Send the link below via email or IMCopy
Present to your audienceStart remote presentation
- Invited audience members will follow you as you navigate and present
- People invited to a presentation do not need a Prezi account
- This link expires 10 minutes after you close the presentation
- A maximum of 30 users can follow your presentation
- Learn more about this feature in our knowledge base article
Transcript of Machine Learning
Why talk about Machine Learning?
What is Machine Learning?
Enable a PC to make decisions using learning by example, NOT make decisions from hard coded rules
The concepts are not that complex, but it uses a lot of complex maths
Types of Machine Learning
Supervised Machine Learning
Unsupervised Machine Learning
Offline vs Online
Training is performed
Is where most of the complexity exists
Can be slow eg 1-2 days
The output is a model
Classification (or prediction) is performed
Use the model from training to classify new data
Typically fast (many times per second)
It is a core software component of most of our products:
Facial feature detection
Really important to train with data which represents all the classes you are interested in
Lots of variation = good!
We often have deficiencies with this
Time consuming to perform:
Collect the data
Label/annotated the data
We need lots of monkeys!
What features to represent ones data?
This is one of the most important design choices
Needs to be compact, yet descriptive
Example 1: Cool Car Detector
Example 2: Face Detector
How do you define a cool/uncool car?
Number of Wheels
How do you define an image of a face?
Pixel values have too much variation.
The classes of your data is already known
Includes methods such as:
Support Vector Machines (SVM)
This is the area we at Seeing Machines have good experience
Unknown classes in your data
The computer automatically learns clusters in your data
The classes may not be what you originally expected
We at Seeing Machines don't have any experience in this area
Example: Face Detection
Two classes of data: 1) face, 2) non-face
Approx 8000 face images, several 100,000 non-faces
Need something which captures the structure in a face.
Boosting is used to learn how to detect faces
Simple features are used to describe the face
Features are only just better than change at classifying faces vs non-faces
But very powerful when you combine many
of them together (Boosting)
We combine several 100 features
Ensuring the data we have contains enough variation
eg different lighting, faces with sunglasses
Currently getting good results with Boosted regression
Seb, Chanop, Farlin and myself currently using
The core of the new automatic initialisation
Light weight head tracking
Ability to classify/annotate new data quickly
Classification vs Regression