English (United Kingdom)

https://curated-unify.zendy.io/wp-json/zendy-region/v1/featured_content/oa?rat=en

https://curated-unify.zendy.io/wp-json/zendy-region/v1/highlighted_journal/

Global: Zendy Plus annual

Presents the access of premium content as premium feature

Premium Content

Presents the keyphrase highlighting as premium feature

Keyphrase Highlighting

Presents the summarisation as premium feature

Summarisation

Insights

Presents the pdf analysis as premium feature

PDF Analysis

Presents the zaia usage as premium feature

ZAIA

Global: Zendy Plus monthly

Global: Zendy Tools annual

Global: Zendy Tools monthly

Global: Zendy Free Plan

Deep learning models suffer from opaqueness. For Convolutional NeuralNetworks (CNNs), current research strategies for explaining models focus on thetarget classes within the associated training dataset. As a result, theunderstanding of hidden feature map activations is limited by thediscriminative knowledge gleaned during training. The aim of our work is toexplain and expand CNNs models via the mirroring or alignment of CNN to anexternal knowledge base. This will allow us to give a semantic context or labelfor each visual feature. We can match CNN feature activations to nodes in ourexternal knowledge base. This supports knowledge-based interpretation of thefeatures associated with model decisions. To demonstrate our approach, we buildtwo separate graphs. We use an entity alignment method to align the featurenodes in a CNN with the nodes in a ConceptNet based knowledge graph. We thenmeasure the proximity of CNN graph nodes to semantically meaningful knowledgebase nodes. Our results show that in the aligned embedding space, nodes fromthe knowledge graph are close to the CNN feature nodes that have similarmeanings, indicating that nodes from an external knowledge base can act asexplanatory semantic references for features in the model. We analyse a varietyof graph building methods in order to improve the results from our embeddingspace. We further demonstrate that by using hierarchical relationships from ourexternal knowledge base, we can locate new unseen classes outside the CNNtraining set in our embeddings space, based on visual feature activations. Thissuggests that we can adapt our approach to identify unseen classes based on CNNfeature activations. Our demonstrated approach of aligning a CNN with anexternal knowledge base paves the way to reason about and beyond the trainedmodel, with future adaptations to explainable models and zero-shot learning.

Wider Vision: Enriching Convolutional Neural Networks via Alignment to External Knowledge Bases