Deep Generative Models in Engineering Design: A Review

Journal of Mechanical Design, Journal Year: 2022, Volume and Issue: 144(7)

Published: Feb. 16, 2022

Abstract Automated design synthesis has the potential to revolutionize modern engineering process and improve access highly optimized customized products across countless industries. Successfully adapting generative machine learning may enable such automated is a research subject of great importance. We present review analysis deep models in design. Deep (DGMs) typically leverage networks learn from an input dataset synthesize new designs. Recently, DGMs as feedforward neural (NNs), adversarial (GANs), variational autoencoders (VAEs), certain reinforcement (DRL) frameworks have shown promising results applications like structural optimization, materials design, shape synthesis. The prevalence skyrocketed since 2016. Anticipating continued growth, we conduct recent advances benefit researchers interested for structure our exposition algorithms, datasets, representation methods, commonly used current literature. In particular, discuss key works that introduced techniques methods DGMs, successfully applied design-related domain, or directly supported development through datasets auxiliary methods. further identify challenges limitations currently seen fields, creativity, handling constraints objectives, modeling both form functional performance simultaneously. discussion, possible solution pathways areas on which target future work.

Language: Английский

---

Ten years of generative adversarial nets (GANs): a survey of the state-of-the-art

Machine Learning Science and Technology, Journal Year: 2024, Volume and Issue: 5(1), P. 011001 - 011001

Published: Jan. 17, 2024

Abstract Generative adversarial networks (GANs) have rapidly emerged as powerful tools for generating realistic and diverse data across various domains, including computer vision other applied areas, since their inception in 2014. Consisting of a discriminative network generative engaged minimax game, GANs revolutionized the field modeling. In February 2018, GAN secured leading spot on ‘Top Ten Global Breakthrough Technologies List’ issued by Massachusetts Science Technology Review. Over years, numerous advancements been proposed, to rich array variants, such conditional GAN, Wasserstein cycle-consistent StyleGAN, among many others. This survey aims provide general overview GANs, summarizing latent architecture, validation metrics, application areas most widely recognized variants. We also delve into recent theoretical developments, exploring profound connection between principle underlying Jensen–Shannon divergence while discussing optimality characteristics framework. The efficiency variants model architectures will be evaluated along with training obstacles well solutions. addition, detailed discussion provided, examining integration newly developed deep learning frameworks transformers, physics-informed neural networks, large language models, diffusion models. Finally, we reveal several issues future research outlines this field.

Language: Английский

---

Reinforcement Learning for Generative AI: State of the Art, Opportunities and Open Research Challenges

Journal of Artificial Intelligence Research, Journal Year: 2024, Volume and Issue: 79, P. 417 - 446

Published: Feb. 6, 2024

Generative Artificial Intelligence (AI) is one of the most exciting developments in Computer Science last decade. At same time, Reinforcement Learning (RL) has emerged as a very successful paradigm for variety machine learning tasks. In this survey, we discuss state art, opportunities and open research questions applying RL to generative AI. particular, will three types applications, namely, an alternative way generation without specified objectives; generating outputs while concurrently maximizing objective function; and, finally, embedding desired characteristics, which cannot be easily captured by means function, into process. We conclude survey with in-depth discussion challenges fascinating emerging area.

Language: Английский

---

Generative AI, Human Creativity, and Art

SSRN Electronic Journal, Journal Year: 2023, Volume and Issue: unknown

Published: Jan. 1, 2023

Recent artificial intelligence (AI) tools have demonstrated their ability to produce outputs traditionally considered creative. One such system is text-to-image generative AI (e.g., Midjourney, Stable Diffusion, Dall-E), which automates humans' execution generate high-quality digital artworks. Utilizing a dataset of over 4 million artworks from more than 50,000 unique users, our research shows that substantially enhances human creative productivity by 25% and increases the value as measured likelihood receiving favorite per view 50% time. While peak artwork content novelty (focal objects object relationships) time, average declines, suggesting an expanding but inefficient space. Additionally, there consistent reduction in both visual (pixel-level stylistic elements). Importantly, AI-assisted artists who can novel ideas, regardless overall before adoption, peers evaluate favorably. The results imply ideation likely filtering are necessary skills process, thus giving rise "generative synesthesia" - harmonious blending senses mechanics discover new workflow. Lastly, adoption decreased capture (favorites earned) concentration among adopted.

Language: Английский

---

Generation of Creative Game Scene Patterns by the Neutrosophic Genetic CoCoSo Method

Studies in Informatics and Control, Journal Year: 2022, Volume and Issue: 31(4), P. 5 - 11

Published: Dec. 19, 2022

Creative procedural generation requires, by its nature, a nondeterministic result.Each set of outputs has to satisfy the required criteria and, at same time, show unique results be considered creative.One keys increasing nondeterminism is use fuzzy neutrosophic sets.A combination CoCoSo multi-criteria decision algorithm with genetic generate game scenes was proposed.The satisfies rules and improves aesthetics value visual scene.The proposed approach seeks produce creative solution which replayability principle, when not optimal in optimization procedure sense.In this paper, improvements method were order work iterative algorithm.

