Dear Marri,

I think when you have a very less amount of data, you should not use ML techniques as your model may not be properly trained with the handful of data.

Thanks,

Sobhan

1. I think you can try with different folds of data. Along with that check the layer wise output cluster. If that doesn't vary, from my opinion it makes sense. Different folds mean from my opinion , pick randomly the train and test data. At the same time check the validation output and testing output results to check whether it is over fitting or not.

2. I think if the extracted features are perfect then not necessary to go for AI/ML model. Conventional techniques are enough. From my point of view, as the features are predefined and correct, AI has not that much to do.

TIA

I completely agree with above response. You cross validate your model. If you want to push it, test the model with some data that are not from the same origin. If your cross validation performance is good and you can achieve good results even in new data then there is no need for sophistication.

https://www.quora.com/Should-a-machine-learning-beginner-go-straight-for-deep-learning

https://medium.com/@integrate.ai/the-pragmatists-guide-to-applying-machine-learning-in-the-enterprise-586adbd3e740

https://medium.com/iotforall/the-difference-between-artificial-intelligence-machine-learning-and-deep-learning-3aa67bff5991

https://developers.google.com/machine-learning/guides/rules-of-ml/

ML and AI are admittedly buzzwords. However, I wouldn't overestimate them. It must be always clear when to use them. Otherwise, you lose calculation speed and end up with many fancy AI models and ML algorithms, but can't explain why you need them.

In regard to your question, I wouldn't use ML/AI, for example:

1. If you have sparse data, you may prefer to use simple rules to explore your data (you have only 5 samples of weather reports, but 10 folds in cross validation);

2. If your data is not controversial (for example, if weather is not foggy then weather is always sunny), you would abstain from ML/AI and and use the corresponding rule.

IMHO the best "fresh" answer is given by Geffner'sPreprint Model-free, Model-based, and General Intelligence

Dear Marri,

I think when you have a very less amount of data, you should not use ML techniques as your model may not be properly trained with the handful of data.

Thanks,

Sobhan

According to Dr. Yaser S. Abu-Mostafa, there are three fundamental prerequisites for ML.

• A pattern exists. If there is no pattern there is nothing to look for.
• We cannot mathematically pin down a general expression for the pattern. Therefore an approximation of the actual function is required.
• We have to have data.

There are some problems that are so well characterized that machine learning adds nothing and may introduce new flaws.

Machine learning problems are generally about fitting a statistical model to a limited number of samples from a high dimensional, unknown distribution.

So, to use machine learning your problem should be:

-data driven

-involving fitting a model to predict, cluster or compress something

-be high dimensional

-one where you done mind an approximate solution

Do not use machine learning if:

-you have an optimization problem for which you have all the information at hand

-you have a model you want to fit on a low dimensional data set (you can use more rigorous statistical methods or maybe an incredibly basic ML method like K-Nearest Neighbours)

-you require that the algorithm give you the correct answer 100% of the time.

-you don't have any data

For instance, we can do sorting with Neural Turing Machines. But if we already know that sorting is needed, the problem is well characterized enough to implement a fast, exact algorithm ourselves.

The map of "Applying Machine Learning " topic

file topic_report.docx = 20 topics from 6532 articles which have words

ti=(Machine Learning) AND ts= (appl*)

in their titles or abstracts or keywords. Each topic is represented through 20 words and 20 thrases with which it is discussed in these articles. Really this terms are the names of methods, objects, properties, laws and so on for topic in question. In addition each topic has quotes from two articles in which it is most manifested.

file a1_basic.xlsm - articles with basic knowledge on the topic

file a3_novelty.xlsm - articles of last years potentially with novelty

We will be grateful for the feedback. Feel free to ask an analysis of a specific topic.

Imo,

According to Tom Mitchell's definition on machine learning: "A computer is said to learn from experience E with respect to some class of tasks T and performance measure P, if its performance at tasks in T as measured by P improves with experience E.

So if it fails to satisfy one of these requirements: not enough training data (E), task T is not well described or if its hypothesis' features are not relevant, it's not the time to use ML.