Batch vs Online Machine Learning — What’s the Difference?
Table of contents:-
· Batch learning
∘ Problems with batch learning:-
∘ Disadvantages of batch learning:-
· Online learning
∘ When to use
∘ Learning rate
∘ Disadvantages
· Comparison between batch learning and online learning
Today’s growing ML culture showcases the usefulness of various types of machine learning, each employing different algorithms to function. We have already delved into the explanations of Supervised, Unsupervised, Reinforcement, and Semi-supervised Machine Learning, but there are still other types worth exploring. These different approaches contribute to the diverse landscape of machine learning and its applications.
- Both batch learning and online learning raise concerns about how machine learning algorithms learn or train, especially in a production environment.
- In simple language production is a server on which your code is going to run
Batch learning:-
Batch learning is a conventional way of training an ML model.In which you use all the data to train the model.you utilize all of your data.
There is no incremental training.incremental training is a learning technique where a model is updated and improved over a time by training it on a new data while retaining from previous learned patterns
When you are working on a real-world problem.The data will be huge.generally it is.when you are going to train such a huge amount of data on a server.It would be costly and time consuming.
In batch learning,you take up your entire data.You train your machine learning model on an offline system.Once the model is trained, you deploy it on the server.
This is the entire flow of batch learning =>
Problems with batch learning:-
Your model is static, which means it doesn’t update or learn from new data.the problem with this approach is your business scenario is evolving
for example:-
consider netflix,
You made a recommendation system for Netflix. You considered all the movies and shows that are available today.
But since Netflix adds new movies and TV shows weekly, your recommendation system should always grow in order to incorporate new movies.
Your machine learning model needs to evolve constantly with new data.If you do not do this, what is the benefit of your recommendation system.
Another example is email spam classifier,
Your email spam classifier is up-to-date as of today’s date.You put it on the server once, and after that, no training took place.You left it like that for one year.
But in one year, marketers will find new techniques. To create spam emails that do not end up in the spam folder.
If you do not keep your system up-to-date, then what's the point? It will be obsolete.
The biggest problem with batch learning is that you need to retrain your model frequently; generally, people do this periodically. What they do is take the updated data, merge it with the old data, and retrain the model with the whole new data. Once the model is trained, you test it again and put it on the server.
Basically this process which happens in cicle Repats again and again in the given period.This period could be 24 hr period, weekly or monthly or 6 monthly
Disadvantages of batch learning:-
- Handling large amounts of data: Batch learning requires loading the entire dataset into memory for training. This becomes a challenge when dealing with large datasets that exceed the available memory capacity. In such cases, it may not be feasible to load the entire dataset at once, leading to potential memory errors and system crashes. Additionally, the processing time for training increases as the dataset size grows, making it impractical for real-time or time-sensitive applications
- Hardware limitations: Batch learning can be computationally expensive, especially when dealing with complex models or large datasets. Training a model on a single machine may take a significant amount of time and may require high-performance hardware, such as GPUs or specialized processing units. However, not all users or organizations have access to such resources, limiting the practicality and scalability of batch learning in resource-constrained environments.
- Availability constraints: In some scenarios, obtaining the entire dataset required for batch learning may not be feasible or practical. For instance, data may be continuously generated or collected in streams, making it impossible to accumulate the complete dataset in advance. Furthermore, there may be situations where the data is distributed across multiple sources, making it difficult to centralize and train on all the data simultaneously. In such cases, alternative learning approaches, like online learning or mini-batch learning, are preferred.
Online learning:-
Online learning is quite unlike batch learning, which is done incrementally.
“Have you ever heard that many companies promote their products in such a way that you will use our products as much. The performance of our product will go on increasing”
Means your performance in terms of usability increase
So, have you heard this then those Companies are talking about “online learning”
So, what you do in online learning is that. You feed data to your model in small batches, sequentially. These batches are called mini batches.
After, each batch of training, your model gets better. since these batches are small chunks of data. so you can perform this training on server (in production)
That’s why it is called online learning means your model is getting trained when your model is on server
example of online learning
chat bots:-
All of you used chatbots. The chatbots by famous companies like Google Assistant, Alexa, and Siri are good. examples of online learning
Because when it is deployed on the server at the time of deployment, it is also doing prediction as chat with new data. and at the same time learning from that new data.
Youtube:-
when you scroll on YouTube and come back to the feed by clicking a particular video. After watching the video, the feed changes. It automatically gets modified according to the video. that has been watched
This is again an example of online learning.
When to use:-
1) where there is a Concept drift:-
Sometimes what happens is that you create a machine learning model for a problem. The nature of the problem keeps changing over time. There, you may use online learning.
for example, the stock market, e-commerce website
2) Cost effective:-
wherever you want to save money. If you are, working on large-scale batch learning will make you spend more money. It would be great if you ase online learning It will be cost effective
3) faster solution-
This thing is fast, no doubt, because you are doing training in mini batches. Training is fast. Overall, you have to make your system fast; this is the solution.
Learning rate:-
There is a concept in online learning called learning rate. The learning rate determines how frequently your model is going to be trained.
You don’t train your model on every incoming piece of data; by doing this, your model drastically changes, and forgets the old learned things.
Generally, you would not like to forget everything old and start learning new things. And you don’t even want to learn slowly, under this is also bad
so correct learning rate has to be kept
Setting the correct learning rate for business in online learning is a difficult task if you do not do it properly. your model. may misbehave
Disadvantages:-
- Tricky to Use: Online learning platforms and tools often require technical proficiency to navigate effectively. Some individuals may find it challenging to adapt to new software, troubleshoot technical issues, or navigate through online resources. This difficulty can hinder the learning experience and create frustration, especially for those who are not familiar with technology or have limited access to reliable internet connections.
- Risky: Online learning poses certain risks, particularly when it comes to security and privacy. Data breaches, unauthorized access to personal information, or cyberattacks on online learning platforms can compromise students’ privacy and confidentiality. In addition, online assessments and exams may be vulnerable to cheating or academic dishonesty, as it can be difficult to monitor students remotely and ensure the integrity of their work
Comparison between batch learning and online learning
