Conceptual alignment deep neural networks

Abstract

Deep Neural Networks (DNNs) have powerful recognition abilities to classify different objects. Although the models of DNNs can reach very high accuracy even beyond human level, they are regarded as black boxes that are absent of interpretability. In the training process of DNNs, abstract features can be automatically extracted from high-dimensional data, such as images. However, the extracted features are usually mapped into a representation space that is not aligned with human knowledge. In some cases, the interpretability is necessary, e.g. medical diagnoses. For the purpose of aligning the representation space with human knowledge, this paper proposes a kind of DNNs, termed as Conceptual Alignment Deep Neural Networks (CADNNs), which can produce interpretable representations in the hidden layers. In CADNNs, some hidden neurons are selected as conceptual neurons to extract the human-formed concepts, while other hidden neurons, called free neurons, can be trained freely. All hidden neurons will contribute to the final classification results. Experiments demonstrate that the CADNNs can keep up with the accuracy of DNNs, even though CADNNs have extra constraints of conceptual neurons. Experiments also reveal that the free neurons could learn some concepts aligned with human knowledge in some cases.

A multimodal deep architecture for traditional Chinese medicine diagnosis

“…A major problem to build a deep architecture for TCM diagnosis is the model interpretability. In order to improve the notorious interpretability of DNN models, we propose to use conceptual alignment deep neural networks (CADNNs), 12 which have been tested with fashion‐MNIST dataset, to embed prior conceptual knowledge and produce interpretable representations in the deep models. By adding constraints to objective function and using a hierarchically supervised training method, CADNNs can align the representation space of some hidden neurons with human‐formed concepts almost without loss of accuracy.…”

“…The extensions of this article are mainly twofold. Firstly, we have improved the multimodal deep architecture and increased its interpretability by introducing CADNNs 12 . Secondly, we have added a text processing channel and enhanced the experiments for an image‐text conjoined dataset.…”

“…For the deep architecture based on DNNs, a big problem is the bad interpretability. In order to improve the interpretability of DNNs, we use a hierarchically supervised training architecture, termed conceptual alignment deep neural networks (CADNNs), 12 to embed the TCM concepts and obtain interpretable representations in the hidden layers.…”

Recognition of Crop Diseases Based on Depthwise Separable Convolution in Edge Computing

“…The basic principle is as follows: with the transfer learning (TL), the “prior knowledge” obtained from “non-single” datasets is applied to CNN training of domain-specific recognition to alleviate the overfitting problem caused by insufficient data volume in a specific domain. In this paper, TL classifies crop diseases to apply the useful skills learned from one or more auxiliary domain tasks to the new targets and tasks [ 17 , 18 , 19 , 20 , 21 , 22 , 23 ]. Most of the research where the TL method is applied to the recognition of crop diseases is based on the TL method for parameter fine-tuning.…”

Soft computing and intelligent systems: Techniques and applications

“…In [36], a top down semi-supervised subspace clustering is proposed to identify a subset of important attributes based on the known label for each data instance. In [37], the authors propose a training method that employs Deep Neural Networks (DNNs) to map the sensory data into a representation space aligned with human concepts. The DNNs become interpretable with human knowledge and become easy to be trained by a small amount of training data with prior knowledge.…”