jo
|
Alan Jo
|
Jun 6, 2023
|
> [Variational Inference ์์๋ณด๊ธฐ - MLE, MAP๋ถํฐ ELBO๊น์ง](https://modulabs.co.kr/blog/variational-inference-intro/)
|
e4eaf4c0bd8046219beb3e8872f459b6
|
|
Automatic Differentiation
|
ML Theory Notion
|
Apr 3, 2023
|
Alan Jo
|
Alan Jo
|
Apr 4, 2023
|
[Loss Function](https://texonom.com/loss-function-e8f6343914494828988137987cf459f9)
|
## AD
> [Automatic Differentiation: Forward and Reverse](https://jingnanshi.com/blog/autodiff.html)
> [2.5. ์๋ ๋ฏธ๋ถ(automatic differentiation) โ Dive into Deep Learning documentation](https://ko.d2l.ai/chapter_crashcourse/autograd.html)
|
368d94f0c6d848ff93a2f10f99bd4e01
|
Closed-form Solution
|
ML Theory Notion
|
Mar 9, 2023
|
Alan Jo
|
Alan Jo
|
Mar 9, 2023
|
minimize the cost explicitly without relying on an iterative algorithm
1. take [Derivation](https://texonom.com/derivation-79d0547aee0d45ce93cdae0f26bc3ce6) of [Cost Function](https://texonom.com/cost-function-66c21423cc7347909016b3423f2ada7e) with respect to the model coefficient $\theta$
2. find [Local extremum point](https://texonom.com/local-extremum-point-f7a7579a29474436b2a8670a28fed921)
### Normal Equation
$$X^TX^\theta - X^T\vec{y} = 0$$
number 1โs result is $d \times 1$ dimension if X is $n \times d$ vector (d is number of coefficient count)
Fancy but Inverse matrix is very expensive
$$\theta = (X^TX)^{-1}X^T\vec{y}$$
|
02dd2781a9334df7ba731da9b9e31ea4
|
|
Cross Validation
|
ML Theory Notion
|
Jan 26, 2020
|
Alan Jo
|
Alan Jo
|
Mar 9, 2023
|
### overfitting removal
'train set์ ๋ค์ train set + validation set์ผ๋ก ๋ถ๋ฆฌํ์ง ์๋๋ค'๋ผ๊ณ ๊ฐ์ ํ๋ฉด, ์ฐ๋ฆฌ๋ ๋ชจ๋ธ ๊ฒ์ฆ์ ์ํด์ test set์ ์ฌ์ฉํ์ฌ์ผ ํ ๊ฒ์ด๋ค
'test set์ ๊ณผ์ ํฉ ๋๋ ๋ฌธ์ '๋ test set์ด ๋ฐ์ดํฐ ์ค ์ผ๋ถ๋ถ์ผ๋ก ๊ณ ์ ๋์ด ์๊ธฐ ๋๋ฌธ์ ๋ฐ์
๋จผ์ ์ ์ฒด ๋ฐ์ดํฐ ์
์ k๊ฐ์ subset์ผ๋ก ๋๋๊ณ k๋ฒ์ ํ๊ฐ๋ฅผ ์คํํ๋๋ฐ, ์ด ๋ test set์ ์ค๋ณต ์์ด ๋ฐ๊พธ์ด๊ฐ๋ฉด์ ํ๊ฐ๋ฅผ ์งํํ๋ค.
### Cross Validation Methods
|Title|
|:-:|
|[Holdout](https://texonom.com/holdout-41cf72ee915d45688bccfc2c16195ea9)|
|[k fole cross validation](https://texonom.com/k-fole-cross-validation-55d6b33241a947fcb94278ed13ab51e2)|
> [๊ต์ฐจ ๊ฒ์ฆ(cross validation)](https://m.blog.naver.com/PostView.nhn?blogId=ckdgus1433&logNo=221599517834&categoryNo=11&proxyReferer=http%3A%2F%2Fwww.google.com%2Furl%3Fsa%3Dt%26rct%3Dj%26q%3D%26esrc%3Ds%26source%3Dweb%26cd%3D1%26ved%3D2ahUKEwjT5Pna59nmAhWyGaYKHaaTAQsQFjAAegQIBhAB%26url%3Dhttp%253A%252F%252Fm.blog.naver.com%252FPostView.nhn%253FblogId%253Dckdgus1433%2526logNo%253D221599517834%2526categoryNo%253D11%2526proxyReferer%253D%26usg%3DAOvVaw1_wYEneH8KBGG_rgcjFSiR)
|
184e7cfc80d6410188922d27ef4ef52c
|
|
Empirical Rule
|
ML Theory Notion
|
Nov 1, 2020
|
Alan Jo
|
Alan Jo
|
Mar 9, 2023
|
> [์ผ๋ฐํต๊ณ 9 - ๊ฒฝํ์ ๊ท์น(68-95-99.7 Empirical Rule)์ ์๋ฏธ์ ์ฐ์์ ์ดํดํ๊ธฐ](https://www.youtube.com/watch?v=28tbS7XAaCA)
|
31cf23b1578c42ab84e160251c3ca38b
|
|
Hyperparameter
| null | null | null | null | null | null |
### Fixed param before train
๊ฐ์ด ํ์ต ๊ณผ์ ์ ์ ์ดํ๋ ๋ฐ ์ฌ์ฉ๋๋ ํ๋ผ๋ฏธํฐ
๋ค๋ฅธ ๋งค๊ฐ๋ณ์์ ๊ฐ์ ํ๋ จ์ ํตํด ํ์
์ฌ๋์ด ์ง์ ์ค์
### Common
- learning rate
- regularization strength
- โฆ
### When you visualize the weight, it should be clean not noisy
|
ef7e34566add4e98b673d4cef59fca90
|
