21 Feb Do Machines Actually Learn?
Author – Dr.. Srinivas Kilambi, Founder & CEO, Sriya DXI, LLC
An article recently was posted on LinkedIn titled “Machines do not learn”. Its contents and some of the comments made me think of writing this article to post my 2 cents on whether “Do machines actually learn?”.
The million $ answer is a “Qualified Yes”. Qualified because the they do learn what is taught and the extent and quality of learning depends upon the quality of teaching and the teacher. I am sure most of us can relate to this. In our school days, we learnt better from better teachers and better teaching methodology. This is the universal truth and the same holds for machines and for another sub class of AI namely neural Networks and Expert Systems.
There are 3 types of machine learning,
- Supervised Learning
- Unsupervised Learning
- Semi-Supervised Learning
Supervised learning is the task of inferring a pattern from pre-labeled training data. The training data consist of a set of numerous examples with known outcomes. In supervised learning, each example consists of an input parameter with a value and a desired output value. The goal of Supervised Learning is to build a model that makes predictions and prescriptions based on teaching material presented. Some of the algorithms include Support Vector Machines, Mutual Information, Naive Bayes, Logistic Regression, Decision Trees and Random Forests. Their principal drawback is that they are as good as what is taught and cannot infer hidden patterns.
Unsupervised Learning is the task of inferring patterns or functions from input data sets without any pre-labeled data or information. The goal here is try to find hidden structure in unlabeled data. This is a “True black box approach”. However its limitation is that there is no error or reward indicator/s to evaluate the outputs, or outcomes or solutions. K-means clustering, EM, PCA, Cobbweb, DBScan and Optics are some of the unsupervised learning algorithms with K-Means being the most widely used.
Semi-Supervised Learning is a combination of both supervised and unsupervised learning, typically a small amount of pre-labeled data with a large amount of unlabeled data. Generative models, graph based models, heuristic models and low density separation models are used in this approach. Semi-Supervised learning is arguably the most complicated of the 3, least used and most practical usage of machine learning as neither 100% teaching nor 100% black box approach is always 100% perfect but a combination is a safer bet.
Finally, data sampling, storage, management and refreshing mechanism or methods also determine which algorithms will work well for a given set of data.
I hope this articles throws some more light on the trendy machine learning topic.