Artificial Intelligence | Online Workshop | English | North-South Americas
Description:
Deep Neural Networks are incredibly data-hungry; the high performance of these algorithms is heavily dependent upon the availability of giant training datasets. In the growing field of Deep Design and Architecture, there is a huge need for the collection/creation of comprehensive labeled datasets (in both the 2D image/plan realm as well as for 3D models).
While this workshop will focus on Dataset Generation (primarily data creation, labeling methodology and tools), we will also instruct students on how to create end-to-end machine learning pipelines. Specifically, how to formulate a desired task into a problem that is solvable with machine vision tools, how to design and generate a dataset that defines the world space for this task/problem space, and finally how one can train and evaluate an algorithm on their dataset.
With these tools, the students of this workshop will be able to participate in a very unique opportunity: they will be able to contribute as both annotators and creators to AR2IL's effort to generate a novel 2D floor plan dataset, called Common House, as well as a novel 3D mesh house dataset, called Model Mine. The students will be credited as contributors to the dataset, and have opportunities to continue as annotators on the dataset in the months after the workshop.
Overview and Syllabus:
Day 1 June Saturday 26th:
Morning:
Introduction to Dataset Creation in architecture (including an overview of AR2IL projects) — 1hr Introduction to Deep Neural Networks and why we need huge datasets — 1hr
High level overview of data-driven feature learning
Data centric vs. model-centric approach to training DNNs — ImageNet example — ImageNet doesn't have enough architectural design semantics to be applied to architecture — List examples of architecture of datasets
Demonstrate that higher quality data is better than improving model quality for performance
The basic anatomy of a dataset
Discuss how the dataset structure (and labels) vary with tasks
2D labeling, 3D labeling
Labeling Noise/Annotator Bias (what problems can arise in dataset taxonomy, labeling, data collection)
Introduction to Common house Dataset, motivation & problem space — 30 min to 1 hr
Afternoon:
Hands-On/Tutorial: Introduction to Floor Plan Annotation — 2 hrs
Day 2 June Sunday 27th
Morning:
Lecture on AI and Creativity, cultural implications — 30min to 1 hr
Q&A/ Discussion — 30min to 1 hr
Hands-On/Tutorial: Continuing Floor Plan Annotation — 2hrs
Afternoon:
Lecture on Dataset Management and Control/Versioning — 1hr
Hands-On/Tutorial: Continuing Floor Plan Annotation — 2hrs
Day 3 June Monday 28th
Morning:
Lecture on the Model Mine Dataset, problem space and motivation – 1hr
Hands-On/Tutorial: How to model objects for a dataset –2 hr
Afternoon:
Q&A with workshop leaders — 30 min to 1 hr
Hands-On/Tutorial: Model Mine Annotation tool and practice (students have the option to do 2D annotation) — 2 hrs
Day 4 June Tuesday 29th
Morning:
Conversation with cultural theorist about AI in Art and Bias in datasets— 30min
Q and A/ Discussion — 30min
Hands-On/Tutorial: continuing modeling for model mine dataset — 2hr
Afternoon:
Hands-On/Tutorial: continuing model mine or floor plan annotation — 2hr
Day 5 June Wednesday 30th
Morning:
Lecture on model evaluation and validation data — 1 hr
Hands-On/Tutorial: continuing modeling for model mine dataset or annotations — 2hr
Afternoon:
Hands-On/Tutorial: Evaluating trained models (cycleGAN trained on labels2floorplans) — 2 hrs
Parting Q and A — 1 hr
Key Words:
Artificial Intelligence,Neural Networks,Dataset Generation,Machine Learning
Required Skills:
N/A
Required Software:
N/A
Required Hardware:
PC
Maximum number of participating students:
unlimited
While this workshop will focus on Dataset Generation (primarily data creation, labeling methodology and tools), we will also instruct students on how to create end-to-end machine learning pipelines. Specifically, how to formulate a desired task into a problem that is solvable with machine vision tools, how to design and generate a dataset that defines the world space for this task/problem space, and finally how one can train and evaluate an algorithm on their dataset.
With these tools, the students of this workshop will be able to participate in a very unique opportunity: they will be able to contribute as both annotators and creators to AR2IL's effort to generate a novel 2D floor plan dataset, called Common House, as well as a novel 3D mesh house dataset, called Model Mine. The students will be credited as contributors to the dataset, and have opportunities to continue as annotators on the dataset in the months after the workshop.
