Advanced Topics Work Force Skills Training: Fill & Download for Free

I would say that first of all you have to decide on which skill you really want to get remarkable skills. You cannot excell in all major development areas so you need to understand what you really like and you think as a future as a web technology for the future years.Then use one of the sevral online training resources out there such as Coursera, tutplus, codecademy, egghead etc. and go from the basics to the more advanced topics.Try to build stuff while studying and learning otherwise you will retain very little of what you read or listen to. Create rapid prototypes of actually working applications to be confronted with real life problems that will force you to understand how things work.Finally, land on some of the online marketplace for frelancers such as Upwork, the world's largest online workplace, Hire Freelancers & Find Freelance Jobs Online, Rocklance.com - freelancers who rock! and create a very detailed profile being first of all VERY HONEST on your actual skills.The first profiles I filter our when searching for freelancers are those that list any kind of technology on earth and say that they are in the top 1% worldwide. I really trust only those profiles that list a reasonable limited amount of skills.

I largely agree with Sajib Hasan on this topic. I would like to elaborate on a couple of things that I consider the main problems.Backdated:The education system is not up to date with the twenty first century. There are completely useless lessons like 'islam shikkha' until SSC while very important information are completely left out.Not diverse:Students are trained to for one and only one kind of job - the so called white collar ones. Not enough care is taken for vocation training. Imagine, if there was a solid vocational education system, there could've been thousands of young men and women graduating SSC and HSC with real life skills like auto work, plumbing, construction, driving, etc. The govt. wouldn't have to go beg for the Saudis to hire a few thousand house maids everyday. Instead the work force could've made it into Europe and other more advanced labor markets.

This is one of the classic challenges in machine learning. One of the greatpapers in this area is: An introduction to variable and feature selectionIt provides a great background on feature selection and some of the classic algorithms - but it's a little dated, and doesn't give quite the attention to the more modern technique of "regularization" that is covered in recent statistical learning books such as: data mining, inference, and prediction. 2nd Edition.One of the best ways to discover 'useful' features is to use "regularization" in the learning algorithm. The "Lasso" method Boris B. mentions is one example of a "regularized" learning algorithm.Intuitively, regularization is a constraint that forces the learning algorithm to limit the contribution that each feature makes to the final prediction. When the learning algorithm is run, a regularization parameter (often called a tuning parameter) controls the amount of constraint. With a tight constraint, the algorithm is forced to chose a small number of features in it's prediction. After the learning step has occurred, the weights can be viewed. Highly weighted features are those that are most important in the prediction.Learning how to best use regularization constraints is a critical skill in machine learning, and can make the difference in creating an algorithm that works well, from one that totally bombs - even on the same training data.The Lasso, mentioned in other posts is a regulation algorithm that uses a L1 norm constraint.Another type of algorithm is the L2 constraint - sometimes called ridge regression.The L2 constraint is different from the L1 Lasso constraint in that that algorithm will typically assign non-zero weight to every feature. You won't get the clean zero-weighted features that the Lasso will give. But often it doesn't matter - you just look at the highly weighted features anyway.There are datasets that will do better with L2, and others that will do better with L1 constraints. Often it depends on how independent the features are (e.g. how much correlation there is between features).But in my experience, L2 constraints often do just as well at prediction but are significantly easier to train - and definitely the thing to try first. Since L2 constraints are easily differentiable, the training algorithms (like ridge regression) can yield exact solutions without using gradient/coordinate descent methods. This all means the training step can be significantly faster.There are other great computational benefits of L2 models:1) you can do some matrix tricks so that you only need to do one expensive matrix inversion even when you test multiple regularization "tuning" parameters2) you can quickly figure out the leave-one-out test score without actually doing it (another matrix trick)3) If features significantly outweigh training examples, you can kernalize the algorithm for computation benefitThese are more advanced topics covered in the Hastie Tibshirani book and also the papers of Michael Jordan (at Berkeley).But simply, I'd start like this:1) Learn how to train and use models like ridge regression and L2 regularized logistic regression2) Experiment with training using different tuning parameters to see how it impacts the weights learned.3) Don't forget the importance of normalizing the training feature vectors when you combine multiple features into a single vector. If the relative scales are different, this can totally screw up the learning process if you don't normalize.When you are comfortable with these things, then I would experiment with more advanced techniques like the Lasso (L1) and see how it works on your data.You'll find that L1 techniques are very popular in the literature, often because it has nice mathematical properties that let researchers prove guarantees about the performance of the algorithm given many assumptions.In reality, the assumptions rarely hold with real data, and the normalization step and choosing the right tuning parameter will have much greater effect than the particular technique you choose (L1 vs L2).Hope this helps.