Binary decision diagrams (BDDs) and zero-suppressed BDDs (ZDDs) are data structures to represent a family of (sub)sets compactly, which can be used as succinct indexes for families. To build a BDD/ZDD representing a family of sets to index, there are various transformation operations whose inputs are BDDs/ZDDs and output is a BDD/ZDD of the family after performing a set operation on the families represented by the inputs. However, the worst-case time complexity of these transformations is unknown for most of the operations. In this paper, we prove that most operations cannot be performed in worst-case polynomial time in the sizes of input BDDs/ZDDs. These include all operations in family algebra raised by Knuth, except for some basic ones. Moreover, these also include some operations that were claimed to be performed in polynomial time by previous works. These results give us guidelines for constructing a BDD/ZDD representing a desired family of sets.

In recruitment interviews, interviewers sometimes make comments that include gender-based judgments, posing a problem for fair recruitment. In this paper, we propose a system that uses GPT-4o to provide feedback to the interviewer during a Web-based interview when the interviewer makes comments that include gender-based judgments. The system provides feedback to the interviewer when a statement by the interviewer includes a gender-based judgment, highlighting which part of the statement included the gender-based judgment, and a comment from the avatar suggesting how the statement should be reinterpreted. To examine avatar presentation, we conducted three surveys on suitable avatar types and their emotional expressions. Results showed that a human avatar was preferable to an animal avatar. Additionally, when the avatar comments, we found that the emotional expression of anger was not appropriate, but there were individual differences in the expression of persuasive emotions.

In recent years, approaches that use the likelihood of a language as a means of distinguishing between human and generative model text have attracted a great deal of attention. In this study, we develop a method to detect unconscious bias in human-generated text by referring to the method developed there. We propose a method that extracts frequency components from a series of log-likelihoods of text and extracts features related to the existence of bias by processing in the frequency domain. The proposed method did not exceed the accuracy by the embedding vector. In the future, we would like to verify the performance of the proposed method by adjusting the frequency components that are input.

In this paper, we introduce a machine learning model to estimate the faulty parts of gas appliances from repair request information and customer data. Traditionally, technicians estimate which parts are faulty based on a customer request, preparing the necessary parts, and visiting the site. However, this approach relies heavily on the experience of the technicians and carries the risk of requiring a return visit. Additionally, a return visit may lead to a decrease in customer satisfaction. To address these challenges, we developed a machine learning model using gradient boosting trees and implemented an algorithm to aggregate necessary parts from past cases. Furthermore, we built a system to display these estimation results to the technicians and put it into operation. Experimental results and evaluations after system implementation suggested the effectiveness of the proposed method.

Large Language Models (LLMs) have low accuracy for long-tail knowledge, and solutions include Retrieval Augmented Generation (RAG) and fine tuning. On the other hand, there are reports that it is not practical due to factors such as differences in the evaluation task. In practical applications, more complex tasks can be solved and it is necessary to understand the context. However, it is not clear how LLMs internally processes texts and captures the context in specific domains such as civil engineering, which is a long-tail domain. We believe that clarifying how LLMs capture context in specific domains will contribute to improving for long-tail knowledge. In this study, we first identified uncertainty in response generation in both the general and specific domains, and then clarified the features of LLM internal processing processes in a specific domain by analyzing the variation of entropy in the intermediate representation of each layer of the LLM.

Unsupervised visual anomaly detection is promising for industry, due to the requirement of training only on normal images. To apply unsupervised approaches in a practical scene, we address covariate shifts by environmental changes such as lighting variations and equipment deterioration. Since embedded machines used in manufacturing processes have limited resources, the machines cannot store many images for retraining. Thus, retraining on a large dataset is impractical. In this paper, we propose an adaptive incremental learning method that avoids retraining the large dataset. The proposed model deals with the mean and covariance matrix. Several sets of mean and covariance matrix are combined by weighted linear sum. Each set is trained on a small dataset stored in the embedded machine. As a result, the proposed method achieves superior AUROC even when only a smaller dataset is trained.

