introduction to artificial intelligence neural networks ... · deep learning i most current machine...

21
Introduction to Artificial Intelligence Neural Networks - Deep Learning for NLP Janyl Jumadinova November 21, 2016

Upload: others

Post on 17-Jun-2020

11 views

Category:

Documents


0 download

TRANSCRIPT

Page 1: Introduction to Artificial Intelligence Neural Networks ... · Deep Learning I Most current machine learning works well because of human-designed representations and input features

Introduction to Artificial IntelligenceNeural Networks - Deep Learning for

NLP

Janyl JumadinovaNovember 21, 2016

Page 2: Introduction to Artificial Intelligence Neural Networks ... · Deep Learning I Most current machine learning works well because of human-designed representations and input features

Neural Networks

2/20

Page 3: Introduction to Artificial Intelligence Neural Networks ... · Deep Learning I Most current machine learning works well because of human-designed representations and input features

Neural Networks

3/20

Page 4: Introduction to Artificial Intelligence Neural Networks ... · Deep Learning I Most current machine learning works well because of human-designed representations and input features

Neural NetworksNeural computing requires a number of neurons, to be connectedtogether into a neural network.

Neurons are arranged in layers.

4/20

Page 5: Introduction to Artificial Intelligence Neural Networks ... · Deep Learning I Most current machine learning works well because of human-designed representations and input features

Activation Functions

I The activation function is generally non-linear.

I Linear functions are limited because the output is simplyproportional to the input.

5/20

Page 6: Introduction to Artificial Intelligence Neural Networks ... · Deep Learning I Most current machine learning works well because of human-designed representations and input features

Activation Functions

6/20

Page 7: Introduction to Artificial Intelligence Neural Networks ... · Deep Learning I Most current machine learning works well because of human-designed representations and input features

Network structures

Feed-forward networks:

I Single-layer perceptrons

I Multi-layer perceptrons

7/20

Page 8: Introduction to Artificial Intelligence Neural Networks ... · Deep Learning I Most current machine learning works well because of human-designed representations and input features

Feed-forward example

8/20

Page 9: Introduction to Artificial Intelligence Neural Networks ... · Deep Learning I Most current machine learning works well because of human-designed representations and input features

Single-layer Perceptrons

Output units all operate separately – no shared weights.

Adjusting weights moves the location, orientation, and steepness ofcliff. 9/20

Page 10: Introduction to Artificial Intelligence Neural Networks ... · Deep Learning I Most current machine learning works well because of human-designed representations and input features

Multi-layer Perceptrons

I Layers are usually fully connected.I Numbers of hidden units typically chosen by hand.

10/20

Page 11: Introduction to Artificial Intelligence Neural Networks ... · Deep Learning I Most current machine learning works well because of human-designed representations and input features

A neural network for learning word vector

I Idea: A word and its context is a posiGve training sample

I A random word in that same context gives a negative trainingsample:

11/20

Page 12: Introduction to Artificial Intelligence Neural Networks ... · Deep Learning I Most current machine learning works well because of human-designed representations and input features

A neural network for learning word vector

12/20

Page 13: Introduction to Artificial Intelligence Neural Networks ... · Deep Learning I Most current machine learning works well because of human-designed representations and input features

A neural network for learning word vector

These are the word features we want to learn .13/20

Page 14: Introduction to Artificial Intelligence Neural Networks ... · Deep Learning I Most current machine learning works well because of human-designed representations and input features

A neural network for learning word vector

14/20

Page 15: Introduction to Artificial Intelligence Neural Networks ... · Deep Learning I Most current machine learning works well because of human-designed representations and input features

Deep Learning

I Most current machine learning works well because ofhuman-designed representations and input features .

I Machine learning becomes just optimizing weights to best makea final prediction.

I Deep learning algorithms attempt to learn multiple levels ofrepresentation of increasing complexity/abstraction.

15/20

Page 16: Introduction to Artificial Intelligence Neural Networks ... · Deep Learning I Most current machine learning works well because of human-designed representations and input features

Deep Learning

I Most current machine learning works well because ofhuman-designed representations and input features .

I Machine learning becomes just optimizing weights to best makea final prediction.

I Deep learning algorithms attempt to learn multiple levels ofrepresentation of increasing complexity/abstraction.

15/20

Page 17: Introduction to Artificial Intelligence Neural Networks ... · Deep Learning I Most current machine learning works well because of human-designed representations and input features

A Deep Architecture

16/20

Page 18: Introduction to Artificial Intelligence Neural Networks ... · Deep Learning I Most current machine learning works well because of human-designed representations and input features

The Need for Distributed RepresentationsCurrent NLP systems are incredibly fragile because of their atomicsymbol representations

17/20

Page 19: Introduction to Artificial Intelligence Neural Networks ... · Deep Learning I Most current machine learning works well because of human-designed representations and input features

Handling the recursivity of human language

18/20

Page 20: Introduction to Artificial Intelligence Neural Networks ... · Deep Learning I Most current machine learning works well because of human-designed representations and input features

Recursive Deep Learning: Building on Word

Vector Space Models

19/20

Page 21: Introduction to Artificial Intelligence Neural Networks ... · Deep Learning I Most current machine learning works well because of human-designed representations and input features

How should we map phrases into a vector

space?

20/20