Artificial Intelligence and Applications

A.EDGE DETECTION

Among many topics in the field of computer vision and image processing research, edge detection plays an important role in a wide range of fields from satellite imaging to medical screening, object recognition. It is an image processing technique to find boundaries/edges in a digital image with discontinuities that present a global view of an image with the most critical outline of the image. Robust edge-based shape features provide more concrete analysis for computer vision applications.

Traditional edge detection methods exploit low-level visual cues to construct hand-crafted features then classify the edge and nonedge pixels using threshold-based methods. The results of conventional approaches lack semantics at the object level. Nowadays, convolutional neural network-based approaches have become mainstream in the image processing domain. Among deep network-based edge detection methods, holistically nested edge detection (HED) [1] is one of the successful frameworks. It produces five intermediate side outputs and performs deep supervision along the network pathway. Its final fused result achieves performance within a 2% gap to human vision. Since then, several approaches use a similar architecture to further improve the accuracy. These efforts mainly focus on improving the quality of intermediate outputs or enhancing the deep supervision strategies. However, these approaches fuse intermediate layers without considering hierarchical edge importance within each side output. This poses the dilemma to the network: to include desired features, it has to accept many unwanted data and vice versa. Consequently, the result often contains more noise and thick edges while missing some key boundaries.

To tackle this issue, Scale-Invariant Salient Edge Detection (SISED) framework [2] can locate and extract the important Scale-Invariant Salient Edge (SISE) as the subset of each side output without increasing the network complexity. The normalized Hadamard Product is the key operation of SISED where a multiplicative operation is applied to promote mutually agreed features across multiscale side outputs while suppressing those with weak scale expression. SISED computes the edge importance hierarchically to enhance the edge results and reaches state-of-the-art performance.

Moving forward, the research of finding robust edges is still ongoing. Recently, vision transformer (ViT)-based model [3] demonstrated better performance and greater efficiency than CNNs on image classification. With its powerful tokenized queries, self-attention mechanism, and encoding-decoding strategy within transformers, it is possible that ViT framework can be applied into edge detection. For example, visual saliency transformer (VST) [4] is able to extract object contours and shows a new paradigm for transformer-based edge detection models.

B.BERT-BASED NEGOTIATION CHATBOT

Business negotiations are often hard due to conflicts of involved parties. Some negotiations can be not only time consuming but also negative, resulting in the damages of business relationships when unexpected negative emotions grow [5]. A solution to these problems is to automate negotiations with a robot, a negotiation chatbot using the BERT framework, which was developed for NLP.

The BERT (Bidirectional Encoder Representations from Transformers) is a deep learning natural language representation model [6] that has a powerful bidirectional prediction and contextual understanding feature. This model involves two main steps: pretraining and fine-tuning data. The BERT model is pretrained on an enormous amount of unlabeled data. The model allows high performance when it is fine-tuned to a specific task through additional training. The BERT model trains data by performing 2 tasks: MLM (Masked Language Model) that is used to predict the missing word(s) in close vicinity within a sentence, and NSP (Next Sentence Prediction) that is effective for the question/answer task.

The first step for this chatbot is fine-tuning the model to the negotiation task, using more than a thousand bilateral negotiations experimentally conducted globally [7]. Next, the MLM is extended to predict whether a negotiation warranted a positive or negative result. Finally, the NSP is used to generate automated responses to “chat” with a human counterpart in a negotiation setting. This BERT-based negotiation chatbot could be utilized as a business representative who can sense and foresee the positive and negative atmosphere during the negotiation, and strategically react to the opponent. The goal is to help the two parties reach a potential agreement that both parties are satisfied with.

C.PHARMBERT: A PRETRAINED LANGUAGE MODEL FOR PHARMACEUTICAL ERROR PREDICTION

Total number of retail prescriptions filled annually in the USA has reached 4.69 billion in 2021 [8]. However, the tracking over the service quality of the dispensation process is still very limited [9]. In an effort to address factors that lead to quality-related events, some healthcare organizations and governments adopt error-reporting systems. Such reporting systems have collected pharmaceutical errors that either reach patients (incident events), such as incorrect drug, dose, or quantity, or are intercepted at pharmacies (near miss events) [10].

To discover common contributing factors that may have led to quality-related events, large-scale analysis of these event is crucial [11]. Many common factors in retail pharmacies that resulted in an incident may not be obvious to the human eye and traditional data-mining solutions. With the progress of deep learning in NLP, the development of effective mining has been boosted, including the field of extracting valuable latent information from pharmaceutical documents.

In this research, BERT is utilized to make predictions on the pharmaceutical transaction data (collected by a Canadian error-reported system). To fit pharmaceutical data with the BERT model, the event information are formatted into Natural Language tokens and fine-tuned on the pretrained BERT model. The trained pharm BERT model is able to achieve an accuracy of ∼84% when predicting whether an event would result in a near miss (caught beforehand) or an incident (caught afterwards). By using this model, it is possible to further predict other aspects of the event, such as what stage of the events (prescribing, transcribing, dispensing, administration, storage, and monitoring) the incident occurs or what category of issues the event falls under. It is optimistic that the findings from this study could lead to solutions to reduce pharmaceutical incidents and provide improvements in patient safety.

Continuing with our strengths in supporting fundamentals in AI development and in industrial applications from the previous issue [12], we selected five papers in these two domains. In the following sections, we present an overview of the papers in this issue that support the education of AI and apply AI to extend and improve performance and capacities in several application domains. These papers include both academic research papers and industrial application papers.