Table 1 Differences between deep learning and traditional ML method [27]
From: Materials descriptors of machine learning to boost development of lithium-ion batteries
Aspects | Traditional ML | Deep learning |
|---|---|---|
Architecture and representation | Typically relies on feature engineering where human experts manually extract relevant features from the data | Neural networks with multiple hidden layers are involved in the process. They automatically learn a hierarchical representation of the data through the training process, with a relatively weak dependence on feature engineering |
Model | Primarily based on supervised learning, unsupervised learning, or semi-supervised learning paradigms | Commonly associated with supervised learning, where the model is trained on labelled large datasets |
Data requirements | Often requires a set of handcrafted and high-quality feature data | After training, the model can automatically identify features from raw data, making it adaptable to diverse and complex datasets and capable of handling large quantities of data effectively |
Computational capability requirement | Lower | Higher |
Interpretability | Models are usually easier to explain and to understand the reasoning behind predictions | Deep neural networks with many layers may be seen as “black boxes” and therefore challenging to explain their decision-making processes |
Applicability | Suited for tasks that require explicit feature engineering and prioritise interpretability | Ideal for tasks involving large-scale complex data |