給創作者的機器學習入門 Basic Machine Learning for Creators
0 likes205 views
This document provides an introduction to machine learning for creators and discusses how creators can collaborate with machine learning systems. It covers several machine learning techniques including supervised learning methods like regression, classification, and support vector machines. It emphasizes that machine learning can empower and support human creators by enabling faster creation, more exploration, and personalized experiences rather than trying to make machines better than humans. The goals of machine learning in arts are different than conventional applications, focusing more on understanding small datasets and enabling real-time interactive systems over predictive accuracy.
給創作者的機器學習入門 Basic Machine Learning for Creators
- 1. 給創作者的機器學習入⾨門 Creators co-work with Machine
- 2. MAIN CONCERNS & KEYWORDS RHYTHM ANALYSIS INTERFACE GENERATIVE MACHINE STRUCTURE REAL-TIME KINETIC OBJECTS AUTOMATED COLLECTIVE SELF- ORGANIZATION SONIFICATIO DATA TIME&SPACE
- 3. BY DESIGNING SYSTEMS THAT MAKE MUSIC. MAKING MUSIC
- 4. BY DESIGNING THE SYSTEM ▸ On one hand… - you give up some controls by leaving artistic decisions unmade ▸ On the other hand… - you still has the ultimate control, as a system designer. They design the generative space the listener explores. ▸ Creator co-works with the software, instead of making a complete control.
- 5. 所以，到底什什麼是「機器學習」？ • 由⼈人去「教」機器該怎麼執⾏行行 • 不同於「設計整個⽣生成系統」，你設計規則 • 創作者能透過展⽰示某種期望被⽣生成的範例例， 讓機器去分析與模仿
- 6. "EVOLVING METAPHORS, IN MY OPINION, IS WHAT ARTISTS DO.” Brian Eno SAN FRANCISCO, JUNE 8, 1996
- 7. BASIC MACHINE LEARNING EXPERIMENTS PREVIOUS EXAMPLES
- 8. WHAT CAN I TEACH TO THE SYSTEM? SIMPLE MACHINE LEARNING PRACTICES ON VARIOUS PURPOSES
- 9. • Inputs could be sent from any real-time input, such as game controllers, webcam, motion tracking, audio input, sensors connected to Arduino, etc. • Outputs could be sent to any real-time process, such as music (Max/MSP, ChucK, PD, SuperCollider, Ableton/Max4Live…), animation (Processing, OpenFrameworks, …), games (Unity3D), robots or physical actuators (e.g. using Arduino), etc.
- 12. MACHINE LEARNING 101 監督式學習 (SUPERVISED LEARNING) ▸ 由訓練資料中學到或建立⼀一個模式（function / learning model），並依此模式推測新的實例例 (instances)。 ▸ 訓練資料（training data）是由輸入物件（通常是向量量）和預 期輸出所組成。函數的輸出，可以是⼀一個連續的值（稱為迴 歸），或是⼀一個預測分類標籤（稱作分類）。 ▸ 線性(linear)及非線性(non-linear)資料 ▸ 監督式學習者的任務，是在觀察完⼀一些訓練範例例（輸入和預期 輸出）後，去預測這個函數對任何可能輸入值的輸出結果。
- 13. MACHINE LEARNING 101 REGRESSION ▸ 迴歸分析。依照所輸入的資訊，輸出連續的結果。 ▸ 最⼩小平⽅方差(Least square regression) ▸ PRACTICE: Draw a line!
- 15. MACHINE LEARNING 101 CLASSIFICATION ▸ 分類法，亦即把輸入做項⽬目分類。 ▸ Nearest-neighbor & Decision stumps. ▸ 聲⾳音相關應⽤用：觸發⾳音檔、偵測風格、判斷說話者性別等 ▸ Occam’s Razor
- 20. Is “overfit” a problem?
- 23. MACHINE LEARNING 101 WORKING WITH TIME ▸ Dynamic Time Warping：帶有時間性的學習判定。 ▸ 如果要學習⼀一段旋律律或節奏的變化？ ▸ 如果要學習我控制器是順時針或逆時針旋轉？ ▸ Calculate feature vectors: Till long enough window in time. ▸ 在Machine learning中決定動作何時開始與結束，叫做segmentation。 ▸ 實務上，可將多個學習模型結合（chain multiple models together）， 例例如讓第⼀一個分類器來來判定動作發⽣生與否，然後⽤用分類輸出來來定義 segment的開始與結束時間。 ▸ Example: Play a game with voice
- 27. MACHINE LEARNING 101 SUPPORT VECTOR MACHINES ▸ SVMs是⼀一種監督式學習 (Supervised Learning)的⽅方法，主要⽤用在分類 (Classification)和回歸 (Regression)上。 ▸ 透過Kernel(內積)將data投射到⾼高維空間。Kernel代表兩兩者之間的相似性。 ▸ 線性Kernel（最不複雜） ▸ Polymonial（較⼤大的指數會使複雜度倍增） ▸ RBF（太複雜了了先不講） ▸ 不同的Kernel以及Kernel參參數將會產⽣生不同結果。
- 29. MACHINE LEARNING 101 ▸ RMS (Root Mean Square): 較⾼高的數值代表聽起來來聲⾳音比較⼤大聲 ▸ Peak FFT：透過快速傅立葉轉換，去偵測最顯著的fft ▸ Constant Q：跟傅立葉轉換類似，但是會變成對數分佈 ▸ Chromagram: 只要⾳音階中的相對⾳音⾼高，不要絕對⾳音⾼高。 ▸ MFCC(Mel-frequency cepstral coefficients): 會輸出許多不直觀訊息 ▸ Spectral Centroid: 適合⾳音⾊色變化很⼤大、多變化的演奏技法等。 FOR DETECTING AUDIO SIGNAL…
- 30. NOT ABOUT TO MAKE A MACHINE EVEN BETTER THAN THE HUMAN MACHINE LEARNING CAN… ▸ Support highly personalized experiences ▸ Empower more people to become creators. ▸ Enable creators to be people with bodies again. ▸ Support human design process. Faster creation means more prototyping, wide exploration, and better outcomes. ▸ Make happy surprises more likely.
- 31. QUOTED FROM DR.REBECCA FIEBRINK HOW IS MACHINE LEARNING IN THE ARTS DIFFERENT? In the arts Conventional ML You might create your own training data ‣ Use the training data to communicate the intensions. Change learning algorithm or features to get better results ‣ Does not make sense to change the data. Goal is to make a useful model Goal of learning is to better understand a given dataset or make predictions Training sets may be very small “Big data” is exciting! ‣ Enable better model, more accurate predictions YOU are the expert on model quality Use metrics like cross-validation to assess quality (i.e., generalization ability) Need real-time computation for interactive systems May not have tight real-time constraints.