Kernel Trick
|
ML Theory Notion
|
Apr 6, 2023
|
Alan Jo
|
Alan Jo
|
Jun 5, 2023
|
[Kernel](https://texonom.com/kernel-13a25bedcee94977b02356d2320c490c) [Dimension Reduction](https://texonom.com/dimension-reduction-15e0474e20ea4645ae90092282391b39)
|
์
๋ ฅ ๋ฐ์ดํฐ๋ฅผ ๊ณ ์ฐจ์ ๊ณต๊ฐ์ผ๋ก ๋งคํํ์ฌ ์ ํ ํจ์๋ก ๋ชจ๋ธ๋งํ ์ ์๋๋ก
๋งคํ๋ ๊ณต๊ฐ์์์ ๋ด์ ์ฐ์ฐ์ ์ฌ์ฉํ์ฌ ์ ํ ํจ์๋ฅผ ๊ณ์ฐํฉ๋๋ค. ์ด๋ ๊ฒ ํจ์ผ๋ก์จ, ๋น์ ํ ๋ฌธ์ ๋ฅผ ํด๊ฒฐํ ์ ์์ต๋๋ค
for Non-separable case, kernel mapping increase the likelihood to find linearly separable but cannot guarantee it
Data โ Feature Map โ Kernel โ Linear Classifier โ Linear Combination
### Kernel Trick Notion
|Title|
|:-:|
|[Kernel Function](https://texonom.com/kernel-function-ff82b05f1c5c4451a683abf0b789aefb)|
|[Feature Map](https://texonom.com/feature-map-ce5417549f464c1aaed69b343eed49c3)|
|[Kernel Matrix](https://texonom.com/kernel-matrix-91a5f7e0f0a5464883b3b4d09283a723)|
|
a3d7cdc0b30d44efbb072a926b0d99aa
|
Learning Hypothesis
|
ML Theory Notion
|
Oct 18, 2020
|
Alan Jo
|
Alan Jo
|
Mar 9, 2023
|
## Set of
- (non) linear
- (non) parametric
- discriminative / generative
|
a4cc437840df4eb99cb311d1ab836258
|
|
Learning Rate
| null | null | null | null | null | null |
[Model Regularization](https://texonom.com/model-regularization-33936588cdef4041a94d0846c43ada98) tradeoff
|
c9a29280131f4f95a6191bde244f6789
|
Logistic Multinomial Classification
|
ML Theory Notion
|
Nov 1, 2020
|
Alan Jo
|
Alan Jo
|
Mar 9, 2023
|
3d7407bc18da4940bdadb5849a8cc3eb
|
||
Maximum Likelihood Estimation
|
ML Theory Notion
|
Oct 18, 2020
|
Alan Jo
|
Alan Jo
|
Mar 9, 2023
|
> [๋จธ์ ๋ฌ๋ - ์ต๋์ฐ๋์ถ์ (Maximum Likelihood Estimation)](https://www.youtube.com/watch?v=KefmJvzvcjM)
|
b30e67a4e47b4b7494aa23667adb6acd
|
|
Unknown Target Function
|
ML Theory Notion
|
Oct 18, 2020
|
Alan Jo
|
Alan Jo
|
Mar 9, 2023
|
a71dd16e5b6943f28827e946ace7914b
|
||
Holdout
|
Cross Validation Methods
|
Jan 26, 2020
| null | null | null | null |
41cf72ee915d45688bccfc2c16195ea9
|
|
k fole cross validation
|
Cross Validation Methods
|
Jan 26, 2020
| null | null | null | null |
55d6b33241a947fcb94278ed13ab51e2
|
|
Feature Map
|
Kernel Trick Notion
|
Apr 17, 2023
|
Alan Jo
|
Alan Jo
|
Jun 7, 2023
|
[Kernel](https://texonom.com/kernel-13a25bedcee94977b02356d2320c490c)
|
### vector to vector $\phi(x)$
์
๋ ฅ ๋ฐ์ดํฐ๋ฅผ ๊ณ ์ฐจ์ ํน์ง ๊ณต๊ฐ(feature space)์ผ๋ก ๋งคํํ๋ **ํจ์**
๋น์ ํ์ผ๋ก ๋งคํ๋ ๋ฐ์ดํฐ๋ 3์ฐจ์ ๊ณต๊ฐ์์ ๋น์ ํ์ ์ธ ํํ๋ฅผ ๊ฐ๊ฒ ๋๋ฉฐ, ์ ํ ๋ถ๋ฅ๊ธฐ๋ก๋ ์ ๋ถ๋ฅํ ์ ์๋ค
๋ง์ฐฌ๊ฐ์ง๋ก ์ปค๋ํจ์ ๊ฒฐ๊ณผ ๋ x, z์ Feature Map ๋ผ๋ฆฌ ๋ด์
|
ce5417549f464c1aaed69b343eed49c3
|
Kernel Function
|
Kernel Trick Notion
|
Apr 17, 2023
|
Alan Jo
|
Alan Jo
|
Jun 7, 2023
|
$$K(x,z) = <ฯ(x), ฯ(z)> = ฯ(x) \cdot ฯ(z)$$
์ด๋ฌํ ํน์ง์ ๊ฐ์ง๊ณ ์ปค๋ ํจ์๋ฅผ ์ ์ํ๋ค
์ปค๋ ํจ์๋ฅผ ์ฌ์ฉํ๋ฉด ์๋ ์
๋ ฅ ๋ฐ์ดํฐ๊ฐ ๊ฐ์ง๊ณ ์๋ ๋น์ ํ์ ์ธ ํน์ง์ linear classifier๋ก ๋ถ๋ฅํ ์ ์๊ฒ ๋ฉ๋๋ค
์
๋ ฅ ๋ฐ์ดํฐ๋ฅผ ๊ณ ์ฐจ์ feature space ์ผ๋ก ๋งคํ์ํค๋ ํจ์
์ปค๋ ํจ์๋ฅผ ์ฌ์ฉํ๋ฉด์๋ ๋น์ ํ์ ์ธ ํน์ง์ ์ ์งํ ์ ์๋ ์ด์ ๋, ๊ณ ์ฐจ์ ํน์ง ๊ณต๊ฐ์์๋ Orthogonal ๊ด๊ณ์ ์๋ ๋ฐ์ดํฐ ์์ด ๋งค์ฐ ๋๋ฌผ๊ธฐ ๋๋ฌธ
์ฌ๊ธฐ์ ์ปค๋ ํจ์๋ ์
๋ ฅ ๋ฐ์ดํฐ x์ z์ [Feature Map](https://texonom.com/feature-map-ce5417549f464c1aaed69b343eed49c3) ๊ฒฐ๊ณผ๊ฐ์ ๋ํ ๋ด์ (inner product)์ ๊ณ์ฐํ๋ ํจ์.
