Week One - Video One - Why ML STrategy
Why we should learn care about ML Strategy
Here when we try to improve the performance of the system we should consider about a lot of things . They are:
-Amount of data
- Amount of diverse data
- Train algorithm longer with gradient descent
-use another optimization algorithm like Adam
- use bigger network or smaller network depending out requirement
- use drop out
- add l2 regularization
- network architecture parameters like number of hidden units, Activation function etc.
Second Video - Orthogonalization
Orthogonalization means in a deep learning network we can change/tune so many things for eg. hyper parameters to get a more performance in the network . So most effective people know what to tune in order to achieve a particular effect.
For every set of problem there is a separate solution. Don't mix up the problems and solutions. For that, first we should find out where is the problem , whether it is with training set/ dev set/ test set/ real world data.
For each problem, the solution is there only. Separate solutions for separate problem. Like each knob for setting height, width or triangle mode in a television. So the difficulty is finding that problem. Expert people can find what is the problem.
Four thing we can do properly and it solutions if some problem is found are given below.
Fit training set well on cost function - Eg. Bigger network, Adam functions
Fit dev set well on cost function - Regularization , Bigger training set
Fit test set well on cost function - Bigger dev set
Performs well in real world problems - Change dev set or cost function
Second Week - Video One - Carrying out error analysis
Error Analysis
If the performance of the system is less and there is an error rate of 10 percent , we can check the development set that on how many of them are getting wrong labels. So we should take a stats on mislabeled images. We can count on the mislabeled images manually and find what made these mislabeling. We can note this in a CSV file format.
The image id , And the reasons..keep on adding reasons as you find
Eg.
Img Name, Like dog, Like lion, Blurry images , Instagram images, Comments etc columns are given to the csv.
We count these images and mark these image counts.
Hence we get a percentage of error reasons ,so that we can work on which error contributed most to the mislabeled images.
Why we should learn care about ML Strategy
Here when we try to improve the performance of the system we should consider about a lot of things . They are:
-Amount of data
- Amount of diverse data
- Train algorithm longer with gradient descent
-use another optimization algorithm like Adam
- use bigger network or smaller network depending out requirement
- use drop out
- add l2 regularization
- network architecture parameters like number of hidden units, Activation function etc.
Second Video - Orthogonalization
Orthogonalization means in a deep learning network we can change/tune so many things for eg. hyper parameters to get a more performance in the network . So most effective people know what to tune in order to achieve a particular effect.
For every set of problem there is a separate solution. Don't mix up the problems and solutions. For that, first we should find out where is the problem , whether it is with training set/ dev set/ test set/ real world data.
For each problem, the solution is there only. Separate solutions for separate problem. Like each knob for setting height, width or triangle mode in a television. So the difficulty is finding that problem. Expert people can find what is the problem.
Four thing we can do properly and it solutions if some problem is found are given below.
Fit training set well on cost function - Eg. Bigger network, Adam functions
Fit dev set well on cost function - Regularization , Bigger training set
Fit test set well on cost function - Bigger dev set
Performs well in real world problems - Change dev set or cost function
Second Week - Video One - Carrying out error analysis
Error Analysis
If the performance of the system is less and there is an error rate of 10 percent , we can check the development set that on how many of them are getting wrong labels. So we should take a stats on mislabeled images. We can count on the mislabeled images manually and find what made these mislabeling. We can note this in a CSV file format.
The image id , And the reasons..keep on adding reasons as you find
Eg.
Img Name, Like dog, Like lion, Blurry images , Instagram images, Comments etc columns are given to the csv.
We count these images and mark these image counts.
Hence we get a percentage of error reasons ,so that we can work on which error contributed most to the mislabeled images.
Comments
Post a Comment