Written by Daniel T. Chang (张遵)
IBM (Retired) [email protected]

Abstract:

Concepts are the foundation of human deep learning, understanding, and knowledge integration and transfer. We propose concept-oriented deep learning (CODL) which extends (machine) deep learning with concept representations and conceptual understanding capability. CODL addresses some of the major limitations of deep learning: interpretability, transferability, contextual adaptation, and requirement for lots of labeled training data. We discuss the major aspects of CODL including concept graph, concept representations, concept exemplars, and concept representation learning systems supporting incremental and continual learning.

1 Introduction

1.1 Human Deep Learning

In human learning, deep learning [1] is an approach that involves the critical analysis of new topics and facts, linking them to already known concepts or forming new concepts, and leads to long term retention of concepts so that they can be used for problem solving in new situations. Deep learning promotes understanding and application for life. This is in contrast to surface learning which is the rote acceptance of facts and memorization as isolated and unlinked facts. It leads to superficial retention of facts and does not promote understanding or long term retention of knowledge.

...

1.2 Machine Deep Learning

In machine learning, deep learning has more than one definition. A useful, though narrow, definition [4] is: deep learning is neural networks with a large number of layers and parameters in one of four fundamental network architectures: unsupervised pretrained networks, convolutional neural networks, recurrent neural networks, and recursive neural networks Automatic feature extraction is one of the major facets, and great advantages, that deep learning has.

...

1.3 Goal and Outline

From the above discussions it should be apparent that conceptual knowledge learning and conceptual understanding are needed to elevate machine deep learning toward the level of human deep learning. We propose Concept-Oriented Deep Learning (CODL) as a general approach to achieve that goal.

2 Concept Graph

The Big Book of Concepts [9] states that “Concepts are the glue that holds our mental world together.” Without concepts, there would be no mental world. As discussed earlier, concepts [3] (e.g., dog) are general ideas derived or inferred from facts. They are abstract and broad, represented by different instances that share common attributes. A concept may contain a set of attributes that describe the concept and a set of sub-concepts that are components of the concept. Concepts may also be related by relationships. The most common relationships include isA relationships.

...

3 Concept Representations

In deep learning, feature representations are generally learned as a blob of ungrouped features. However, an increasing number of visual applications nourish from inferring knowledge from imagery which requires scene understanding. Semantic segmentation is a task that paves the way towards scene understanding. Deep semantic segmentation [17] uses deep learning for semantic segmentation.

...

4 Concept Exemplars

As is the case for deep semantic segmentation, it can be difficult to gather and create labeled concept representation datasets to use for training in supervised learning. Due to the semantically-segmented nature of concepts, a good alternative is to use concept exemplars.

A concept exemplar set is a set of one or more typical instances of a concept, possibly augmented with instances generated from the typical instances using identity-preserving transformations. Each concept is associated with at most one concept exemplar set.