Single Neuron. Hung-yi Lee

Similar documents
First Lecture of Machine Learning. Hung-yi Lee

What does the data look like? Logistic Regression. How can we apply linear model to categorical data like this? Linear Probability Model

Unit 6: Solving Exponential Equations and More

ECE602 Exam 1 April 5, You must show ALL of your work for full credit.

Chapter 13 GMM for Linear Factor Models in Discount Factor form. GMM on the pricing errors gives a crosssectional

Kernels. ffl A kernel K is a function of two objects, for example, two sentence/tree pairs (x1; y1) and (x2; y2)

3-2-1 ANN Architecture

N1.1 Homework Answers

A. Limits and Horizontal Asymptotes ( ) f x f x. f x. x "±# ( ).

Partial Derivatives: Suppose that z = f(x, y) is a function of two variables.

Status of LAr TPC R&D (2) 2014/Dec./23 Neutrino frontier workshop 2014 Ryosuke Sasaki (Iwate U.)

First derivative analysis

Derangements and Applications

Applied Statistics II - Categorical Data Analysis Data analysis using Genstat - Exercise 2 Logistic regression

High Energy Physics. Lecture 5 The Passage of Particles through Matter

Machine Learning Basics Lecture 7: Multiclass Classification. Princeton University COS 495 Instructor: Yingyu Liang

Exam 1. It is important that you clearly show your work and mark the final answer clearly, closed book, closed notes, no calculator.

Basic Polyhedral theory

Definition1: The ratio of the radiation intensity in a given direction from the antenna to the radiation intensity averaged over all directions.

EEO 401 Digital Signal Processing Prof. Mark Fowler

SER/BER in a Fading Channel

Radiation Physics Laboratory - Complementary Exercise Set MeBiom 2016/2017

Estimation of odds ratios in Logistic Regression models under different parameterizations and Design matrices

CS553 Lecture Register Allocation I 3

Lecture 37 (Schrödinger Equation) Physics Spring 2018 Douglas Fields

The Theory of Small Reflections

A Gentle Introduction to Neural Networks (with Python)

Learning Spherical Convolution for Fast Features from 360 Imagery

What are those βs anyway? Understanding Design Matrix & Odds ratios

Inheritance Gains in Notional Defined Contributions Accounts (NDCs)

Lecture #15. Bipolar Junction Transistors (BJTs)

SCHUR S THEOREM REU SUMMER 2005

EXST Regression Techniques Page 1

11: Echo formation and spatial encoding

Selective Mass Scaling (SMS)

Image Filtering: Noise Removal, Sharpening, Deblurring. Yao Wang Polytechnic University, Brooklyn, NY11201

y = 2xe x + x 2 e x at (0, 3). solution: Since y is implicitly related to x we have to use implicit differentiation: 3 6y = 0 y = 1 2 x ln(b) ln(b)

Logistic Regression. Sargur N. Srihari. University at Buffalo, State University of New York USA

Lagrangian Analysis of a Class of Quadratic Liénard-Type Oscillator Equations with Exponential-Type Restoring Force function

Solution of Assignment #2

Logistic Regression Review Fall 2012 Recitation. September 25, 2012 TA: Selen Uguroglu

Calculus II (MAC )

CSE 373: More on graphs; DFS and BFS. Michael Lee Wednesday, Feb 14, 2018

Hydrogen Atom and One Electron Ions

Addition of angular momentum


W considr a wight runing schm summarid as follows: th ntwork is traind to a minimum of th larning rror, th wight that would caus th smallst incras in

Data Assimilation 1. Alan O Neill National Centre for Earth Observation UK

Multilayer neural networks

Intro to Neural Networks and Deep Learning

The function y loge. Vertical Asymptote x 0.

Duffing Oscillator. Mike Brennan (UNESP) Bin Tang (Dalian University of Technology, China)

[ ] [ ] DFT: Discrete Fourier Transform ( ) ( ) ( ) ( ) Congruence (Integer modulo m) N-point signal

Ch. 24 Molecular Reaction Dynamics 1. Collision Theory

Principles of active remote sensing: Lidars. 1. Optical interactions of relevance to lasers. Lecture 22

CSC321 Lecture 4: Learning a Classifier

General Notes About 2007 AP Physics Scoring Guidelines

1997 AP Calculus AB: Section I, Part A

Why is a E&M nature of light not sufficient to explain experiments?

