Machine Learning (ML), as one of the key driver of Arti cial Intelligence, has demonstrated disruptive results in numerous industries. However one of the most fundamental problem of applying ML, and particularly Arti cial Neural Network models, in critical systems is its inability to provide a rational of their decisions. For instance a ML system recognizes an object to be a warfare mine through comparison with its similar observations. No human-transposable rationale is given, mainly because common sense knowledge or reasoning is out-of-scope of ML systems. We developed an asset, combining ML and knowledge graphs to expose a human-like explanation when recognizing an object of any class in a knowledge graph of 4,233,000 resources.
The current hype of Arti cial Intelligence (AI) mostly refers to the success of Machine Learning (ML) and its sub-domain of deep learning. However industries operating with critical systems are either highly regulated, or require high level of certi cation and robustness. Therefore, such industry constraints do limit the adoption of non deterministic and ML systems. Answers to the question of explainability will be intrinsically connected to the adoption of AI in industry at scale. Indeed explanation, which could be used for debugging intelligent systems or deciding to follow a recommendation in real-time, will increase acceptance and (business) user trust. Explainable AI (XAI) is now referring to the core backup for industry to apply AI in products at scale, particularly for industries operating with critical systems.
This is particular valid for object detection task in ML, as objet detection is usually performed from a large portfolio of Arti cial Neural Networks (ANNs) architectures such as YOLO trained on large amount of labelled data. In such contexts explaining object detections is rather di cult due to the high complexity (i.e., number of layers, lters, convolutions phases) of the most accurate ANNs. Therefore explanations of an object detection task are limited to features ? Copyright c 2019 for this paper by its authors. Use permitted under Creative Commons License Attribution 4.0 International (CC BY 4.0). involved in the data and model e.g., saliency maps [1] or at best to examples [2], or prototypes [3]. They are the best state-of-the-art approaches but explanations are limited by data frames feeding the ANNs.
We present a system expanding and linking initial (training, validation and test) data (for a ML object detection task) with entities in knowledge graphs, in order (i) to encode context in data, (ii) to capture complex relations among objects, classes and properties and even (iii) to support inference and causation, rather than only correlation. 2
Explaining Objet Detection with Knowledge Graphs Training Dataset
Data: Open
Image
Dataset v4 Task
Object Detection
Task
Image s Labels
Training Process
CNN: Faster
RCNN Pre-trained inception R model esnet v2
Object Detection
Model
Knowledge Graph Selection Knowledge
Graphs
First step detections Selected KG,
Labels Vocabulary: Wikidata, DBpedia, YAGO.
50% Dictionnary of Context
Augmentati on of 'Paddle' Score
Semantic Augmentation of Confidence
Context
Generation
74
%
Augmented
detections
" 'Paddle' confidence is
augmented as class 'Boat' and
'Canoe’. are in both (
Explainable Layer (ML) Training Process: We selected Faster RCNN (Region CNN, designed for object detection) Inception Resnet v2 [4]. Due to resources required to train a neural network on this dataset, we used pre-trained detection model. Among the pre-trained models on OID v4 available online, the faster RCNN with Inception Resnet v2 is the best tradeo between detection performance and speed. (ML) Object Detection Model: The con guration of the faster-RCNN is as following: The region proposal networks suggest 100 regions, with non maximum suppression IOU (Intersection-over-union) threshold at :7, and no NMS (Nonmaximum suppression) hyper-parameters score threshold. The second stage of the RCNN infers detections for these 100 regions, with no additional NMS. These 100 predictions are the baseline of our work. (ML) Object Detection Task: Our method is tested on a subset of the Open Image v4 Validation Dataset (described below). We designed our system to detect objects among the 600 categories of the Open Image challenge. All those categories are used to drive the search in Knowledge graph, and extract contextual information for each one of these categories to augment the baseline object detection approach.
Knowledge Graph Selection: Knowledge graphs are selected based on overage of labels of Open Image v4 Validation Dataset and an initial set of knowledge graphs: Wikidata, DBpedia, YAGO. Fuzzy match are used to optimize coverage, and the knowledge graph with the highest coverage is used for the further steps. Context Generation: The task is to detect and localize objects out of a set C of 600 categories. For each category in c 2 C we determine a list of other categories C0 that are `close enough' (i.e., through identi cation of direct link among entities in the knowledge graph) to c so that if they are detected simultaneously, then we can con dently increase detections scores. The main di culty is to identify a unique resource in the graph that will correspond to the given category. But often, several web resources can represent the same category, and categories can also have several homonym in the knowledge graph. Towards this challenge DBpedia is the most e cient graph to extract a unique resource associated to a category. Indeed, duplicates are quasi-null, disambiguate pages enabling to di erentiate between homonyms, and redirection property enable to deal with synonyms issues. Moreover, it has a propriety that redirects every dbpedia resource toward YAGO and Wikidata. This is extremely interesting as it is hard to obtain a unique resource from a category name using wikidata itself, and it can be useful to combine several knowledge graph. The output of this process is a dictionary, where every key is a label of our detection task, and the value is the subset of the labels that are contextually linked to it.
Semantic Augmentation of Con dence: We obtain: 100 predictions with
bounding boxes and a contextual dictionary extracted from the knowledge graph.
The process is as following: (
Demonstration We illustrate our demonstration (Figure 2) on an image from the open image validation dataset. We feature detections with con dence score higher than :4 before and after the semantic augmentation of con dence.
Evaluation: We evaluate detection performance with the Open Images mean Average Precision @0.5 score used for the Open Image detection challenge3. The Average Precision score for a category is derived as the era under the Precisionrecall curve for a speci c IoU threshold (here 0.5). Precision measures how accurate is the predictions. i.e. the percentage of the predictions are correct and recall measures how good the identi cation of positive is. The mean Average Precision or mAP score is calculated by taking the mean AP over all classes. Results: mAP 0:5 = 49:5% with baseline, mAP 0:5 = 49:9% with our approach. Our semantic augmentation con dence slightly improves the average detection performance of the model, while providing an interpretable layer due to the use of external information extracted from knowledge graphs. 3 https://storage.googleapis.com/openimages/web/challenge.html