์ฌ๊ธฐ์ ์๋ก ์ถ๊ฐ๋๋ ์ฐจ์์ ๋ํ input์ ์์๋ก ๋๋จธ์ง ์ฐจ์์ ์ด์ฉํด ๋ง๋ค์ด์ค๋ค
Kernels also can be interpreted as similarity metrics (especially [Gaussian Kernel](https://texonom.com/gaussian-kernel-0098c457bdc049568e8daf9fda306d49))
### Kernel Functions
|Title|
|:-:|
|[Gaussian Kernel](https://texonom.com/gaussian-kernel-0098c457bdc049568e8daf9fda306d49)|
|[Polynomial Kernel](https://texonom.com/polynomial-kernel-aec7cc809f4546dab71c5bc92d176d7d)|
|
ff82b05f1c5c4451a683abf0b789aefb
|
|
Kernel Matrix
|
Kernel Trick Notion
|
Apr 17, 2023
|
Alan Jo
|
Alan Jo
|
Jun 5, 2023
|
<> means [Vector Similarity](https://texonom.com/vector-similarity-846675a408e74235acc836302110ab07) and $\phi(x)$ means vector kernel mapping function
$$K_{nm} = K_{mn} = K(x_n, x_m) = <\phi(x_n), \phi(x_m)>$$
- If K is a valid kernel, K must be symmetric
- If K is a valid kernel, K must be semi-definite
### Kernel Matrix Notion
|Title|
|:-:|
|[Mercer Kernel Theorem](https://texonom.com/mercer-kernel-theorem-cebcf7555bf44413907751aa58633ee0)|
|
91a5f7e0f0a5464883b3b4d09283a723
|
|
Gaussian Kernel
|
Kernel Functions
|
Apr 17, 2023
|
Alan Jo
|
Alan Jo
|
Apr 17, 2023
|
$$K(x, z) = exp(-ฮณ ||x - z||^2)$$
ฮณ๋ ํ์ดํผํ๋ผ๋ฏธํฐ(hyperparameter)๋ก์, ์
๋ ฅ ๋ฐ์ดํฐ x์ z ์ฌ์ด์ ๊ฑฐ๋ฆฌ๋ฅผ ์ผ๋ง๋ ์ค์ํ๊ฒ ์๊ฐํ ์ง๋ฅผ ๊ฒฐ์ \
๊ฐ์ฐ์์ ์ปค๋์ 0์ ๋ชป๋ง๋๋๊น ๋ ๋ฒกํฐ๋ฅผ orthogonalํ๊ฒ ๋ชป๋ณํ
|
0098c457bdc049568e8daf9fda306d49
|
|
Polynomial Kernel
|
Kernel Functions
|
Apr 17, 2023
|
Alan Jo
|
Alan Jo
|
Apr 17, 2023
|
aec7cc809f4546dab71c5bc92d176d7d
|
||
Mercer Kernel Theorem
|
Kernel Matrix Notion
|
Apr 17, 2023
|
Alan Jo
|
Alan Jo
|
Jun 5, 2023
|
[Symmetric positive semi-definite](https://texonom.com/symmetric-positive-semi-definite-16ce5c2766e94f3d803ea6ddf2742660)
|
1. Symmetric: K(x, z) = K(z, x)
2. Positive semi-definite: for any finite set of points {x1, x2, ..., xn}, the corresponding kernel matrix K must be symmetric and positive semi-definite
|
cebcf7555bf44413907751aa58633ee0
|
Learn Machine Learning
|
Machine Learning Usages
|
Jul 6, 2021
|
Alan Jo
|
Alan Jo
|
Jun 6, 2023
|
[Linear Algebra](https://texonom.com/linear-algebra-a5879da463c442d89c51891d124f97ac) [Statistics](https://texonom.com/statistics-561b1e8231f44a45bb6284033db682fb) [Optimization](https://texonom.com/optimization-984100dcbc454057bffd1b97be6ef134) [Calculus](https://texonom.com/calculus-8e37e7dac61e435da6a6970da7296f73)
|
### Korean Study
> [0000 Machine Learning Basic - Deepest Documentation](https://deepestdocs.readthedocs.io/en/latest/000_machine_learning/0000_machine_learning_basic/)
> [๊ฒฝ์ฌํ๊ฐ๋ฒ(gradient descent) - ๊ณต๋์ด์ ์ํ์ ๋ฆฌ๋
ธํธ](https://angeloyeo.github.io/2020/08/16/gradient_descent.html)
> [PRML](http://norman3.github.io/prml/)
|
1c6270ed8f544e52b0fa19044a459c6a
|
Machine Learning Method
|
Machine Learning Usages
|
May 13, 2021
|
Alan Jo
|
Alan Jo
|
Jul 6, 2023
|
[Parameter Estimation](https://texonom.com/parameter-estimation-2c5667f2707b4d3082e96a5928464029) [Machine Learning Theory](https://texonom.com/machine-learning-theory-c69317fb5dc24f05a29532fa27d27cb1)
|
### Machine Learning Techniques
|Title|
|:-:|
|[ML Compiler Optimization](https://texonom.com/ml-compiler-optimization-011e7bd0ba8f417bb111ec5ea2171c8e)|
|[Quantum Machine Learning](https://texonom.com/quantum-machine-learning-8f5276045f4d43b8b96f3b4ec6646f66)|
|[Parallel Training](https://texonom.com/parallel-training-4a8896bc837b4dddb47c3700b715cdc8)|
|[ Weight Initialization](https://texonom.com/weight-initialization-6cfc10eb06f948528aa76a9814a9ac85)|
### Machine Learning Methodology
|Title|
|:-:|
|[MLOps](https://texonom.com/mlops-70d0a9a4ab154f528d364f0198f064d2)|
|[Auto ML](https://texonom.com/auto-ml-2e283b8fb3b34a148979fe3feefa53f9)|
|[Ray Architecture](https://texonom.com/ray-architecture-8c8b1acd717c4989a65230bc90d92b7f)|
|[Pathway Architecture](https://texonom.com/pathway-architecture-70fd437c1f71460a9f31b28d690cc90c)|
|[Vessl](https://texonom.com/vessl-467995baadef4f4e94bdfb8f236157f9)|
