{"id":975,"date":"2018-11-12T02:54:23","date_gmt":"2018-11-11T23:54:23","guid":{"rendered":"http:\/\/kusuaks7\/?p=580"},"modified":"2021-12-15T05:21:25","modified_gmt":"2021-12-15T05:21:25","slug":"ways-in-which-machines-learn","status":"publish","type":"post","link":"https:\/\/www.experfy.com\/blog\/ai-ml\/ways-in-which-machines-learn\/","title":{"rendered":"Ways in Which Machines Learn"},"content":{"rendered":"

Ready to learn Machine Learning? Browse<\/em><\/strong> Machine Learning Training and Certification courses<\/a> developed by industry thought leaders and Experfy in Harvard Innovation Lab.<\/em><\/strong><\/p>\n

There are four major ways to train deep learning networks:\u00a0supervised<\/em>,unsupervised<\/em>,\u00a0semi-supervised<\/em>, and\u00a0reinforcement learning<\/em>.\u00a0We\u2019ll explain the intuitions behind each of the these methods. Along the way, we\u2019ll share terms you\u2019ll read in the literature in parentheses and point to more resources for the mathematically inclined. By the way, these categories span both traditional machine learning algorithms and the newer, fancier deep learning algorithms.<\/p>\n

For the math-inclined, see\u00a0this Stanford tutorial which covers supervised and unsupervised learning<\/a>\u00a0and includes code samples.<\/p>\n

Supervised Learning<\/h3>\n

Supervised Learning<\/strong>\u00a0trains networks using examples where we already know the correct answer. Imagine we are interested in training a network to recognize pictures from your photo library that have your parents in them. Here\u2019s the steps we\u2019d take in that hypothetical scenario.<\/p>\n

Step 1: Data Set Creation and Categorization<\/h3>\n

We would start the process by going through your photos (the data set) and identifying all the pictures that have your parents in them, labeling them. We would then take the whole stack of photos and split them into two piles. We would use the first pile to train the network (training data) and the second pile to see how accurate the model is at picking out photos with our parents (validation data).<\/p>\n

Once the data sets are ready, we\u2019d feed the photos to the model. Mathematically, our goal is for the deep network to find a function whose input is a photo and whose output is a 0 when your parents are not in the photo or 1 when they are.<\/p>\n

This step is usually called the\u00a0categorization task<\/em>. In this case we\u2019re training for results that are yes-no, but supervised learning can also be used to output a set of values, rather than just a 0 or 1. For example, we might train a network to output the probability that someone will repay a credit card loan, in which case the output is anywhere between 0 and 100. These tasks are called regressions.<\/p>\n

Step 2:\u00a0Training<\/h3>\n

To continue the process, the model makes a prediction for each photo by following rules (activation function) to decide whether to light up a particular node in the work. The model works from left to right one layer a time \u2014 we will ignore more complicated networks for the moment. After the network calculates this for every node in the network, we\u2019ll get to the rightmost node (output node) which lights up or not.<\/p>\n

Since we already know which pictures have your parents in them, we would be able to tell the model whether its prediction is right or wrong. We would then\u00a0feed back<\/em>\u00a0this information to the network.<\/p>\n

The algorithm uses this feedback, which is the result of a function that quantifies \u201chow far off from the real answer is from the model\u2019s prediction\u201d. This is called a\u00a0cost function<\/em>, also known as\u00a0objective function<\/em>,\u00a0utility function<\/em>or\u00a0fitness function<\/em>. The result of the function is then used to modify the strength of connections and biases between nodes in a process called\u00a0backpropagation<\/em>\u00a0since the information travels \u201cbackwards\u201d from the result nodes.<\/p>\n

We\u2019d repeat this for each of the pictures, and in each case the algorithms try to minimize the cost function.<\/p>\n

There are a variety of mathematical techniques to use this knowledge of whether the model was right or wrong back into the model, but a very common method is gradient descent.\u00a0Algobeans<\/a>\u00a0has a good layman\u2019s explanation of how this works. Michael Nielsen\u00a0adds the math<\/a>\u00a0which involves calculus and linear algebra (and a friendly demon!).<\/p>\n

Step 3:\u00a0Verify<\/h3>\n

Once we\u2019ve processed all the photos from our first stack we will be ready to test the model. We would grab the second stack of photos and use them to see how accurately the trained model can pick up photos of your parents.<\/p>\n

Steps 2 and 3 would typically be repeated by tweaking various things about the model (hyperparameters), such as how many nodes there are, how many layers there are, which mathematical function to use to decide whether a node lights up, how aggressively to train the weights during the backpropagation phase, and so on. This\u00a0Quora answer<\/a>\u00a0has a good explanation of the knobs you can turn.<\/p>\n

Step 4:\u00a0Use<\/h3>\n

Finally, once you have an accurate model, you deploy that model into your application. Your expose the model as an API call, such as\u00a0ParentsInPicture(photo)<\/code>, and you can call that method from your software, causing the model to make an inference and giving you the result.<\/p>\n

We\u2019ll go through this exact process later to write an iPhone application that recognizes business cards.<\/p>\n

It can be hard (that is, expensive) to get a labeled data set, so you need to make sure the value of the prediction justifies the cost of getting the labeled data and training the model in the first place. For example, getting labeled X-rays of people who might have cancer is expensive, but the value of an accurate model that generates few false positives and few false negatives is obviously very high.<\/p>\n

Unsupervised Learning<\/h3>\n

Unsupervised Learning<\/strong>\u00a0is for situations where you have a data set but no labels. Unsupervised learning takes the input set and tries to find patterns in the data, for instance by organizing them into groups (clustering) or finding outliers (anomaly detection). For example:<\/p>\n