Minireviews
Copyright ©The Author(s) 2022.
Artif Intell Med Imaging. Jun 28, 2022; 3(3): 55-69
Published online Jun 28, 2022. doi: 10.35711/aimi.v3.i3.55
Table 1 Comparison of symbolic artificial intelligence, statistical learning and deep learning (Nadkarni P & Merchant SA)

Symbolic AI
Statistical learning (SL)
Deep learning (DL)
Entities manipulatedBoth symbols and numbersNumbers (most representing interval data, but some representing categories)Same as SL, can be applied to the same problems
Algorithm designRequires computer-science knowledge & traditional software skills, including user-interface designLess customization needed, but problem-specific pre-processing of data (e.g., statistical standardization is necessary)Same as SL
Domain expert roleWork closely and extensively with software developer, Evaluate output of algorithm for a set of test cases against desired outputTo identify variables/features of interest, annotating training data, and evaluating results and individual features’ relative importance. Must evaluate results for noveltySame as SL, but features can be discovered from raw data, so may not need designation. Annotation is more burdensome because much more data is typically needed
Data inputsExpert and software work closely to design software and create test casesRows of data, annotated text, or images. For supervised learning, the output variable’s value for each instance is also suppliedSame as SL, in some forms of DL, notably for image processing, features are discovered from raw data
Partitioning of input data(Not applicable)Divided into training data and test dataSame as SL
GeneralizabilityLimited to modest: Typically required tailored solutions, especially for the user interfaceMore generalizable than symbolic AI, but success depends on careful feature selection, choice of method and whether the data matches the method’s assumptions (e.g., Gaussian distribution, additive effects)DL methods are “non-parametric” and rely on few or no assumptions about the variables/features in the data