This paper discussed the problem of image classification on a subset of the ImageNet data set, the LSRVC-2010. It introduces AlexNet, a deep Convolutional Neural Network architecture, which won the 2012 ImageNet competition, significantly advancing deep learning.
This paper discusses about one of the components of personalized gaming, dynamic difficulty adjustment (DDA). While existing DDA systems have a more greedy/ short term solution, this paper discusses the long term goal of increasing player engagement.
This paper presents a method called DreamFusion, for generating 3D models from text descriptions by leveraging pre-trained 2D diffusion models. Models like ImageGen can generate 2D images based on text descriptions. DreamFusion adapts these models to create 3D objects.
This paper uses a diffusion-based generative model to present BrepGen, a method for directly generating 3D CAD models in the Boundary Representation (B-rep) format. The key to BrepGen is using a hierarchical tree structure to represent the Brep geometry and topology. The root of the tree is the entire solid structure and the child nodes are the components such as its edges, vertices, and faces. Thus the geometry is maintained in the nodes of the tree and the topology is encoded implicitly through node duplication. BrepGen opens up the possibility of generating more complex and realistic shapes, as it can generate free-form and doubly curved surfaces. This opens up new possibilities for automating and improving the CAD design workflow. You can find the original paper here
This paper introduces a new model by combining deep learning with reinforcement learning, specifically to handle high dimensional inputs like raw images. The researchers developed the Deep Q-networks, which uses convolutional neural networks to process high-dimensional data. This is used to train an agent to play Atari games, without any game specific information.
This paper focuses on the development and performance of the AlphaZero algorithm and how it achieves tabula rasa learning, that is learning from a blank slate or without any domain specific knowledge.
The paper discusses the working and achievements of the AlphaGo model, trained specifically for the complex game of Go. AlphaGo was able to achieve superhuman levels of playing Go and was able to beat world champions.
This paper tackles the challenge of training very deep neural networks, which often suffer from vanishing and exploding gradients, leading to a degradation problem where performance worsens as layers increase. The researchers address this problem by using a residual learning framework. Instead of approximating a function H(x), they approximate F(x) = H(x) - x.