### Machine Learning Methods
|Title|
|:-:|
|[Supervised Learning](https://texonom.com/supervised-learning-eaf91e20b3d04aadb4fc3ea512c73aa1)|
|[Unsupervised learning](https://texonom.com/unsupervised-learning-8cb6e253bfa845b5931d22963ea93019)|
|[Reinforcement Learning](https://texonom.com/reinforcement-learning-3e6e1e5b806941b6b586ce2c3d1bbd02)|
|
4d27fd7ddf4247e6b98df2ac09c2e25a
|
Machine Learning Tool
|
Machine Learning Usages
|
Aug 5, 2021
|
Alan Jo
|
Alan Jo
|
Mar 5, 2023
|
[AI Library](https://texonom.com/ai-library-30022d621bfd4843b908559bf1f6c1ac)
|
### Machine Learning Tools
|Title|
|:-:|
|[ML Accelerator](https://texonom.com/ml-accelerator-0df3a271237b4e63a342aa7ce704870d)|
|[ML Analyze Tool](https://texonom.com/ml-analyze-tool-5aeb48a9850245eb97e20ec56448a15f)|
|[ML Platform](https://texonom.com/ml-platform-9d4142db8db042ed9e4a79085348cc55)|
|[Function Transformers](https://texonom.com/function-transformers-c4396d81a01a4425ba0c6702501c911a)|
|[ML Feature Store](https://texonom.com/ml-feature-store-f53069a7083b4719b1e5fab18a5a9bbd)|
|[ML Container Tool](https://texonom.com/ml-container-tool-14223d2021c748268fba092dee1fa357)|
> [ํ ์คใ
ฃSLASH 22 - ๋ฌผ ํ๋ฅด๋ฏ ์์ฐ์ค๋ฌ์ด ML ์๋น์ค ๋ง๋ค๊ธฐ](https://youtu.be/EEsYbiqqcc0)
|
4183d7b4a9d44149932142731c91ca1d
|
Machine Unlearning
|
Machine Learning Usages
|
Jul 10, 2023
|
Alan Jo
|
Alan Jo
|
Jul 10, 2023
|
> [Announcing the first Machine Unlearning Challenge](https://ai.googleblog.com/2023/06/announcing-first-machine-unlearning.html)
|
7b6060981f15488e8708533a77b88a2d
|
|
ML Meta Algorithm
|
Machine Learning Usages
|
Oct 6, 2021
|
Alan Jo
|
Alan Jo
|
Mar 7, 2023
|
### ML Meta Algorithms
|Title|
|:-:|
|[AdaBoost](https://texonom.com/adaboost-79feac6204384e7299e08e3cfa40d05e)|
|
29290f722f8c442a9e8c9a3528420be1
|
|
Neural Network
|
Machine Learning Usages
|
Nov 5, 2019
|
Alan Jo
|
Seong-lae Cho
|
Jul 6, 2023
|
[Synapse](https://texonom.com/synapse-955f5ab318f54b85a23f1240e9b93a0d) [Computational Graph](https://texonom.com/computational-graph-51f54bce30ed44569be80f56d596340f) [Back Propagation](https://texonom.com/back-propagation-18f4493692ad43449d4271f1bb293781)
|
## N**arrowly defined AI**
### [Data Representation](https://texonom.com/data-representation-81d5d9fe876c4c95a6bb3cdbc01fab2d) + [Data Classification](https://texonom.com/data-classification-0ed6ffa3555b4879be09f0501f0b580c) + [Algorithm](https://texonom.com/algorithm-678d8c51d1d3479badb557af87762562)
### Neural Network Notion
|Title|
|:-:|
|[Neural Network History](https://texonom.com/neural-network-history-2127bdb56da54a659e570f47704e40b1)|
|[Deep Learning](https://texonom.com/deep-learning-7d3c8b9ce05b49cf9eed92dbcdc80cfd)|
|[Neural Network Structure](https://texonom.com/neural-network-structure-400bbea8029c4eb1a97c0dd063735551)|
### Visualization
> [Tensorflow โ Neural Network Playground](https://playground.tensorflow.org/)
### **Zero to Hero**
> [Neural Networks: Zero To Hero](https://karpathy.ai/zero-to-hero.html)
|
86f54f9f1de848c1a29c56c24f7d5094
|
Auto ML
|
Machine Learning Methodology
|
Sep 5, 2020
|
Alan Jo
|
Alan Jo
|
Apr 19, 2023
|
### Data Prep โ Feature Engineering โ Modeling โ HyperParameter Tuning
can pick us tool
### Auto MLs
|Title|
|:-:|
|[Microsoft NNi](https://texonom.com/microsoft-nni-5df59b29fe4c4dcba6f738bf2e89e1df)|
> [[6์ฃผ์ฐจ] Introduction to Meta Learning](https://velog.io/@tobigs_xai/6์ฃผ์ฐจ-Introduction-to-Meta-Learning)
> [google-research/google-research](https://github.com/google-research/google-research/tree/master/automl_zero)
|
2e283b8fb3b34a148979fe3feefa53f9
|
|
MLOps
|
Machine Learning Methodology
|
Dec 26, 2020
|
Alan Jo
|
Alan Jo
|
Mar 8, 2023
|
[MLOPs-Primer](https://github.com/dair-ai/MLOPs-Primer)
|
## Machine Learning + Operation
### MLOps Usages
|Title|
|:-:|
|[MLOps Framework](https://texonom.com/mlops-framework-3a7b16d0dd284c7d90f3ae7ceb288b81)|
> [Online gradient descent written in SQL โข Max Halford](https://maxhalford.github.io/blog/ogd-in-sql/)