In training deep generative models, it is essential how to approach the generated distribution to the data distribution. The most used measure is KL divergence, however, it has the problem that the gradent based optimization is difficult when the two distributions do not have any intersections. Flow matching is a deep generative model whose application has been widely studied, it can be regarded as minimizing the upper bound of KL divergence as diffusion models do. Then, we propose a method to minimize Wasserstein distance in flow matching formulation. We introduce a discriminator to calculate Wasserstein distance, and approach the time derivative of the distance to 0. In 2d dataset experiments, we verify that the proposed method minimizes Wasserstein distance by training. Although the proposed method would not outperform the original flow matching, we investigate the effect of the discriminator performance by using entropic regularization.

In the parameter adjustment of machine tools, it is useful to estimate the parameter region where evaluation values, such as the quality of the workpiece, exceed a predetermined threshold. This approach provides parameters that are robust to external factors (e.g., temperature) that affect processing. The quality standards have two thresholds: a good machining standard and a poor machining standard. If the machining quality exceeds the good machining standard, it is classified as the upper region; if it falls below the poor machining standard, it is classified as the lower region; and anything in between is classified as the intermediate region. In this study, we define termination criteria for three-class level set estimation that considers estimation errors in the intermediate region. We then propose a new acquisition function based on these termination criteria. We validated the termination criteria through numerical experiments by comparing them with accuracy. Additionally, compared to conventional methods, the proposed method reached the termination criteria with fewer search iterations. This is expected to improve machining quality and reduce costs in machine tools.

The emergence of "generative AI" that produces text, images, and music has intensified discussions about creativity. Analyzing the essence of generative AI models through Floridi's critique reveals that the "works" generated by these models are merely ectypes derived from existing works that serve as archetypes. This clarifies the position of current generative AI models within Boden's three categories of creativity (combinational, exploratory, and transformational creativity): their creativity remains primarily at the combinational level. Given that humans not only change technology but are also changed by it, it is crucial to investigate the possibilities of transformational creativity through human-AI blending rather than questioning AI's creativity in isolation. From this perspective, this paper proposes a new design concept for AI tools that could lead to transformational creativity: the potential for creative appropriation or misuse, which allows for flexible use beyond its intended purposes. Investigating this design concept requires an integrated approach of theory and practice through collaboration among philosophers of technology, artists, and developers.

This presentation reconsiders creativity in view of the development of generative AI. Creativity can be broadly categorized into two types: allopoiesis and autopoiesis. In allopoiesis, the creative subject and object are distinct, whereas in autopsies, the creative subject and object are identical. However, this identification is not self-sameness but ever process of becoming itself. While AI ethics concerning creativity often focus on allopoiesis, they are fundamentally rooted in issues related to autopoiesis. Since autopoiesis concerns not the creation of products but the ongoing becoming of the self, the essence of autopoietic creativity lies in the time spent engaging in creative activities that foster self-transformation. Although technical automation and streamlining of production processes may shorten this time, from the perspective of autopoiesis, what is critical are technics that enable the generation of such time. This is not merely about reducing working hours to increase leisure but about the genesis of time itself: the life of time where time emerges through processes of creation. This resonates with concepts such as the embodied mind and enactivism, for addressing this question unites creation and knowledge. This presentation, therefore, raises the issue of creating time for fulfillment within the context of artificial intelligence and creativity.

Creativity in people and their creations involves two key concepts: novelty and value. A creation is creative when it meets both. With LLM-based generative models, we can produce diverse, novel music via language manipulation. However, a musical work’s value is determined externally. To address the value aspect of creativity, the generative model must learn value criteria through iterative music generation and evaluation. We propose a system using LLM-based in context learning to learn and apply such criteria in repeated loops. The system forms hypotheses about potentially high-rated music, generates accordingly, and refines its approach via feedback. Experiments under multiple value criteria showed it can produce high-value music by inferring and hypothesizing those criteria. However, inferred values may not match actual criteria, reflecting dependence on the LLM’s training. Experimental results suggest that, by trial and error, a system can interpret external standards and generate valuable music.