Difference -Analytical Method of The One-Dimensional Convection-Diffusion Equation

Chapter 2 Infinite Series Page 1 of 11. Chapter 2 : Infinite Series

Mock Exam 2 Section A

Final Exam Solutions

Gradebook & Midterm & Office Hours

Computing and Communications -- Network Coding

Chapter 1. Analysis of a M/G/1/K Queue without Vacations

Introduction to Condensed Matter Physics

6.1 Integration by Parts and Present Value. Copyright Cengage Learning. All rights reserved.

VII. Quantum Entanglement

PARTICLE MOTION IN UNIFORM GRAVITATIONAL and ELECTRIC FIELDS

Problem Set #2 Due: Friday April 20, 2018 at 5 PM.

PHA 5127 Answers Homework 2 Fall 2001

Addition of angular momentum

Fourier Transforms and the Wave Equation. Key Mathematics: More Fourier transform theory, especially as applied to solving the wave equation.

de/dx Effectively all charged particles except electrons

10. The Discrete-Time Fourier Transform (DTFT)

Network Congestion Games

Random Access Techniques: ALOHA (cont.)

nd the particular orthogonal trajectory from the family of orthogonal trajectories passing through point (0; 1).

Chapter 16. 1) is a particular point on the graph of the function. 1. y, where x y 1

CO-ORDINATION OF FAST NUMERICAL RELAYS AND CURRENT TRANSFORMERS OVERDIMENSIONING FACTORS AND INFLUENCING PARAMETERS

University of Illinois at Chicago Department of Physics. Thermodynamics & Statistical Mechanics Qualifying Examination

Large Systems (Section 2.4)

A Recommendation Model Based on Multi-Emotion Similarity in the Social Networks

MA 262, Spring 2018, Final exam Version 01 (Green)

The pn junction: 2 Current vs Voltage (IV) characteristics

Lecture 15: Logistic Regression

Appendix. Kalman Filter

Recursive Estimation of Dynamic Time-Varying Demand Models

ESE (Prelims) - Offline Test Series ELECTRICAL ENGINEERING SUBJECT: Power Electronics & Drives, Power Systems and Signals & Systems SOLUTIONS

Stochastic gradient descent; Classification

Propositional Logic. Combinatorial Problem Solving (CPS) Albert Oliveras Enric Rodríguez-Carbonell. May 17, 2018

1973 AP Calculus AB: Section I

Dealing with quantitative data and problem solving life is a story problem! Attacking Quantitative Problems

Quantum manipulation and qubits

Least Favorable Distributions to Facilitate the Design of Detection Systems with Sensors at Deterministic Locations

22/ Breakdown of the Born-Oppenheimer approximation. Selection rules for rotational-vibrational transitions. P, R branches.

VALUING SURRENDER OPTIONS IN KOREAN INTEREST INDEXED ANNUITIES

Extraction of Doping Density Distributions from C-V Curves

3. Stereoscopic vision 2 hours

Transcription:

Singl Nuron Hung-yi L

Singl Nuron 2 w w 2 Activation function a N w N ias

Larning to say ys/no Binary Classification

Larning to say ys/no Sam filtring Is an -mail sam or not? Rcommndation systms rcommnd th roduct to th customr or not? Malwar dtction Is th softwar malicious or not? Stock rdiction Will th futur valu of a stock incras or not with rsct to its currnt valu? Binary Classification

Eaml Alication: Sam filtring f : X Y { ys, no} E-mail Sam Not sam 2 f ys 2 (htt://sam-filtr-rviw.totnrviws.com/) f no

Eaml Alication: Sam filtring f : X Y { ys, no} What dos th function f look lik? y f ys no P P ys ys 0.5 0.5 How to stimat P(ys )?

Eaml Alication: Sam filtring To stimat P(ys ), collct amls first.. Earn fr fr Ys (Sam) Som words frquntly aar in th sam.g., fr 2 Win fr Ys (Sam) Us th frquncy of fr to dcid if an -mail is sam 3. Talk Mting. No (Not Sam) Estimat P(ys fr = k) fr is th numr of fr in -mail

Rgrssion In training data, thr is no - mail containing 3 fr. (ys fr ) (ys fr = ) = 0.4 (ys fr = 0 ) = 0. Frquncy of Fr ( fr ) in an -mail Prolm: What if on day you rciv an -mail with 3 fr.

