What is the difference between dlnetwork and serisenetwork about deep learning?

Hi

I am studying about neural networks. I am not sure of the difference between dlnetwork and serisenetwork. Please tell me what the difference is between them.

Answer 1

[Shantanu Dixit]

Hi Jun,

Both 'dlnetwork' and 'SeriesNetwork' are used to specify deep learning architectures in MATLAB. However, starting from MATLAB R2024a, 'SeriesNetwork' objects are not recommended. Instead, MathWorks recommends using 'dlnetwork' objects due to the following advantages:

• Unified Data Type: 'dlnetwork' objects provide a unified data type that supports a comprehensive range of functionalities, including network building, prediction, built-in training, visualization, compression, verification, and custom training loops. This makes them highly versatile for various deep learning tasks.
• Support for Complex Architectures: 'dlnetwork' objects can accommodate a wider range of network architectures, which you can either create or import from external platforms, offering greater flexibility in model design.
• Efficient Training with 'trainnet': The 'trainnet' function is compatible with 'dlnetwork' objects, allowing you to easily specify loss functions. You have the option to choose from built-in loss functions or define custom ones, facilitating tailored training processes.
• Faster Training and Prediction: Training and prediction processes with 'dlnetwork' objects are typically faster compared to the 'LayerGraph' and 'trainNetwork' workflows, enhancing performance and efficiency.

I hope this helps clarify the difference between 'dlnetwork' and 'SeriesNetwork' and the recommended function for creating neural network architectures.

Additionally you can refer to the following MathWorks documentation on 'dlNetwork' and 'SeriesNetwork'

'dlNetwork': https://www.mathworks.com/help/deeplearning/ref/dlnetwork.html

'SeriesNetwork': https://www.mathworks.com/help/deeplearning/ref/seriesnetwork.html

[俊廈]

Is it possible that there is a difference in efficiency when studying with trainnet and trainnetwork?

(Assuming that the conditions for options and layers and the number of data are the same)