|
70d0a9a4ab154f528d364f0198f064d2
|
Pathway Architecture
|
Machine Learning Methodology
|
Feb 9, 2022
|
Alan Jo
|
Alan Jo
|
Mar 5, 2023
|
### Google Search AI Architecture
> [Introducing Pathways: A next-generation AI architecture](https://blog.google/technology/ai/introducing-pathways-next-generation-ai-architecture/)
|
70fd437c1f71460a9f31b28d690cc90c
|
|
Ray Architecture
|
Machine Learning Methodology
|
Sep 11, 2020
|
Alan Jo
|
Alan Jo
|
Mar 5, 2023
|
Ray๋ ๋จ์ผ ๋์ ์์
๊ทธ๋ํ์์ ์์
๋ณ๋ ฌ ๋ฐ Actor ํ๋ก๊ทธ๋๋ฐ ๋ชจ๋ธ์ ํตํฉํ๊ณ ๊ธ๋ก๋ฒ ์ ์ด ์ ์ฅ์ ๋ฐ ์ํฅ์ ๋ถ์ฐ ์ค์ผ์ค๋ฌ๋ฅผ ํตํด ์ง์๋๋ ํ์ฅ ๊ฐ๋ฅํ ์ํคํ
์ฒ๋ฅผ ์ฌ์ฉ
์ํคํ
์ฒ์์ ๋์์ ๋ฌ์ฑ๋๋ ํ๋ก๊ทธ๋๋ฐ ์ ์ฐ์ฑ, ๋์ ์ฒ๋ฆฌ๋ ๋ฐ ๋ฎ์ ๋๊ธฐ ์๊ฐ์ ๋ฆฌ์์ค ์๊ตฌ ์ฌํญ, ๊ธฐ๊ฐ ๋ฐ ๊ธฐ๋ฅ์ด ๋ค์ํ ์์
์ ์์ฑํ๋ ์๋ก์ด ์ธ๊ณต ์ง๋ฅ ์ํฌ๋ก๋์ ํนํ ์ค์
Ray๋ ๋ฏธ๋์ AI ์ ํ๋ฆฌ์ผ์ด์
๊ฐ๋ฐ์ ์ํด ์ ์ฐ์ฑ, ์ฑ๋ฅ ๋ฐ ์ฌ์ฉ ํธ์์ฑ์ ๊ฐ๋ ฅํ ์กฐํฉ์ ์ ๊ณต
> [Ray ์ํคํ
์ณ ๋ฐฑ์ (๋ฒ์ญ)](https://brunch.co.kr/@chris-song/106)
|
8c8b1acd717c4989a65230bc90d92b7f
|
|
Vessl
|
Machine Learning Methodology
|
Apr 3, 2023
|
Alan Jo
|
Alan Jo
|
May 17, 2023
|
[AI ํด๋ฌ์คํฐ VESSL ์ฌ์ฉ์ ๋งค๋ด์ผ](https://texonom.com/ai-vessl-1d134eb8f9f440f78ef87c08761ebe34)
[VESSL for Academics](https://texonom.com/vessl-for-academics-fa47bf5e69b44e92b5daaead758cb057)
### Canโt run docker inside workspace
> [Tips & Limitations](https://docs.vessl.ai/user-guide/workspace/tips-and-limitations)
### Wroksapce custom image
```typeRUN apt-get install -y -q \ libsqlite3-dev \ build-essential```
> [Building Custom Images](https://docs.vessl.ai/user-guide/workspace/building-custom-images)
|
467995baadef4f4e94bdfb8f236157f9
|
|
Microsoft NNi
|
Auto MLs
|
Sep 5, 2020
| null | null | null | null |
## Neural Network Intelligence
automate machine learning lifecycle
### Feature
- Feature Engineering
- Neural Architecture Search
- Hyperparameter Tuning
- Model Comperssion
### Windows
```type# installpython -m pip install --upgrade nni# testgit clone -b v1.8 https://github.com/Microsoft/nni.gitnnictl create --config nni/examples/trials/mnist-tfv1/config_windows.yml --port 9999```
### Github
> [microsoft/nni](https://github.com/microsoft/nni)
> [AutoML ๋๊ตฌ Microsoft NNI](https://eunguru.tistory.com/241)
|
5df59b29fe4c4dcba6f738bf2e89e1df
|
MLOps Framework
|
MLOps Usages
|
Oct 9, 2021
|
Alan Jo
|
Alan Jo
|
Oct 9, 2021
|
### MLOps Frameworks
|Title|
|:-:|
|[ZenML](https://texonom.com/zenml-db8d78d303ba4e5c90e7cf4a99bcb7b0)|
|[Open MLOps](https://texonom.com/open-mlops-78dadc74bbcd4609bc7f1d7267ea9fa7)|
|[WanDB](https://texonom.com/wandb-1c57514e8a1f476f8451738d3486d77a)|
|
3a7b16d0dd284c7d90f3ae7ceb288b81
|
|
Open MLOps
|
MLOps Frameworks
|
Oct 9, 2021
|
Alan Jo
|
Alan Jo
|
Feb 18, 2022
|
[GNPS](https://texonom.com/gnps-a1f892efaeec4950ae30b7fcf90cd09d)
> [Production ML solution for MS using Open MLOps](https://www.datarevenue.com/en-blog/production-machine-learning-omigami-open-mlops)
> [GitHub - datarevenue-berlin/OpenMLOps](https://github.com/datarevenue-berlin/OpenMLOps)
|
78dadc74bbcd4609bc7f1d7267ea9fa7
|
|
WanDB
|
MLOps Frameworks
|
Apr 25, 2023
|
Alan Jo
|
Alan Jo
|
Apr 25, 2023
|
> [Pytorch wandb (Weight & Biases) ์ ์ฉ](https://velog.io/@nawnoes/Pytorch-Wandb-Weight-Biases-์ ์ฉ)
|
1c57514e8a1f476f8451738d3486d77a
|
|
ZenML
|
MLOps Frameworks
|
Dec 26, 2020
|
Alan Jo
|
Alan Jo
|
Oct 9, 2021
|
[ZenML Init](https://texonom.com/zenml-init-47e9d508a53f401ea8838e02464e5128)
> [maiot-io/zenml](https://github.com/maiot-io/zenml)
|
db8d78d303ba4e5c90e7cf4a99bcb7b0
|
|
GNPS
|
Open MLOps
| null | null | null | null | null |
> [Collaborative MS Analysis with the GNPS Dashboard](https://www.datarevenue.com/en-blog/gnps-dashboard)
|
a1f892efaeec4950ae30b7fcf90cd09d
|
ZenML Init
|
ZenML
| null | null | null | null | null |
```typepip install zenmlcd repozenml init```
> [maiot-io/zenml](https://github.com/maiot-io/zenml)
|
47e9d508a53f401ea8838e02464e5128
|
Reinforcement Learning
|
Machine Learning Methods
|
Nov 5, 2019
|
Alan Jo
|