Language: Английский

---

Open-endedness in synthetic biology: a route to continual innovation for biological design

Published: April 12, 2023

Design in synthetic biology is typically goal-oriented: aiming to repurpose and optimize existing biological functions for new contexts, augmenting with new-to-nature capabilities, or creating life-like systems from scratch. While the field has seen many advances over past two decades, bottlenecks complexity of built are emerging designs that function lab often fail when used real-world contexts. Here, we propose an open-ended approach design aims continuously generate innovative solutions capable overcoming such hurdles. The fundamental principle put forward novelty designed biology, be it a protein, genetic construct organism, at least as important how well fulfils its goal. Designing mind, rather than sole focus on optimization towards single ``best'' design, may allow us move beyond diminishing returns see performance most engineered date. Research artificial life evolutionary computing communities demonstrated embracing can automatic generation surprising challenging problems local optima. Synthetic now offers ideal playground test these ideas living explore more creative approaches design.

Language: Английский

---

Creativity in Generative Musical Networks: Evidence From Two Case Studies

Frontiers in Robotics and AI, Journal Year: 2021, Volume and Issue: 8

Published: Aug. 2, 2021

Deep learning, one of the fastest-growing branches artificial intelligence, has become most relevant research and development areas last years, especially since 2012, when a neural network surpassed advanced image classification techniques time. This spectacular not been alien to world arts, as recent advances in generative networks have made possible creation high-quality content such images, movies or music. We believe that these novel models propose great challenge our current understanding computational creativity. If robot can now create music an expert cannot distinguish from composed by human, musical entities were known at training time, exhibit conceptual leaps, does it mean machine is then creative? emergence clearly signals much more needs be done this area. would like contribute debate with two case studies own: TimbreNet, variational auto-encoder trained generate audio-based chords, StyleGAN Pianorolls, adversarial capable creating short excerpts, despite fact was images data. discuss assess terms their creativity we show they are practice learning concepts obvious based on data, hypothesize deep models, robots machines, considered, fact, creative.

Language: Английский

---

Copyright in Generative Deep Learning

arXiv (Cornell University), Journal Year: 2021, Volume and Issue: unknown

Published: Jan. 1, 2021

Machine-generated artworks are now part of the contemporary art scene: they attracting significant investments and presented in exhibitions together with those created by human artists. These mainly based on generative deep learning techniques, which have seen a formidable development remarkable refinement very recent years. Given inherent characteristics these series novel legal problems arise. In this article, we consider set key questions area for arts, including following: is it possible to use copyrighted works as training models? How do legally store their copies order perform process? Who (if someone) will own copyright generated data? We try answer considering law force both United States America European Union, potential future alternatives. then extend our analysis code generation, an emerging learning. Finally, also formulate practical guidelines artists developers working art, well some policy suggestions policymakers.

Language: Английский

---

Threshold Designer Adaptation: Improved Adaptation for Designers in Co-creative Systems

Proceedings of the Thirty-First International Joint Conference on Artificial Intelligence, Journal Year: 2022, Volume and Issue: unknown, P. 4943 - 4949

Published: July 1, 2022

To best assist human designers with different styles, Machine Learning (ML) systems need to be able adapt them. However, there has been relatively little prior work on how and when an ML system a co-designer. In this paper we present threshold designer adaptation: novel method for adapting creative model individual designer. We evaluate our approach subject study using co-creative rhythm game design tool. find that prefer proposed produce higher quality content in comparison existing baseline.

Language: Английский

---

Towards Creativity Characterization of Generative Models via Group-Based Subset Scanning

Proceedings of the Thirty-First International Joint Conference on Artificial Intelligence, Journal Year: 2022, Volume and Issue: unknown, P. 4929 - 4935

Published: July 1, 2022

Deep generative models, such as Variational Autoencoders (VAEs) and Generative Adversarial Networks (GANs), have been employed widely in computational creativity research. However, models discourage out-of-distribution generation to avoid spurious sample generation, thereby limiting their creativity. Thus, incorporating research on human into deep learning techniques presents an opportunity make outputs more compelling human-like. As we see the emergence of directed toward research, a need for machine learning-based surrogate metrics characterize creative output from these is imperative. We propose group-based subset scanning identify, quantify, processes by detecting anomalous node-activations hidden layers models. Our experiments standard image benchmarks ``creatively generated'' variants reveal that proposed scores distribution useful novelty activation space rather than pixel space. Further, found samples generate larger subsets anomalies normal or non-creative across datasets. The node activations highlighted during decoding process are different those responsible generation. Lastly, assess if images selected our method were also evaluators, presenting link between perception humans within neural nets.

Language: Английский

---