I want an accurate scientific explanation.
If I have got question correctly, you ask about prioritization between computational performance (e.g. time for single inference) and prediction quality.
This question is controversial. As shown, for example, in the YoloV7 paper (Preprint YOLOv7: Trainable bag-of-freebies sets new state-of-the-art ...
) while increasing the quality of prediction, the computation cost increases too, but it has a place to be during comparison between models with similar or close architecture. There is not always same for models from different generation, but the main Idea that in a real task, it is impossible to follow maximization of an only single parameter.Thus, the computation performance is critical for real-time application (for example, face recognition, object detection for vehicle), meanwhile, if inference time is not strongly limited by end-task (for example, medical image analysis (CT, MRI, etc.)) higher priority may be given to prediction quality.
Summing up, the limitation of target computation performance and prediction quality (accuracy or other metrics) should be defined as requirements during the research design stage.
To evaluate a machine learning model you have to calculate many metrics using the confusion matrix. The accuracy and the sensitivity are considered the most importnt metrics among them.
Hi, you could check the following research using unsupervised machine learning in plant diseases detection. In methods, you could see how we could define the accuracy model.
Article Abaxial leaf surface-mounted multimodal wearable sensor for ...
As it is the case for most of such questions, the answer is: it depends. You need to consider the requirements of the application that your machine learning model should serve:
- For safety-critical applications, the value of accuracy might be higher than performance.
- For time-critical systems or systems that run on mobile devices, speed might be valued higher.
Please also note that neither accuracy nor performance are well-defined metrics. There are various ways of measuring accuracy, and the correct evaluation metric is dependent on your problem at hand. For performance, there are requirements in terms of hardware, memory, execution speed and so on.
Finally, there are other terms that should be looked into when designing a machine learning model for a specific application:
- Robustness: Is the model able to deal with unexpected input?
- Trustworthiness: Can we trace how the model arrived at its prediction?
- Fairness: Are all minority groups treated equally by the machine learning model?
Thanks a lot Dear Dr. Aleksandr Sinitca I really appreciate your contribution.
Thank you very much for your valuable ideas Dr. Dina A. Ragab .
Thank you very much for your valuable ideas Dr. Aparna Sathya Murthy .
Dr. Sina Jamalzadegan Thanks for this important answer, I hope this link is very useful for me.
Thank you very much for your valuable ideas Dr. Manuel Günther .
Thank you very much for your valuable ideas Dr. António José Rodrigues Rebelo
.
Thank you very much my dear Dr. Thaer Dawood Salman for your kindly words, once again thank you for your comment.
Assis.Prof.Dr. Zubaidah Abdulwahab Aldabbagh Zubaidah, I’m going to suggest thinking about António José Rodrigues Rebelo
's answer. What he posed may seem unexpected but in my experience it addresses your question. And, gives you a way to learn about the questions vs underlying assumptions. I’ve found, across many areas research and model /prototype design, the moment we separate two intertwined concepts we also lose sight of how they work individually and together. Focusing on one part, usually for ease of study, more often because we’re taught to do research on parts vs the whole, we diminish the chance of the model either performing well or being accurate.Also, I’m quite dismayed when you say, “I want an accurate scientific explanation” while your initial question says you’ll consider performance over accuracy. So, if you don’t mind, I’m changing my opinion; with no context, in its current form, this is not a useful question. For example; is this related to life saving information, predicting social interactions, analyzing a survey or finding good coffee ??
Thank you very much for your valuable ideas Lee Clawson .
- Badges
- Science topic
- Similar topics
- Linguistics
- Composition Studies
- Publication