The Importance of Classical Machine Learning Algorithms in Machine Learning and AI
In recent years, there has been a growing trend towards using deep learning algorithms for various machine learning and artificial intelligence (AI) problems. However, when it comes to applying these algorithms to every problem that comes our way, especially those with large amounts of data, we need to consider the limitations and constraints of deep learning.
One of the main reasons why we still need to study classical machine learning algorithms like logistic regression, tree-based algorithms, and others is because they tend to work better in certain situations. As of 2018, deep learning models are particularly effective when working with data such as images, video, audio, or more complex tasks like machine translation from one language to another. However, for simple problems like predicting the probability of an ad being clicked, using a simple model like logistic regression or a slightly more complicated tree-based model is often sufficient.
In fact, studies have shown that the incremental benefit of using deep learning models for these types of problems is minuscule, and in many cases, it's not worth it. For example, if we want to use a model that is highly interpretable, meaning we can understand why the model decided one class label over another, non-deep learning models like logistic regression or tree-based methods are often better suited for this task. This is because deep learning models can be complex and difficult to interpret, making them less suitable for applications where transparency and explainability are crucial.
Another important consideration when choosing a machine learning algorithm is the need for low latency in certain applications. In these situations, the model needs to respond quickly, often within just a few milliseconds. However, deep learning models can take significantly longer to train and evaluate than simpler algorithms like tree-based methods or simple logistic regression. This makes them less suitable for applications with low latency requirements.
In addition to these considerations, designing and training deep learning models can be complex and requires a good mix of scientific knowledge and artistic intuition. While there have been significant advances in deep learning over the years, making it work is still a challenging task, especially when dealing with extremely deep models.
So, why do we still need to study classical machine learning algorithms? The answer lies in the fact that there are many situations where these algorithms are better suited for a particular problem or application. By choosing the right algorithm for the right problem, we can often achieve better results and greater efficiency than using a deep learning model. Furthermore, having knowledge of multiple machine learning algorithms is essential for being able to choose the best approach for each specific situation.
One common mistake that many people make when learning machine learning and AI is to develop a favorite algorithm and try to apply it to every problem they come across. This can lead to suboptimal solutions and wasted time and resources. Instead, we should strive to be objective and choose the best algorithm for the particular problem at hand. By doing so, we can make better use of our knowledge and skills, and achieve greater success in our machine learning endeavors.
In conclusion, while deep learning has made significant progress over the years, it's not a silver bullet that can be applied to every machine learning problem. Classical machine learning algorithms like logistic regression, tree-based methods, and others still have their place in the world of machine learning and AI. By understanding the strengths and limitations of each type of algorithm, we can make more informed decisions about which approach to use for each specific situation, and achieve greater success in our endeavors.
"WEBVTTKind: captionsLanguage: enso the question here is why not use deep learning for every problem especially when you have lots of data because as we discussed in our deep learning chapters when you have lots of data deep learning tends to work better typically then the question here is why are we even studying classical machine learning algorithms like logistic regression tree based algorithms all of that why can't we just move to deep learning so I will answer this question with lot of cases one by one number one deep learning as of 2018 works best when you have data like images video audio or more complex tasks like machine translation of text from one language to other language etc right if you already have simple features if you already example let's take the simple problem of predicting the probability of an ad being clicked right a very simple problem right so based on the features that you already have simple models like logistic regression and slightly more complicated ones like tree based models work fairly well and the incremental benefit of using a deep learning model is minuscule is very small and for solving a real-world problem there will be many real-world constraints for example you might want the model to be highly interpretable which means you don't want the model to be black box you want to understand why a model decided one class label and not the other class label when interpretability is important using non deep learning models like logistic regression like tree based methods like radiant booster addition trees random forests etcetera are much better number one number two there will be instances where you want to respond back you get a query point X cube and you want to respond back with Y Q very very quickly in just a few milliseconds those are called low latency applications that we have discussed in our course multiple times so when you have low latency requirements deep learning evaluation time significantly more then simple algorithms like tree based algorithms or simple logistic regression right so these are just a couple of cases and even training deep learning models can get fairly complex especially when you have an extremely deep model making it work designing the right architecture is non-trivial it's it's actually it's it's it's a mix of science and art today as of 2018 by the way so there is still there are still lots of applications in in machine learning and AI where we still continue to use simpler more basic algorithms like logistic regression like boost addition trees like like random forest because it's much easier to train them they're much more relevant especially if you want interpretability of a model or when you want low latency requirements right so the actual model to be used is actually to be decided based on the real-world constraints and the real-world problem and requirements of the problem that we solve it's very very important that we don't have favorite algorithms in machine learning and deep learning a lot of people especially youngsters when they learn they have one favorite algorithm and they try to apply that algorithm to every problem please don't do it please because I I came through that pipeline myself it's very important to choose the right algorithm for the right problem that is as important as knowing more algorithms trust meso the question here is why not use deep learning for every problem especially when you have lots of data because as we discussed in our deep learning chapters when you have lots of data deep learning tends to work better typically then the question here is why are we even studying classical machine learning algorithms like logistic regression tree based algorithms all of that why can't we just move to deep learning so I will answer this question with lot of cases one by one number one deep learning as of 2018 works best when you have data like images video audio or more complex tasks like machine translation of text from one language to other language etc right if you already have simple features if you already example let's take the simple problem of predicting the probability of an ad being clicked right a very simple problem right so based on the features that you already have simple models like logistic regression and slightly more complicated ones like tree based models work fairly well and the incremental benefit of using a deep learning model is minuscule is very small and for solving a real-world problem there will be many real-world constraints for example you might want the model to be highly interpretable which means you don't want the model to be black box you want to understand why a model decided one class label and not the other class label when interpretability is important using non deep learning models like logistic regression like tree based methods like radiant booster addition trees random forests etcetera are much better number one number two there will be instances where you want to respond back you get a query point X cube and you want to respond back with Y Q very very quickly in just a few milliseconds those are called low latency applications that we have discussed in our course multiple times so when you have low latency requirements deep learning evaluation time significantly more then simple algorithms like tree based algorithms or simple logistic regression right so these are just a couple of cases and even training deep learning models can get fairly complex especially when you have an extremely deep model making it work designing the right architecture is non-trivial it's it's actually it's it's it's a mix of science and art today as of 2018 by the way so there is still there are still lots of applications in in machine learning and AI where we still continue to use simpler more basic algorithms like logistic regression like boost addition trees like like random forest because it's much easier to train them they're much more relevant especially if you want interpretability of a model or when you want low latency requirements right so the actual model to be used is actually to be decided based on the real-world constraints and the real-world problem and requirements of the problem that we solve it's very very important that we don't have favorite algorithms in machine learning and deep learning a lot of people especially youngsters when they learn they have one favorite algorithm and they try to apply that algorithm to every problem please don't do it please because I I came through that pipeline myself it's very important to choose the right algorithm for the right problem that is as important as knowing more algorithms trust me\n"