Seong-lae Cho
|
Jul 15, 2023
|
[gym](https://github.com/openai/gym) [Reinforcement](https://texonom.com/reinforcement-a04826d130484e65b22afbf08ca96800) [baselines](https://github.com/openai/baselines)
|
### Its label is a real-valued reward signal (possibly delayed)
์ด๋ค ํ๊ฒฝ ์์์ ์ ์๋ ์์ด์ ํธ๊ฐ ํ์ฌ์ ์ํ๋ฅผ ์ธ์ํ์ฌ, ์ ํ ๊ฐ๋ฅํ ํ๋๋ค ์ค ๋ณด์์ ์ต๋ํํ๋ ํ๋ ํน์ ํ๋ ์์๋ฅผ ์ ํํ๋ ๋ฐฉ๋ฒ
์์ด์ ํธ๊ฐ ํ๊ฒฝ๊ณผ ์ํธ์์ฉํ์ฌ ๋ฐ์ดํฐ๋ฅผ ์์ง ํน์ human feedback
Special form of Supervised learning
- Agent
- Environment
- Action
- State
- Reward
<details><summary>
**Basic idea - **random** **action โcheck reward โ observation given state โ learn โ make policy</summary>
- Receive feedback in the form of rewards
- Agentโs utility is defined by the reward function
- Must (learn to) act so as to maximize expected rewards
- All learning is based on observed samples of outcomes
</details>
### Reinforcement Learning Notion
|Title|
|:-:|
|[Markov decision process](https://texonom.com/markov-decision-process-d01c5c23d56648ce843727514008e25c)|
|[RL Exploration](https://texonom.com/rl-exploration-3b4038d0640f4fe184558e3d477cbec2)|
|[RL Exploitation](https://texonom.com/rl-exploitation-caee227c108e4139b66ec1e7f71ab80c)|
|[Reinforcement Learning Term](https://texonom.com/reinforcement-learning-term-c0e49da62c5a4690b9e6d653fc63d45a)|
|[Passive reinforcement learning](https://texonom.com/passive-reinforcement-learning-5ff410c59608455380eb53627655fc1b)|
|[Evolutionary Learning](https://texonom.com/evolutionary-learning-b63b9a0ef63845d3bfea5b9e0abb9692)|
|[Model based learning](https://texonom.com/model-based-learning-7e1b7de3d4b54e7a8281dc4e5cd6dbb4)|
|[Policy Iteration](https://texonom.com/policy-iteration-55cbe9219f8a48fbb850f64b677e847a)|
|[Bellmanย Expectation Equation](https://texonom.com/bellmanexpectation-equation-1135fc0eb9bb46388bb1453399d8cca6)|
|[State-value function](https://texonom.com/state-value-function-d9d3467fea994443b2336fd7958dc00b)|
|[Q Function](https://texonom.com/q-function-115a8a9d62554ebf9fbb46efb3f263ca)|
### Reinforcement Learnings
|Title|
|:-:|
|[RLHF](https://texonom.com/rlhf-4b184f9c9e8b4c7a8861fb6374e91aa6)|
|[RRHF](https://texonom.com/rrhf-073c6bdf48444a1b9acad9e65057f3d0)|
|[DQN](https://texonom.com/dqn-d3f9810ec70640a0b940b2d0c214adbb)|
|[PPO](https://texonom.com/ppo-87ce8ebe81f84ac1a7617b1f6def9e26)|
|[ACER](https://texonom.com/acer-a0eafd5e768c4f62955b12fc428e8719)|
|[TRPO](https://texonom.com/trpo-23c0f917124d419cbfe478d20b804276)|
|[A3C](https://texonom.com/a3c-eb1de7310234443d986317171949eaff)|
|[SARSA](https://texonom.com/sarsa-e5d847a4fb6e41cdad5b9ffb6e974a10)|
|[Q-Learning](https://texonom.com/q-learning-6fb81e4a53ab4e3097784cde99c8c038)|
> [What is Reinforcement Learning ยท Fundamental of Reinforcement Learning](https://dnddnjs.gitbooks.io/rl/content/what_is_reinforcement_learning.html)
|
3e6e1e5b806941b6b586ce2c3d1bbd02
|
Supervised Learning
| null | null | null | null | null | null |
## Each example in a dataset is labeled in advance
pairs of output y, input x $p(y|x, \theta)$. But [Data Labeling](https://texonom.com/data-labeling-dda48db348de44a9bafa1ae69ca18133) is expensive
|
eaf91e20b3d04aadb4fc3ea512c73aa1
|
Unsupervised learning
| null | null | null | null | null | null |
### x is output $p(x|\theta)$
How data is composed
> [Using Large Language Models Effectively](https://unsupervisedlearning.substack.com/p/using-large-language-models-effectively)
|
8cb6e253bfa845b5931d22963ea93019
|
**Bellmanย Expectation Equation**
|
Reinforcement Learning Notion
|
Jul 18, 2023
|
Alan Jo
|
Alan Jo
|
Jul 18, 2023
|
### recursive equation that relates the value of a state to the values of its neighboring states
$$V(s)=โaโฯ(aโฃs)โsโฒ,rโp(sโฒ,rโฃs,a)[r+ฮณV(sโฒ)]$$
Reinforcement learning targets to solve this equation
๋ชฉํ๋ย value func.์ ์ฐธ ๊ฐ์ ์ฐพ๋ ๊ฒ์ด ์๋ย **์ต๋์ reward๋ฅผ ์ป๋ policy๋ฅผ ์ฐพ๋ ๊ฒ**
ํ**์ฌ ์ํ์ ๊ฐ์นํจ์์ ๋ค์ ์ํ์ ๊ฐ์นํจ์ ์ฌ์ด์ ๊ด๊ณ**
