Artificial Intelligence and the idea of it has always been around be it research or sci-fi movies. But the advances in AI wasn’t drastic until recently. Guess what changed? The focus moved from vast AI to components of AI such as machine learning, natural language processing, and other technologies that make it possible.
Learning models which form the core of AI started being used extensively. This shift of focus to Machine Learning gave rise to various libraries and tools which make ML models easily accessible. Here are some common myths surrounding Machine Learning:
Machine Learning, Deep Learning, Artificial Intelligence are all the same
In a recent survey by TechTalks, it was discovered that more than 30% of the companies wrongly claim to use Advance Machine Learning models to improve their operations and automate the process. Most people use AI and ML synonymously. How different are AI, ML and Deep Learning?
Machine Learning is a branch of Artificial Intelligence which has learning algorithms powered by annotated data which learn through experiences. There are primarily two types of learning algorithms.
Supervised Learning algorithms draw patterns based on the input and output values of the datasets. It starts predicting the outputs from the training data sets with possible input and output values.
Unsupervised learning models look at all the data fed into the model and find out patterns in the data. It uses unstructured and unlabeled data sets.
Artificial Intelligence, on the other hand, is a very broad area of Computer Science, where robust engineering and technological advances are used to build systems that need minimal or no human intelligence. Everything from the auto-player in video games to predictive analytics used to forecast sales fall under the same roof using some Machine Learning algorithms
Deep Learning uses a set of ML algorithms to model abstraction in data sets with system architecture. It is an approach used to build and train neural networks.
All data is useful to train a Machine Learning model
Another common myth around Machine learning models is that all the data is useful to improve the outputs of the model. The raw data is never clean and representative of the outputs.
To train the Machine Learning models to learn the accurate outputs expected, data sets need to be labeled with relevance. Irrelevant data needs to be removed.
The accuracy of the model is directly correlated to the quality of the data sets. The quality of the trained data sets results in better accuracy rather than a huge amount of raw/unlabelled data.
Building an ML system is easy with unsupervised learning and ‘Black Box Models’
The most business decision will require very specific evaluation, to make strategic data-driven decisions. Unsupervised and ‘Black Box’ models use algorithms randomly and highlight data patterns making it biased towards patterns which aren’t relevant.
The usability and relevance of these patterns to the objective the business the focus is on are a lot less when these models are used. Black box systems do not reveal what patterns they have used to arrive at certain conclusions. Supervised or Reinforcement learning trained with curated, labeled data sets can surgically investigate the data and give us the desired outputs.
ML will replace people and kill jobs
The usual notion around any advanced technology is that it will replace people and make people jobless. According to Erik Brynjolfsson and Daniel Rock, with MIT, and TomMitchell of Carnegie Mellon University, ML will kill the automated or painfully redundant tasks, not jobs.
Humans will spend more time on decision making jobs rather than repetitive tasks which ML can take care of. The job market will see a significant reduction in repetitive job roles but the wave of ML, AI will create a new sector of jobs to handle the data, train it and derive outcomes based on the ML systems.
Machine Learning can only discover correlations between objects and not causal relationships
A common perception of Machine Learning is that it discovers easy correlations and not insightful outputs. Machine Learning used in conjunction with thematic roles and relationship models of NLP will provide rich insights. Contrary to common belief, ML can identify causal relationships. This is commonly used to try out different use cases and observing the consequences of the cases.
Machine learning can work without human intervention
Most decisions from the ML models will need human intelligence and intervention. For examples, an airlines company may adopt ML algorithms to get better insights and influence best ticket prices. Data sets are constantly updated and complex algorithms may be run on it.
But, to decide the price of a flight by the system itself has a lot of loopholes, the company will hire an analyst who will analyze the data and sets prices with the help of models and their analytical skills, not just relying on the model alone.
The reasoning behind the decision making is still a human intelligence one. Complete control should not be rested on models for optimal results.
Machine Learning is the same as Data mining
Data mining is a technique to examine databases and discover the properties of data sets. The reasons its often confused is because Data Analytics uses these data sets using data visualization techniques. Whereas, Machine Learning is a subfield which uses curated data sets to teach systems the desired outputs and make predictions.
There is similarity when unsupervised learning Ml models use datasets to draw insights from them, which is precisely what data mining does. Machine Learning can be used for data mining.
The common confusion between the two arises due to a new term being used extensively, Data Science. Most Data mining-focused professionals and companies are leaning towards using Data science and analytics now causing more confusion.
ML takes a few months to master and is simple
To be an efficient ML Engineer, a lot of experience and research is needed. Contrary to the hype, ML is more than importing existing libraries in languages and using Tensor Flow or Keras. These can be used with minimal training but takes an experienced hand to provide accuracy.
A lot of intense Machine Learning focussed products require intense research on topics and even coming up with approaches using methods that are in discussion at a university or research level. Already existing libraries solve very generic problems people are trying to solve and not really insightful data. A deeper understanding of algorithms is needed to create an accurate model with an improved f1(accuracy) score.
To sum up, there is an overlap of concepts and models in Machine Learning, Artificial Intelligence, Data Science and Deep Learning. However, the goal and science of the subfields vastly vary. To build completely automated AI systems, all the fields become crucial and play a distinct role.