tag:blogger.com,1999:blog-7994087232040033267.post4954116798443987082..comments2019-09-20T21:53:44.924-07:00Comments on Pragmatic Programming Techniques: Characteristics of Machine Learning ModelRicky Hohttp://www.blogger.com/profile/03793674536997651667noreply@blogger.comBlogger5125tag:blogger.com,1999:blog-7994087232040033267.post-70422341229700091922012-04-26T07:58:16.641-07:002012-04-26T07:58:16.641-07:00Machine learning? In the end, it all comes with a ...Machine learning? In the end, it all comes with a concept of <a href="http://www.360training.com/" rel="nofollow">online training solutions</a>. The approaches are different but the end result are almost the same.Leonard Vicehttps://www.blogger.com/profile/02718719192132004268noreply@blogger.comtag:blogger.com,1999:blog-7994087232040033267.post-10091815360657121012012-03-02T19:17:32.421-08:002012-03-02T19:17:32.421-08:00Hi Ricky,
Few comments:-
Linear Regressio...Hi Ricky, <br /> Few comments:-<br />Linear Regression:-The objective as you mentioned is to find the best weights by minimizing the loss function ||Xw-Y||^2. The best way to solve this is by minimizing this loss function by taking the derivative w.r.t w which is X^{T}Xw-X^{T}y (Putting in Latex Symbols). Regularization helps us in getting sparse values i.e by selecting few columns which correspondingly also helps in storing data efficiently.L1 is quite efficient in making this sparse unlike L2. But the optimization problem of L1 is quite complicated (not differentiable at the origin) so you need to frame the optimization problem efficiently.<br />I am not sure if i understand what you mean by linear assumption of features ? <br />Yes if the data is highly non linear then SVM with RBF kernel should be good (based on experiments but no thereotical guarentees).<br />Neural Networks:- You also have single perceptron but yes they do not solve non linear problems. The good thing about Neural Networks are that they solve this non linear issues problem is that they are very non convex. You can never have a global solution and need to work on heurisitcs to make sure that the solution you have is good for your problem. Again there are algorithms like Gradient Descent, Stochastic Gradient Descent, etc. Also you can make sure that Neural Networks have real values i.e non binary values. <br />Support Vector Machines:-I am not sure what do you say binary values you could change your loss function to get non binary values for say a task like multi classification problem where the taks for SVM is to give a value of k={1...K} classes. <br /><br />Overall nice article..<br /><br />Cheers<br />TanmoyTanmoyhttps://www.blogger.com/profile/08373959294561464615noreply@blogger.comtag:blogger.com,1999:blog-7994087232040033267.post-78481315676197728652012-02-21T02:21:16.272-08:002012-02-21T02:21:16.272-08:00Thank you for this summary. What about Markov Logi...Thank you for this summary. What about Markov Logic Network and Conditional Random Fields?<br /><br />Thank you again. Following your blog is always a pleasure.<br /><br />michele.Michele Filanninohttps://www.blogger.com/profile/16791100144449477090noreply@blogger.comtag:blogger.com,1999:blog-7994087232040033267.post-11416147354689916572012-02-20T17:59:20.684-08:002012-02-20T17:59:20.684-08:00Excellent summary of the different classification ...Excellent summary of the different classification techniques in machine learning.Pranabhttps://www.blogger.com/profile/13640683687903025427noreply@blogger.comtag:blogger.com,1999:blog-7994087232040033267.post-83424068655475295352012-02-20T06:36:00.429-08:002012-02-20T06:36:00.429-08:00Hi Ricky,
This is a nice and useful summary of di...Hi Ricky,<br /><br />This is a nice and useful summary of different ML algorithms, thanks for posting it!<br /><br />However, it was pointed out in the discussion on the LinkedIn group that logistic regression is a very popular model. It would round out this post nicely if you put a couple of paragraphs about that, as well.<br /><br />Thanks!<br />Fred.Fredhttps://www.blogger.com/profile/15150639324968431297noreply@blogger.com