**์ํ๊ฐ ๋ง์์ง์๋ก ์๋นํ ๋นํจ์จ์ ์ธ ๋ฐฉ๋ฒ**
> [(3) ๊ฐ์นํจ์์ ๋ฒจ๋ง๋ฐฉ์ ์](https://jang-inspiration.com/bellman-equation)
> [Untitled](https://sumniya.tistory.com/5)
|
1135fc0eb9bb46388bb1453399d8cca6
|
|
Evolutionary Learning
|
Reinforcement Learning Notion
|
Nov 12, 2021
|
Alan Jo
|
Alan Jo
|
Jul 15, 2023
|
[DERL](https://texonom.com/derl-f01c4bd87fe447f8871b24ca09e4b654)
|
b63b9a0ef63845d3bfea5b9e0abb9692
|
|
Markov decision process
|
Reinforcement Learning Notion
|
Mar 5, 2023
|
Alan Jo
|
Alan Jo
|
Jul 18, 2023
|
## Find most effective Policy from random
Reinforcement Learning is process of resolving MDP
> New twist - donโt know R and T (different to traditional MDP) โ **(Step 1)**
>> โ Must actually try actions and states out to learn
> - A set of states s โS
> - A set of actions (per state) A
> - A **model T(s,a,sโ) **(probability)
> - A **reward function** **R(s,a,sโ)**
> Bellman equation - do not need to know environment model use action -value function
### MDP
> [OpenRL - ๊ฐํํ์ต ๊ทธ๋ฆฌ๊ณ OpenAI - 2: Intro to Reinforcement Learning (1) MDP &Value Function](http://www.modulabs.co.kr/RL_library/2136)
> MDP is Markov Decision Process and tuple of states, actions, station probability matrix (probability to go certain state to another state by certain action), reward function, discounted factor
|
d01c5c23d56648ce843727514008e25c
|
|
Model based learning
|
Reinforcement Learning Notion
|
Nov 5, 2019
|
Alan Jo
|
Seong-lae Cho
|
Jul 15, 2023
|
- Learn an approximate model based on experiences
- Solve for values as if the learned model were correct?
> Model free learning do not instersted in R and T
**Step 1: Learn empirical MDP model**
- Count outcomes sโ for each s, a
- Normalize to give an estimate of `math: T\hat(s, a, s')` - ๋จ์ํ๋ฅ ํ๊ท of action
**Step 2: Solve the learned MDP**
use value iteration, as before
> State value function - sum of rewards get from states passed (all time)
> Action value function - expected reward from action in this state (that time)
|
7e1b7de3d4b54e7a8281dc4e5cd6dbb4
|
|
**Passive reinforcement learning**
|
Reinforcement Learning Notion
|
Nov 5, 2019
|
Alan Jo
|
Seong-lae Cho
|
Jul 15, 2023
|
Do not consider T and R โ simplified task - **just evaluate policy by state value**
<details><summary>
**So goal is compute value of each state under policy (input is policy โ (episodes training by observation) โ output is value per state)**</summary>
<details><summary>
**direct evaluation** - just sum of after state's value and divide by episode number</summary>
</details>
<details><summary>
**Sample-based policy evaluation **- improve V per transition by estimate T and R by sample of outcome</summary>
<details><summary>
**Temporal difference learning** - ๊ทธ์ํ๋ง ๊ณ ๋ ค</summary>
</details>
<details><summary>
**Exponential moving average** - ์ต๊ทผ๊บผ ๋ ์ค์ํ๊ฒ ๊ณ ๋ ค</summary>
</details>
<details><summary>
**Active reinforcement learning **- ์คํด๋ก ํํ๋ฉด์</summary>
> Fundamental tradeoff: exploration vs. exploitation - since it is online playing
</details>
<details><summary>
**Q-learning **- exponential + ์ ๋ถ๋ค = Q-value-iteration</summary>
</details>
</details>
</details>
> Optimal value function means high Q value
> Q value is prediction of all future R by action
> V value is prediction of all future R in state
> Policy is action per state which maximize V
> Alpha means learning rate
> So most reinforce learning is Q - MDP is not used - just for understanding
|
5ff410c59608455380eb53627655fc1b
|
|
****Policy Iteration****
|
Reinforcement Learning Notion
|
Jul 18, 2023
|
Alan Jo
|
Alan Jo
|
Jul 18, 2023
|
[Dynamic Programming](https://texonom.com/dynamic-programming-810a9aa590434f02b9fe4d869fb9c371)
|
## Generalized Policy Iteration (GPI)
### 1. Policy Evaluation
### 2. Policy Improvement
### Limitation
- Computation Complexity
- Canโt get perfect information about the environment
> [(5) ์ ์ฑ
์ดํฐ๋ ์ด์
, ๊ฐ์น ์ดํฐ๋ ์ด์
](https://jang-inspiration.com/policy-value-iteration)
|
55cbe9219f8a48fbb850f64b677e847a
|
