Neural networks used in machine learning (i)

This series of blogs is summarized according to Geoffrey Hinton course neural Network for machine learning. The course website is:

Https://www.coursera.org/course/neuralnets

1. Some examples The most applicable field example of the tasks best solved by learning machine learning

-Recognizing patterns: pattern recognition
–objects in real scenes object recognition
–facial identities or facial expressions face detection
–spoken Words language
Recognizing anomalies: identifying anomalies
–unusual sequences of credit card transactions unusual sequence of cards transactions
–unusual patterns of sensor readings in a nuclear power plant nuclear power plant sensors unusual readings
Prediction: Forecast
–future stock prices or currency exchange rates future stock price or currency exchange rate
–which movies Would a person like? What kind of movies does one like to watch?

2, use a machine learning standard example to explain many machine learning algorithms, many disciplines use this way
In the case of genetics, a lot of genetics was done on fruit flies ( fly Class).

–they is convenient because they breed fast.
–we already know a lot about them

The MNIST database of hand-written digits is the and the machine learning equivalent of fruit flies

–they is publicly available and we can get machine learning algorithm to learn what to recognize these handwritten digits , so it's easy to try lots of variations. them quite fast in a moderate-sized neural net.

–we know a huge amount about what well various machine learning methods does on MNIST. and particular, the different machine learning methods were implemented by people who believed in them, so we can rely on Those results.

So, we chose the Mnist database as our standard test task.

For example, some handwritten numbers in Mnist are as follows:

For example, the second row of 2, with one of the other to cover any other, it is difficult to match very well. So the template can't do the work. It is difficult to find the template suitable for the green box in these 2 and red boxes in the 2, so the handwritten numbers are more appropriate for machine learning.

Beyond mnist:imagenet Task

Mnist is now relatively simple for machine learning. We now have neural networks close to 1 million parameters that identify different object categories in the 1.3 million HD pixel training picture

Jitendra Malik (An eminent neural net sceptic) said the This competition was a good test of whether deep neural networks W Ork well for object recognition

–a Very deep neural net (Krizhevsky et al.) gets less, 40% error for it first choice and less than 20% for its Top 5 choices

Early versions of Neural networks

Neural networks used in machine learning (i)