ICLR 2023 has announced the four recipients of the ICLR 2023 Outstanding Paper Awards.The award is given to the paper that is judged to be of high quality, originality, and significance, and have the potential to make a significant contribution to the field of machine learning.
The International Conference on Learning Representations (ICLR) is an annual conference that brings together researchers and professionals from the fields of machine learning and artificial intelligence to discuss and share the latest advances and research in the area of machine learning and deep learning.
ICLR 2023 takes place from May 1 to May 5, 2023 in Kigali, Rwanda and offers various activities, including paper presentations, workshops, tiny papers, and social events.
The Outstanding Paper Award is highly competitive, and only a few papers are selected each year. As for the 2023 Outstanding Paper Award, the following four papers have been selected:
1. “Universal Few-shot Learning of Dense Prediction Tasks with Visual Token Matching”
from School of Computing, KAIST & Microsoft Research Asia
The paper describes a method for learning dense prediction tasks (pixel-wise predictions for an input image) like semantic segmentation (label each pixel in an image with a semantic class label, such as “person”, “car”, “building”), depth estimation (predicting the depth or distance of each pixel in an image), edge detection (the boundaries of objects in an image), and keypoint detection (detecting and localizing specific points or landmarks in an image) using a few labeled examples.
The proposed approach is seen in picture below.
VTM can quickly learn to solve new image-related tasks with very little labeled data. To do this, it uses the image encoder and label encoder to extract important features from small patches of the images and their corresponding labels.
Then, it compares these patch features to the label features to match parts of the image with the corresponding labels. Finally, it combines these parts to create a full prediction of the label for the whole image.
What’s really outstanding about this method is that it only needs a tiny amount of labeled data (less than 0.004% of what other methods might need) to get started, and it can learn new tasks quickly and easily with just a small amount of additional labeled data (around 0.1% of what other methods might need). This makes it very useful for situations where labeled data is scarce or expensive to obtain.
2. “Rethinking the Expressive Power of GNNs via Graph Biconnectivity”
from National Key Laboratory of General Artificial Intelligence, School of Intelligence Science and Technology, Peking University and Center for Data Science, Peking University
This research investigates the ability of the Graph Neural Networks (GNNs) to represent complex relationships between nodes in a graph and to solve graph biconnectivity problem (the existence of two independent paths between any two nodes in the graph).
The authors discovered that only one type of GNN is capable of effectively representing these relationships, and they propose a new algorithm called GD-WL that utilizes the distance between nodes.
They also created a new powerful and efficient architecture called Graphormer-GD. Their new algorithm was tested on both synthetic and real-world data, with amazing results.
3. “DreamFusion: Text-to-3D using 2D Diffusion”
from Google Research & UC Berkeley
The paper proposes a novel technique for generating 3D models from text descriptions, without requiring any pre-existing 3D models for training. It uses a diffusion model (DreamFusion) previously trained for generating images from text.
Given a caption, DreamFusion generates reliable 3D objects with high-fidelity appearance, depth, and normals.
Video source: GitHub
4. “Emergence of Maps in the Memories of Blind Navigation Agents”
from Georgia Institute of Technology, FAIR, Meta AI, Simon Fraser University, Google Research Atlanta, and Oregon State University
The article introduces innovative interdisciplinary research that draws from cognitive science and machine learning to explore the representations acquired by artificial “blind” navigating agents and their potential for enhancing navigation.
Initially, the agent that exhibits wall-following behavior, navigates (blue path, a blue Long Short-Term Memory LSTM) from a starting point denoted by a green sphere (S) to a target location marked by a red sphere (T). Subsequently, a probe with collected spatial information, represented by a purple LSTM, is assigned to repeat the same navigation episode.
By taking shortcuts, the probe network navigates more efficiently, which is demonstrated by the shorter purple path.
We hope you found these papers as interesting as we did!
Learn more:
- Story source: “Announcing the ICLR 2023 Outstanding Paper Award Recipients” (on ICLR)
- Paper 1: “Universal Few-shot Learning of Dense Prediction Tasks with Visual Token Matching” (on arXiv)
- Paper 2: “Rethinking the Expressive Power of GNNs via Graph Biconnectivity” (on arXiv)
- Paper 3: “DreamFusion: Text-to-3D using 2D Diffusion” (on arXiv)
- Paper 4: “Emergence of Maps in the Memories of Blind Navigation Agents” (on arXiv)
- GitHub repository: “DreamFusion: Text-to-3D using 2D Diffusion” (paper, project, gallery)
