I was simply doing some late night perusing and ran over this article:
http://gigaom.com/information/why-turning into an information researcher may be-simpler than-you-might suspect/
TL;DR - You can take the ML seminar on Coursera and you’re mystically an information researcher, since three extremely wise individuals did it. I oppose this idea.
I’m not guaranteeing the individuals referenced in this article are not information researchers who score high in Kaggle rivalries. They’re most likely extremely clever individuals who got another aptitude and exceeded expectations at it (albeit one was at that point a statistician, so he is fundamentally doing AI in some structure as of now).
Click here
Here is my concern with it - being an information researcher for the most part requires an a lot bigger range of abilities than an essential comprehension of a couple of learning calculations. I’m taking the Coursera ML course at this moment, and I think it is extraordinary! Here is the thing that I didnt adapt however: Programming Languages and Other Technologies: Most information researchers and the organizations that utilize them are not utilizing Matlab/Octave. They have backend web administrations written in Java, Python, Scala, or Ruby. These dialects are not secured. Python has libraries like Scipy, Numpy, and Scikit-discover that are extraordinary for taking care of numerical issues. Java has a lot of libraries too like the Mahout math library [2]. R is utilized by most analysts (again not canvassed in the course). At the point when your chief (or a client) comes to you and says you have to incorporate a calculation into a previous web administration ( model - they need a proposal motor), and you state “I just know Matlab” that is going be an immense issue. You don’t simply get Java/Python/C++/Scala/whatever in a couple of days at work. You must be to some degree acquainted with these dialects to see enormous, prior code bases. It wouldn’t damage to have a not too bad comprehension of existing advancements like Django, ROR, Groovy, Lift, and so on in light of the fact that you will need to coordinate your astonishing calculation into one of them. On the off chance that you just know Python yet the remainder of the organization is utilizing Java, you better think about Thrift, Avro, Google ProtoBufs or something smiliar. I deviate … Enormous Data Software: Most information researchers are dealing with issues that can’t be kept running on single 512MB RAM machine (the informational collections on Coursera are minor). They have enormous informational collections that require conveyed preparing. To do this, you have to comprehend map-diminish, appropriated records frameworks and have the option to use Hadoop. Regardless of whether you don’t know Java, despite everything you have to realize that Hadoop gushing exists, how use it, and know a scripting language (Hadoop spilling doesn’t as of now support Matlab or Octave). Again - not something you simply get in a couple of days. In case you will do appropriated AI, you will presumably need to utilize Mahout. A portion of the calculations shrouded in the Coursera course have appropriated forms executed into Mahout (Clustering, PCA, Regression), yet most don’t exist yet (SVM, Distributed-User-Recommendations, and cross-approval/execution measurements for circulated renditions of the calculations). You will need to know Java (and the Hadoop/Mahout APIs) to execute them without any preparation, or utilize an alternate calculation that you could possibly be comfortable with. Regardless of whether you don’t have to utilize a disseminated calculation, it would presumably be great to realize how to turn up a 64GB occurrence on ec2, login, introduce some product, and run your calculation on the cloud. Other helpful learning calculations Coursera skirted Bayesian learning [1]. A ton of frameworks utilize this (or some type of this) underway, yet you would think nothing about it (I’m not saying you couldn’t learn it, yet I am clearly contradicting this article). Highlight Extraction: You can utilize each calculation from the ML course and construct many various classifiers to take care of a true issue (you could even join them!), yet on the off chance that your highlights suck, the presentation of your classifier is going to suck moreover. Separating great highlights typically requires a profound comprehension of the issue, the hidden appropriations of the information, and additionally a commonality of how the information is being created. It may help to likewise think about Convolution, Wavelets, Time Series Analysis, Digital Signal Processing, Fourier Transforms, and so on. Highlight extraction could be an entire Coursera course independent from anyone else. Information cleaning: Information preprocessing - Coursera sets up