Core Concepts
Natural Learning (NL) is a novel prototype-based machine learning algorithm that elevates the explainability and interpretability of classification models to an extreme level. NL discovers sparse prototypes that serve as human-friendly decision rules, enabling simple and intuitive explanations of its predictions.
Abstract
The paper introduces Natural Learning (NL), a novel machine learning algorithm that aims to achieve a high level of explainability and interpretability. NL is inspired by prototype theory from cognitive psychology, which suggests that people categorize objects based on their similarity to prototypical examples.
The key aspects of NL are:
NL discovers a single prototype sample for each class, along with the minimal set of features that characterize the prototype. This aligns with the principles of prototype theory, which states that people rely on sparse prototypes for categorization.
NL employs locality-sensitive hashing (LSH) to efficiently find the nearest neighbors of each sample, addressing the curse of dimensionality. It also uses a recursive feature pruning method to identify the core features of the prototypes.
The training process of NL is optimization-free and involves simple operations like nearest neighbor search and element-wise vector comparisons. This results in NL models that are extremely sparse, both in terms of the number of prototypes and the number of features used.
The prediction process of NL is also straightforward - it assigns a new sample to the class whose prototype it is closest to, based on the selected prototype features.
The paper presents an extensive empirical evaluation of NL on 17 benchmark datasets, including high-dimensional gene expression data and low-dimensional healthcare datasets. The results show that NL achieves performance comparable to state-of-the-art black-box models like deep neural networks and random forests in 40% of the cases, with only a 1-2% lower average accuracy. Importantly, NL models are highly interpretable, using only a few prototypes and features, and exhibit the lowest model variance among all classifiers.
The paper concludes by discussing the potential applications of NL in domains that prioritize interpretability and explainability, such as healthcare, finance, and criminal justice, where the existence of prototypical cases is common.
Stats
The paper reports the following key figures:
In a colon cancer dataset with 1545 patients and 10935 genes, NL achieves 98.1% accuracy by analyzing just 3 genes of test samples against 2 discovered prototypes.
In the UCI's WDBC dataset, NL achieves 98.3% accuracy using only 7 features and 2 prototypes.
On the MNIST dataset (0 vs. 1), NL achieves 99.5% accuracy with only 3 pixels from 2 prototype images.
Quotes
"NL simplifies decisions into intuitive rules, like 'We rejected your loan because your income, employment status, and age collectively resemble a rejected prototype more than an accepted prototype.'"
"NL efficiently discovers the sparsest prototypes in O(n^2pL) with high parallelization capacity in terms of n."
"Evaluation of NL with 17 benchmark datasets shows its significant outperformance compared to decision trees and logistic regression, two methods widely favored in healthcare for their interpretability."