Rgrssion f( fr ) = w fr + (f( fr ) is an stimat of (ys fr ) ) Stor w and (ys fr ) Rgrssion Frquncy of Fr ( fr ) in an -mail

Rgrssion f( fr ) = w fr + Th outut of f is not twn 0 and (ys fr ) Rgrssion Frquncy of Fr ( fr ) in an -mail Prolm: What if on day you rciv an -mail with 6 fr.

Logit ln fr vrtical lin: Proaility to sam (ys fr ) () is always twn 0 and vrtical lin: logit() logit ln

Logit f ( fr ) = w fr + (f ( fr ) is an stimat of logit() ) fr fr vrtical lin: Proaility to sam (ys fr ) () is always twn 0 and vrtical lin: logit() logit ln

Logit Stor w and fr f 3 fr w 3.5 logit ln.5 0.87 > 0.5, so ys f ( fr ) = w fr + (f ( fr ) is an stimat of logit() ) f fr w fr ln 0 0.5 ys 0 (rctron) vrtical lin: logit() logit fr ln

Multil Varials Considr two words fr and hllo comut (ys fr, hllo ) () logit ln hllo fr

Multil Varials Considr two words fr and hllo comut (ys fr, hllo ) () logit ln f fr, hllo Rgrssion w fr w2 hllo hllo fr

Multil Varials Of cours, w can considr all words {t, t 2, t N } in a dictionary w w w f N N t N t t t t t 2 2 2, w w N w w w 2 t N t t 2 is to aroimat logit() t N t t ys 2, P :

Logistic Rgrssion w aroimat logit ln : P ys If th roaility = or 0, ln(/-) = +infinity or infinity Can not do rgrssion t, t 2 t N Th roaility to sam is always or 0. t aars 3 tims t 2 aars 0 tim t N aars tim P ys t t 2 t N 3 0

Logistic Rgrssion w w w ln Sigmoid Function

Logistic Rgrssion Ys (Sam) 7 0 3 2 t N t t w clos to w No (not Sam) 2 2 w 2 clos to 0

Logistic Rgrssion Ys No t t2 t N fatur w w 2 w N w ias w 0

Logistic Rgrssion Ys No t t2 t N w w 2 w N w w 0

Rfrnc for Linar Classifir htt://grgor.chruala.m/ars/ml4nl/linarclassifirs.df

Mor than saying ys/no Multiclass Classification

Mor than saying ys/no Handwriting digit classification This is Multiclass Classification

Mor than saying ys/no Handwriting digit classification Simlify th qustion: whthr an imag is 2 or not Dscri th charactristics of inut ojct 2 Each il corrsonds to on dimnsion in N th fatur fatur of an imag

Mor than saying ys/no Handwriting digit classification Simlify th qustion: whthr an imag is 2 or not 2 2 or not N

Mor than saying ys/no Handwriting digit classification Binary classification of, 2, 3 If y 2 is th ma, thn th imag is 2. y or not 2 y 2 2 or not N y 3 3 or not

Limitation of singl nuron Motivation of cascading nurons

Limitation of Logistic Rgrssion w w 2 2 a ys no a a 0.5 0.5 ys no w w2 2 0 0 Inut Outut 2 0 0 No 0 Ys 0 Ys No Ys No

Limitation of Logistic Rgrssion w w 2 2 Inut Outut 2 0 0 No 0 Ys 0 Ys a 0 0 XNOR gat NOT 0 0 0 0 AND AND 0 0 OR No

Nural Ntwork NOT XNOR gat AND Nural Ntwork AND OR a 2 2 a 2 a Hiddn Nurons

a =0.73 a =0.27 a 2 a =0.27 a =0.05 2 a 2 =0.05 a 2 =0.27 2 a 2 2 a 2 =0.27 a 2 =0.73

2 a =0.73 a =0.27 a w a a =0.27 a =0.05 a 2 w 2 a 2 =0.05 a 2 =0.27 (0.73, 0.05) 2 a 2 a 2 =0.27 a 2 =0.73 (0.27, 0.27) (0.05,0.73) a

Thank you for your listning!

Andi

Mor rfrnc htt://www.ccs.nu.du/hom/vi/tach/mlcours /2_GD_REG_ton_NN/lctur_nots/logistic_rgr ssion_loss_function/logistic_rgrssion_loss.df htt://mathgotchas.logsot.tw/20/0/why-isrror-function-minimid-in.html htts://cs.nyu.du/~yann/talks/lcun-2007207- nonconv.df htt://www.cs.columia.du/~li/fogm/lcturs/g lms.df

Logistic Rgrssion t N t t 2? P ys w ln w