What situation is it preferable to make predictions using machine learning rather than traditional statistical methods.
- Machine learning: The core to create an algorithm which a computer then uses to find a model that fits the data as best as possible. And makes very accurate predictions based on that.
- How is that different from conventional methods. We don’t give the machine instructions on how to find that model. We provided with algorithm which give the machine directions on how to learn on its own.
- Machine learning algorithm is like a trial and error process. Each consecutive trial is at least as good as the previous one. Technically speaking there are four ingredients, data, model, objective function and optimization algorithm. For example, an robot need to archery. The data is the arrow. The model is the usage of the bow. Objective function is calculate how far from the target. The optimization algorithm is mechanics that will improve the model’s performance.
- Training your model, no need to set of rules that is you won’t have programmed a set of instructions like place the arrow in the middle bow, pull the bow string and so on. Instead you will have given the machine a final goal. Place the arrow in the center of the target so you don’t care if it places the hero in the middle or in the bottom of the bow. So long as it hits the target.
- Training your model, model is trained. objective function is need to minimized. The optimization algorithm has done its job.
- The benefit of using machine learning is that the robot can learn to fire more effectively than a human. It might even discover that we’ve been holding bows in a wrong way for centuries. To conclude we must say that machine learning is not about robots. What people use it for is to improve complex computational models that can find infinite applications in our daily lives. Especially in the business world and that complex computational models we are talking about step on the fundaments of regressions and cluster analyses models.
- There are three major types of machine learning. The most common one appplied in the majority of cases we call supervised learning which the robot archery. It’t name dervies from the fact that training an algorithm resembles a teacher supervising her students to elaborate. Apart from the final goal you set to the robot. It is important to mention you have been dealing with label data. The data, in other words you can assess the accuracy of each shot. In fact there isn’t a single target different arrows have their own targets. Labelled data = associating, or “labelling”, a target to a type of arrow.
- Labelled data means we know the target prior to the shot and we can associate that shot with the target. This allows us to measure the inaccuracy of the shot through the objective function and improve the way the robot shoots through the optimization algorithm.
- In practive though it might happen that you won’t have the time or the resources to associate the arrows with targets before giving them to the robot.
- Unsupervised learning: you will just give your robot a bag of arrows with unknown physical properties unlabeled data. This means neither you nor the robot will have separated the arrows into groups. Then you would ask the machine to simple fire in a direction without providing it with targets. Therefore in this case you won’t be looking for a model that helps you shoot better rather you will be looking for one which divides the arrows in a certain way.
- Reinforcement learning plus a reward system. Instead of minimizing an error we are maximizing a reward or in other words maximizing the objective function. If you put yourself in the shoes of the machine you will be reasoning in the following way.
- SVMs = support vector machines.
- NNs = neural networks
- supervised learning: SVMs, NNs, deep learning, random forests, and bayesian networks.
- There are neural networkd that can be applied to an unsupervised type of machine learning but K-means is the most common unsupervised approach.
- Deep learning: New, revolutionary approach, is the state of the art ML. Describing it briefly we can say it is fundamentally different from the other approaches. However, it has a broad practical scope of application in all ML areas because of the extremely high accuracy of its models. The deep learning is still divided and supervised, unsupervised and reinforcement. So it solves the same problems but in a concequently different way.