every one of the informational indexes for you. They even compose the contents to stack the information. (see week 6 SVM email arrangement, they composed the majority of the regex articulations to clean the messages for you) That doesn’t work in reality. True information is terrible, and unstructured. You have to know ordinary articulations and UNIX directions like sed, grep, tr, cut, sort, awk, and map/diminish to tidy these informational collections up and put them into “Coursera” position. Notice I said UNIX directions, which infers you should be to some degree agreeable on UNIX/LINUX, which might be a precarious expectation to learn and adapt in case you’re at present utilizing Windows. Likelihood and Statistics The ML course addresses a portion of these subjects, yet genuine issues more often than not require an a lot further understanding. Models: Are your highlights needy or autonomous (Chi Squared test). How would you translate p-values? How would you set up certainty interims? What is the F-test? What is standard mistake? Would it be advisable for me to utilize AROCs to test the exhibition of this calculation? What is a ROC bend? What is Hypothesis trying and when can you/do you use it. I could go on - however you get the point . It’s essential to know this stuff and skill and when to apply it. Databases: Where is this information being put away? It’s fine in level documents for the reasons for the ML course, however when you appear at your first day at the particular employment, it will be put away in MySQL, Postgres, MongoDB, Casandra, CouchDB, or potentially on the HDFS. Your must extemporize. Perception: On the off chance that you think Matlab 2D plots are amazing, look at D3.js. Gracious, I neglected to make reference to, you have to know javascript, utilitarian programming, and the general expectation to absorb information for the D3.js APIs. Troubleshooting: Troubleshooting is quickly canvassed in the ML course (Neural Networks - Gradient Checking, which was a discretionary programming excerise), yet when a calculation isn’t working effectively, you will need to stop treating it like a black-box and make a plunge. That is when poop gets genuine - you really need to think about Conjungate Gradients, Partial Differential Equations, Numerical Analysis, Lagrange Multipliers, Numerical Linear Alegbra, Convex Optimization, Vector Calculus, Stochastic Processes, and possibly a great deal of different subjects. The ML course is just two months, and they can’t cover these subjects, since they require long periods of experience, and presumably some type of specialized preparing past a BS in whatever. You could simply ask the person with a couple of years more experience that is going to take your activity however
Click here
Here is my concern with it - being an information researcher for the most part requires an a lot bigger range of abilities than an essential comprehension of a couple of learning calculations. I’m taking the Coursera ML course at this moment, and I think it is extraordinary! Here is the thing that I didnt adapt however: Programming Languages and Other Technologies: Most information researchers and the organizations that utilize them are not utilizing Matlab/Octave. They have backend web administrations written in Java, Python, Scala, or Ruby. These dialects are not secured. Python has libraries like Scipy, Numpy, and Scikit-discover that are extraordinary for taking care of numerical issues. Java has a lot of libraries too like the Mahout math library [2]. R is utilized by most analysts (again not canvassed in the course). At the point when your chief (or a client) comes to you and says you have to incorporate a calculation into a previous web administration ( model - they need a proposal motor), and you state “I just know Matlab” that is going be an immense issue. You don’t simply get Java/Python/C++/Scala/whatever in a couple of days at work. You must be to some degree acquainted with these dialects to see enormous, prior code bases. It wouldn’t damage to have a not too bad comprehension of existing advancements like Django, ROR, Groovy, Lift, and so on in light of the fact that you will need to coordinate your astonishing calculation into one of them. On the off chance that you just know Python yet the remainder of the organization is utilizing Java, you better think about Thrift, Avro, Google ProtoBufs or something smiliar. I deviate … Enormous Data Software: Most information researchers are dealing with issues that can’t be kept running on single 512MB RAM machine (the informational collections on Coursera are minor). They have enormous informational collections that require conveyed preparing. To do this, you have to comprehend map-diminish, appropriated records frameworks and have the option to use Hadoop. Regardless of