Distilling a Neural Network Into a Soft Decision Tree

By Nicholas Frosst, Geoffrey E. Hinton. 2017

715 citations recorded by the source corpus.

Deep neural networks have proved to be a very effective way to perform classification tasks. They excel when the input data is high dimensional, the relationship between the input and the output is complicated, and the number of labeled training examples is large. But it is hard to explain why a learned network makes a particular classification decision on a particular test case. This is due to their reliance on distributed hierarchical representations. If we could take the knowledge acquired by the neural net and express the same knowledge in a model that relies on hierarchical decisions instead, explaining a particular decision would be much easier. We describe a way of using a trained neural net to create a type of soft decision tree that generalizes better than one learned directly from the training data.

Subfields: Artificial Intelligence, Efficiency & Compression

How researchers see this paper

The researchers whose trusted networks rank it highest, strongest advocate first. Every researcher gets a different order.

In Artificial Intelligence

In Efficiency & Compression

Citations in the corpus

Dynamic Neural Networks: A Survey (2021, background): “a) Soft decision tree (SDT) [98], [99], [100] adopts neural units as routing nodes (blue nodes in Fig.”
Born Again Neural Networks (2018, background): “Further, Frosst & Hinton (2017) proposed distilling deep networks into decision trees for the purpose of explaining decisions.”
FEED: Feature-level Ensemble for Knowledge Distillation (2019, background): “This motivated researchers to develop many variants of it to various domains, and many researchers studied ways to better teach the student (Frosst and Hinton 2017; Pham et al. 2018; Radosavovic et al. 2018; Tan et al. 2018; Tarvainen and Valpola 2017).”
Neural Prototype Trees for Interpretable Fine-grained Image Recognition (2020, background): “In contrast to most SDTs that require a ﬁxed tree structure, including ours, ANTs greedily learn a binary tree topology.”
Designing Interpretable Approximations to Deep Reinforcement Learning with Soft Decision Trees (2020, positive): “(Frosst and Hinton 2017) demonstrated that the target student need not be a smaller neural network, introducing soft decision trees (SDTs) which essentially feature a single layer perceptron at each inner node.”
Interpreting Blackbox Models via Model Extraction (2017, methodology): “There has been work using model compression [5] to learn decision trees [4, 15], but they use rejection sampling, whereas our active sampling strategy directly targets paths most in need of additional data, thereby substantially improving accuracy.”
A Benchmark for Interpretability Methods in Deep Neural Networks (2018, background): “For example, one can distill or constrain a model into a functional form that is considered more interpretable [4, 12, 38, 28].”
Learning Decision Trees Recurrently Through Communication (2021, methodology): “To improve their performance, combining decision trees with neural networks has been explored by building hierarchical classifiers [46, 4, 65, 64], by transferring models [25, 55, 17, 23], and through regularization [61].”
Born-Again Tree Ensembles (2020, neutral): “Recent works focused on creating soft decision trees from a neural network (Frosst & Hinton, 2017), decomposing the gradient in knowledge distillation (Furlanello et al., 2018), deriving a class of models for self-explanatory neural networks (Melis & Jaakkola, 2018), or speciﬁed knowledge…”
Definitions, methods, and applications in interpretable machine learning (2019, methodology): “In addition to the areas below, we note that there are other emerging techniques, such as model distillation (51, 52).”
Improving the Interpretability of Deep Neural Networks with Knowledge Distillation (2018, methodology): “Perhaps the most related work is the model in [15] which uses a type of soft decision tree to mimic the input-output functions of a trained DNN.”
Abduction-Based Explanations for Machine Learning Models (2018, background): “Although the scalability of approaches for computing explanations is an important issue, the existence of a principled approach can serve to benchmark other heuristic approaches proposed in recent work (Ribeiro, Singh, and Guestrin 2016; Frosst and Hinton 2017).”
LIT: Learned Intermediate Representation Training for Model Compression (2019, methodology): “KD has inspired a variety of related methods, such as for cross-modal distillation or faster training (Gupta et al., 2016; Chen et al., 2016; Frosst & Hinton, 2017; Romero et al., 2015; Furlanello et al., 2018), but we focus on compression in this work.”
Trustworthy AI: From Principles to Practices (2021, background): “With the rise of deep learning in the past decade, explainer approximation on the DNN has advanced as the knowledge distillation problem on explainers such as trees [125, 384].”