Overview and Syllabus:
Day 1 June Saturday 26th:
Morning:
Introduction to Dataset Creation in architecture (including an overview of AR2IL projects) — 1hr Introduction to Deep Neural Networks and why we need huge datasets — 1hr
High level overview of data-driven feature learning
Data centric vs. model-centric approach to training DNNs — ImageNet example — ImageNet doesn't have enough architectural design semantics to be applied to architecture — List examples of architecture of datasets
Demonstrate that higher quality data is better than improving model quality for performance
The basic anatomy of a dataset
Discuss how the dataset structure (and labels) vary with tasks
2D labeling, 3D labeling
Labeling Noise/Annotator Bias (what problems can arise in dataset taxonomy, labeling, data collection)
Introduction to Common house Dataset, motivation & problem space — 30 min to 1 hr
Afternoon:
Hands-On/Tutorial: Introduction to Floor Plan Annotation — 2 hrs
Day 2 June Sunday 27th
Morning:
Lecture on AI and Creativity, cultural implications — 30min to 1 hr
Q&A/ Discussion — 30min to 1 hr
Hands-On/Tutorial: Continuing Floor Plan Annotation — 2hrs
Afternoon:
Lecture on Dataset Management and Control/Versioning — 1hr
Hands-On/Tutorial: Continuing Floor Plan Annotation — 2hrs
Day 3 June Monday 28th
Morning:
Lecture on the Model Mine Dataset, problem space and motivation – 1hr
Hands-On/Tutorial: How to model objects for a dataset –2 hr
Afternoon:
Q&A with workshop leaders — 30 min to 1 hr
Hands-On/Tutorial: Model Mine Annotation tool and practice (students have the option to do 2D annotation) — 2 hrs
Day 4 June Tuesday 29th
Morning:
Conversation with cultural theorist about AI in Art and Bias in datasets— 30min
Q and A/ Discussion — 30min
Hands-On/Tutorial: continuing modeling for model mine dataset — 2hr
Afternoon:
Hands-On/Tutorial: continuing model mine or floor plan annotation — 2hr
Day 5 June Wednesday 30th
Morning:
Lecture on model evaluation and validation data — 1 hr
Hands-On/Tutorial: continuing modeling for model mine dataset or annotations — 2hr
Afternoon:
Hands-On/Tutorial: Evaluating trained models (cycleGAN trained on labels2floorplans) — 2 hrs
Parting Q and A — 1 hr
Schedule:
Jun 26 - Jun 30
-
Day 1 / Jun 26
-
Day 2 / Jun 27
-
Day 3 / Jun 28
-
Day 4 / Jun 29
-
Day 5 / Jun 30
Instructors:
-
Matias del Campo Taubman College for Architecture and Urban Planning, University of Michigan,Associate ProfessorDr. Matias del Campo is a registered architect, designer, and educator. He is Associate Professor at Taubman College for Architecture and Urban Planning, University of Michigan, and director of the AR2IL – The Architecture and Artificial Intelligence Laboratory. Here he conducts research on advanced design methods in architecture, primarily through the application of Artificial Intelligence techniques in collaboration with Michigan Robotics and the Computer Science department. Co-founded by Matias del Campo and Sandra Manninger, the architecture practice SPAN is a globally acting practice best known for its application of contemporary technologies in architectural production. Their award-winning architectural designs are informed by advanced geometry, computational methodologies, and philosophical inquiry. SPAN gained wide recognition for its winning competition entry for the Austrian Pavilion at the 2010 Shanghai World Expo, as well as the new Brancusi Museum in Paris. Most recently Matias del Campo was awarded the Accelerate@CERN fellowship, the AIA Studio Prize, and was elected into the boards of directors of ACADIA and IJAC the International Journal of Architectural Computing. SPAN’s work is in the permanent collection of the FRAC, the MAK in Vienna, the Benetton Collection, the Albertina, and several private collections. -
Sandra Manninger ARILab,CoFDr. Sandra Manninger Sandra Manninger is an architect, researcher, and educator. Born and educated in Austria, she co-founded SPAN Architecture together with Matias del Campo in 2003. SPAN Architecture’s research highlights how to go beyond beautiful data to discover something that could be defined voluptuous data. This coagulation of numbers, algorithms, procedures, and programs uses the forces of thriving nature and, passing through the calculation of multi-core processors, knits them with human desire. Sandra Manninger has taught internationally and her award-winning projects have been published and exhibited, for example, at La Biennale di Venezia 14/16/18/21, the MAK, the Autodesk Pier 1 and have been included in the permanent collections of the FRAC Centre-Val de Loire, The Design Museum/Die Neue Sammlung in Munich, or the Albertina in Vienna. -
Danish Syed University of Michigan,Research AssistantDanish is a Computer Vision Researcher at the Architecture and Artificial Intelligence Lab ( AR2IL ) working with Matias del Campo and Justin Johnson. His research interests lie in developing agents that can learn to understand the underlying representation of our 3D world and power real-world applications for steerable 3D scene generation. He received his Master's degree in Electrical & Computer Science from the University of Michigan, Ann Arbor. Previously, he worked at ZS Associates, where he applied the concepts of Computer Vision and NLP to innovate solutions across Healthcare, Gaming and Software industry. -
Alexa Carlson University of Michigan,PhD CandidateAlexandra Carlson is a PhD Student in the University of Michigan Robotics Institute. Her graduate research focuses on robust computer vision for autonomous vehicles, specifically on intelligent data generation and feature modeling in images to improve detection and segmentation algorithm performance. She has served as a graduate student mentor for the Taubman AI+Architecture Master’s thesis studio for the past several years, and has collaborated on numerous architecture projects that involve modeling style in both images and 3D models. She has a BA in psychology from the University of Chicago, where she performed research in both computational neuroscience and physics. -
Janpreet Singh University of Michigan,Research AssistantCurrently pursuing Masters in Electrical and computer engineering department at University of Michigan, Ann Arbor. My research interests are in the field of Computer Vision.