whether you don’t know Java, despite everything you have to realize that Hadoop gushing exists, how use it, and know a scripting language (Hadoop spilling doesn’t as of now support Matlab or Octave). Again - not something you simply get in a couple of days. In case you will do appropriated AI, you will presumably need to utilize Mahout. A portion of the calculations shrouded in the Coursera course have appropriated forms executed into Mahout (Clustering, PCA, Regression), yet most don’t exist yet (SVM, Distributed-User-Recommendations, and cross-approval/execution measurements for circulated renditions of the calculations). You will need to know Java (and the Hadoop/Mahout APIs) to execute them without any preparation, or utilize an alternate calculation that you could possibly be comfortable with. Regardless of whether you don’t have to utilize a disseminated calculation, it would presumably be great to realize how to turn up a 64GB occurrence on ec2, login, introduce some product, and run your calculation on the cloud. Other helpful learning calculations Coursera skirted Bayesian learning [1]. A ton of frameworks utilize this (or some type of this) underway, yet you would think nothing about it (I’m not saying you couldn’t learn it, yet I am clearly contradicting this article). Highlight Extraction: You can utilize each calculation from the ML course and construct many various classifiers to take care of a true issue (you could even join them!), yet on the off chance that your highlights suck, the presentation of your classifier is going to suck moreover. Separating great highlights typically requires a profound comprehension of the issue, the hidden appropriations of the information, and additionally a commonality of how the information is being created. It may help to likewise think about Convolution, Wavelets, Time Series Analysis, Digital Signal Processing, Fourier Transforms, and so on. Highlight extraction could be an entire Coursera course independent from anyone else. Information cleaning: Information preprocessing - Coursera sets up every one of the informational indexes for you. They even compose the contents to stack the information. (see week 6 SVM email arrangement, they composed the majority of the regex articulations to clean the messages for you) That doesn’t work in reality. True information is terrible, and unstructured. You have to know ordinary articulations and UNIX directions like sed, grep, tr, cut, sort, awk, and map/diminish to tidy these informational collections up and put them into “Coursera” position. Notice I said UNIX directions, which infers you should be to some degree agreeable on UNIX/LINUX, which might be a precarious expectation to learn and adapt in case you’re at present utilizing Windows. Likelihood and Statistics The ML course addresses a portion of these subjects, yet genuine issues more often than not require an a lot further understanding. Models: Are your highlights needy or autonomous (Chi Squared test). How would you translate p-values? How would you set up certainty interims? What is the F-test? What is standard mistake? Would it be advisable for me to utilize AROCs to test the exhibition of this calculation? What is a ROC bend? What is Hypothesis trying and when can you/do you use it. I could go on - however you get the point . It’s essential to know this stuff and skill and when to apply it. Databases: Where is this information being put away? It’s fine in level documents for the reasons for the ML course, however when you appear at your first day at the particular employment, it will be put away in MySQL, Postgres, MongoDB, Casandra, CouchDB, or potentially on the HDFS. Your must extemporize. Perception: On the off chance that you think Matlab 2D plots are amazing, look at D3.js. Gracious, I neglected to make reference to, you have to know javascript, utilitarian programming, and the general expectation to absorb information for the D3.js APIs. Troubleshooting: Troubleshooting is quickly canvassed in the ML course (Neural Networks - Gradient Checking, which was a discretionary programming excerise), yet when a calculation isn’t working effectively, you will need to stop treating it like a black-box and make a plunge. That is when poop gets genuine - you really need to think about Conjungate Gradients, Partial Differential Equations, Numerical Analysis, Lagrange Multipliers, Numerical Linear Alegbra, Convex Optimization, Vector Calculus, Stochastic Processes, and possibly a great deal of different subjects. The ML course is just two months, and they can’t cover these subjects, since they require long periods of experience, and presumably some type of specialized preparing past a BS in whatever. You could simply ask the person with a couple of years more experience that is going to take your activity however