Alexandre Kirchmeyer

Hello! I am a 1st year PhD student at Princeton, advised by Prof. Jia Deng. I am interested in topics related to reasoning and lifelong learning - my research goal is to come up with a learning algorithm that can learn about the world and grow like a human baby would.

I was a research intern at Cartesia, where I worked on training multimodal foundation models using state-space models.

I have a Master’s degree in Machine Learning from Carnegie Mellon University. At CMU, I did research with Prof. Deepak Pathak on learning controllable representations from robotic simulators and 2D generative models.

Before CMU, I completed the Ingénieur Polytechnicien BEng/MS degree at Ecole Polytechnique, majoring in Mathematics and Computer Science.

I love to challenge myself - I was reserve member in the team that ranked 4th in the ACM ICPC Europe Regional competition (SWERC) in 2020/2021, built an electric rocket with 5 other people that I presented to the International Astronautical Congress in 2021, interned at a cybersecurity startup to work on operating system internals, and spent 1-2 years writing my own 3D RTS video game in C++ and OpenGL in middle school.

News

Oct 4, 2023	Presented Convolutional Networks with Oriented 1D Kernels poster at ICCV 2023!
Sep 28, 2023	Released code for Convolutional Networks with Oriented 1D Kernels.
Jul 13, 2023	Our work Convolutional Networks with Oriented 1D Kernels was accepted at ICCV 2023!

Research

Convolutional Networks with Oriented 1D Kernels

Kirchmeyer A., and Deng J.

IEEE/CVF International Conference on Computer Vision (ICCV), 2023

Abs PDF Code Poster

In computer vision, 2D convolution is arguably the most important operation performed by a ConvNet. Unsurprisingly, it has been the focus of intense software and hardware optimization and enjoys highly efficient implementations. In this work, we ask an intriguing question: can we make a ConvNet work without 2D convolutions? Surprisingly, we find that the answer is yes – we show that a ConvNet consisting entirely of 1D convolutions can do just as well as 2D on ImageNet classification. Specifically, we find that one key ingredient to a high-performing 1D ConvNet is oriented 1D kernels: 1D kernels that are oriented not just horizontally or vertically, but also at other angles. Our experiments show that oriented 1D convolutions can not only replace 2D convolutions but also augment existing architectures with large kernels, leading to improved accuracy with minimal FLOPs increase. A key contribution of this work is a highly-optimized custom CUDA implementation of oriented 1D kernels, specialized to the depthwise convolution setting. Our benchmarks demonstrate that our custom CUDA implementation almost perfectly realizes the theoretical advantage of 1D convolution: it is faster than a native horizontal convolution for any arbitrary angle. Code is available at this https URL: https://github.com/princeton-vl/Oriented1D.
Zero-shot Image to 3D using Diffusion models

Kirchmeyer A., Duggal S., and Pathak D.

Independent Study Spring 2023, 2023

Abs Code

Text-to-3D methods have made significant strides in the past year, by leveraging the power of diffusion models to optimize the representation of 3D shapes conditioned on a text prompt. These methods can be used for 3D asset creation and find applications in the animation and gaming industries. Current state-of-the-art open-source methods have difficulty getting cross-view consistency, leading to inconsistent depths and the so-called multiple Janus head problem. To tackle this challenge, we introduce a diffusion model trained on novel view synthesis and conditioned on previous views. To overcome the lack of 3D data, we design our approach to be category-free by generating sparse novel views of the input image and augmenting these views with inpainting using diffusion models.
Multi-Modal Concept Learning using Auxiliary Learning and Diffusion models

Kirchmeyer A., and al

CMU 11-777 Multi-Modal Machine Learning Fall 2022 Course, 2022

Abs Code Poster

Recent developments, especially in transformer-based models, have led to impressive progress on multimodal tasks such as visual question-answering. However, it is still unclear whether these large pretrained models know the concepts needed to reason about the problem and whether they can combine these concepts to produce predictions instead of simply relying on surface patterns in the data. We study the performance of multimodal models, specifically ViLT, on VQA dataset Visual7W, and implement 3 approaches through the lens of our concept-driven framework. We find that concept information can be useful as auxiliary information via multitask learning, their latent representations can be useful in uncovering latent concepts, and causal generative augmentation of images can increase the robustness of ViLT towards conceptual variations.