Praneet Kuber, a seasoned Manager Data Scientist at Michelin India Pvt Ltd, graced the Machine Learning Developers Summit (MLDS) 2024 in Bengaluru, unraveling the possibilities of an “End-to-End Point Cloud-Based Generative Model for Multi-Part Engineering Designs.” With a wealth of experience, Praneet has spent seven years in the field, three of which have been dedicated to the realm of data science. His focus on applied AI research for the engineering design domain, particularly in AI automation and generative modeling, marked him as a key figure at India’s biggest generative AI conference.
Praneet commenced his talk by providing a comprehensive understanding of point cloud data, describing it as a collection of points in space or on a plane, characterized by XYZ coordinates or simply XY coordinates. This data, acquired through methods like photogrammetry and time-of-light sensors, has applications ranging from depth-sensing cameras to self-driving cars.
However, Praneet emphasized the challenges associated with point cloud data, including its voluminous nature, potential noise, and the complexities introduced during the registration of data from multiple sources. He seamlessly bridged this introduction to the realm of deep learning applications in handling point cloud data.
Transitioning to the challenges in point cloud-based deep learning, Praneet shed light on past solutions. Traditional voxelization, the process of converting point cloud data into a 3D grid, had its limitations, leading to high memory consumption and loss of finer details.
Referencing past architectures like PointNet, Praneet highlighted their attempts to extract meaningful features from point cloud data. He discussed the significance of architectures like TNet and Max Pool layer in preserving permutational invariance, ensuring that the sequence of points doesn’t impact the recognition of objects.
The crux of Praneet’s talk lay in the introduction of their end-to-end point cloud-based generative model. Addressing the problem statement they set for themselves, Praneet elucidated on the need for a generative model capable of handling compound shapes in engineering designs. The model aimed not only to generate individual parts but also to maintain relationships between these parts, forming cohesive compound shapes.
Praneet delved into the architecture they developed, which showcased two groundbreaking contributions. First, they leveraged the PointNet architecture for multi-part structures, a novel application not seen in prior works. Second, they introduced the use of structural embedding, eliminating the need for post-processing or optimization steps after generating reconstructions.
The talk also touched upon the attention block added to capture finer details and the experiments with beta annealing, adjusting the weight on the regularization term of the variational autoencoder loss function.
The experimental results, validated through FID scores, demonstrated a progressive decrease, indicating the model’s capability to generate data closer to the original distribution. Praneet concluded that VAs are robust candidates for generating point cloud data and affirmed the successful learning of interpart relations within the proposed network.
In terms of future scope, Praneet hinted at exploring architectures like the diffusion model and incorporating hierarchically feature-extracting architectures like PointNet Plus+. These endeavors showcase a commitment to continual improvement and exploration within the realm of generative AI for engineering designs.
Praneet Kuber’s talk at MLDS 2024 showcased not only the current state of generative AI in engineering design but also hinted at the promising future. His insights into the challenges of traditional methods and the innovative solutions presented by their end-to-end generative model instilled a sense of optimism in the audience. As the AI landscape evolves, Praneet’s work at the intersection of point cloud data and deep learning paves the way for transformative advancements in engineering design.