Q Function
|
Reinforcement Learning Notion
|
Jul 18, 2023
|
Alan Jo
|
Alan Jo
|
Aug 31, 2023
|
### value function
Generalizing Across States - a table of all q-values hard โ Too many state โ **Generalize that experience to new, similar situations โ Feature-Based Representations**
This is a fundamental idea in machine learning, and weโll see it over and over again
1. linear value function
> 
> 
> + least squares or regression then minimize error
ํ๋๊ฐ์นํจ์
**์ด๋ค ์ํ์์ ์ด๋ค ํ๋์ด ์ผ๋ง๋ ์ข์์ง**
> [(3) ๊ฐ์นํจ์์ ๋ฒจ๋ง๋ฐฉ์ ์](https://jang-inspiration.com/bellman-equation)
|
115a8a9d62554ebf9fbb46efb3f263ca
|
|
Reinforcement Learning Term
|
Reinforcement Learning Notion
|
Mar 5, 2023
|
Alan Jo
|
Alan Jo
|
Jul 15, 2023
|
### States
- environment state - environment representation
- **agent state** - agent representation - most used state
- information state (Markov state) - probability from start to this state = probability from previous state to this state โ state is Markov (independant)
### Others
- history - sequence of observation, action, reward
- value iteration
> 
- policy iteration
- MDP - policy and value
> 
> Both value iteration and policy iteration compute the same thing (all optimal values)
|
c0e49da62c5a4690b9e6d653fc63d45a
|
|
RL Exploitation
|
Reinforcement Learning Notion
|
Nov 5, 2019
|
Alan Jo
|
Seong-lae Cho
|
Jul 15, 2023
|
**Regret** - Regret is a measure of your total mistake cost
= the difference between your (expected) rewards, including youthful sub-optimal, and optimal (expected) rewards
> policy evaluation is prediction in policy iteration
> policy improvement is control in policy iteration
|
caee227c108e4139b66ec1e7f71ab80c
|
|
RL Exploration
|
Reinforcement Learning Notion
|
Nov 5, 2019
|
Alan Jo
|
Seong-lae Cho
|
Jul 15, 2023
|
Simplest: random actions (ฮต-greedy)
> With (small) probability ฮต, act randomly (if random is small then threshold) โ** threshold become lower greedly by learning **โ become zero
> With (large) probability 1-ฮต, act on current policy
> - can keep thrashing around once learning is done
>> One solution: lower ฮตover time **(decaying epsilon greedy) **
>> explore areas whose badness is not (yet) established **(optimism for uncertainty)**
|
3b4038d0640f4fe184558e3d477cbec2
|
|
State-value function
|
Reinforcement Learning Notion
|
Jul 18, 2023
|
Alan Jo
|
Alan Jo
|
Jul 18, 2023
|
## V-function
expected return starting from a particular state under a given policy
$$V(s) = E[G |S=s]$$
|
d9d3467fea994443b2336fd7958dc00b
|
|
DERL
|
Evolutionary Learning
| null | null | null | null | null |
## Deep Evolutionary Reinforcement Learning)
[Unimal](https://texonom.com/unimal-1dc7297fd5014e0b81b8de92950c0dce)
๊ฐํํ์ต์ ํ ๋๋ก ๋ค์ํ ์งํ์ ์ด๋ํ ์ ์๊ณ ,ย ๋ฌผ์ฒด๋ฅผ ์์ง์ด๋๋ฐ ํ์ํ ์ ์ฒด ์์ฑ
> [Stanford University Deep Evolutionary RL Framework Demonstrates Embodied Intelligence via Learning...](https://medium.com/syncedreview/stanford-university-deep-evolutionary-rl-framework-demonstrates-embodied-intelligence-via-learning-686c63e18dc9)
|
f01c4bd87fe447f8871b24ca09e4b654
|
Unimal
|
DERL
| null | null | null | null | null |
### Universe + Animal
|
1dc7297fd5014e0b81b8de92950c0dce
|
**A3C**
|
Reinforcement Learnings
|
Jun 17, 2023
|
Alan Jo
|
Alan Jo
|
Jul 15, 2023
|
[Deepmind](https://texonom.com/deepmind-5eb171c77b344d4786a9a5b23ae70eca)
|
### Asynchronous Advantage actor-critic
|
eb1de7310234443d986317171949eaff
|
ACER
|
Reinforcement Learnings
|
Jul 15, 2023
|
Alan Jo
|
Alan Jo
|
Jul 15, 2023
|
2016๋
11์, DeepMind.
> [Sample Efficient Actor-Critic with Experience Replay](https://arxiv.org/abs/1611.01224)
|
a0eafd5e768c4f62955b12fc428e8719
|
|
**DQN**
|
Reinforcement Learnings
|
Jun 17, 2023
|
Alan Jo
|
Alan Jo
|
Jul 15, 2023
|
[Deepmind](https://texonom.com/deepmind-5eb171c77b344d4786a9a5b23ae70eca)
|
d3f9810ec70640a0b940b2d0c214adbb
|
Subsets and Splits
No community queries yet
The top public SQL queries from the community will appear here once available.