NeurIPS 2023 Poster Session 3 (Wednesday Evening)

1. Graph of circuits is used to explore the optimal design space.

🥈85 00:20

The graph of circuits is a representation of a circuit design problem, where each node corresponds to a circuit component and the edges represent the connections between them.

• The graph is used to solve the pre-layout designing problem and optimize the design parameters.
• It helps in minimizing the objective function while satisfying the specifications of the circuit.

2. Graph surrogate model approximates the behavior of a simulator.

🥈88 01:00

The graph surrogate model is trained using a graph neural network (GNN) on labeled circuit instances.

• The GNN predicts labels for unlabeled circuit instances, approximating the behavior of a simulator.
• This approach reduces the computational cost of generating labeled samples for optimization.

3. Constraint optimization is used to find the most optimal design parameters.

🥇92 05:10

Constraint optimization algorithms, such as ESASO and ASO, are applied to the graph surrogate model to find the design parameters that satisfy the specifications and minimize the objective function.

• ESASO is a static optimization framework, while ASO is a dynamic learning framework.
• These algorithms iteratively optimize the design parameters based on the labels predicted by the graph surrogate model.

4. NeurIPS 2023 Poster Session 3 (Wednesday Evening)

🥉78

The content is a transcript of the NeurIPS 2023 Poster Session 3 held on Wednesday evening.

• The session likely covered various topics related to machine learning and artificial intelligence.
• The transcript provides insights into the discussions and research presented during the session.

5. Empowering Collaborative Filtering with Principled Adversarial Contrastive Loss

🥇92 15:20

Principled adversarial contrastive loss can improve generalization ability in recommendation systems by finding a better loss function.

• The loss function should achieve better generalization ability, meaning it performs well even when the testing and training data have different distributions.
• Adversarial training can be used to learn the loss function, which is equivalent to solving a distributional robust problem.

6. Cluster-aware Semi-supervised Learning: Relational Knowledge Distillation Provably Learns Clustering

🥈88 22:11

Relational knowledge distillation (RKD) can be used to learn clustering structure in semi-supervised learning.

• RKD helps the student model learn the clustering structure revealed by the teacher model.
• By leveraging the clustering structure, the student model can achieve good generalization and reduce the number of labeled samples needed.

7. Privacy Assessment on Reconstructed Images: Are Existing Evaluation Metrics Faithful to Human Perception?

🥈82 28:30

Existing evaluation metrics for privacy assessment on reconstructed images may not align with human perception.

• Metrics that suggest similarity between reconstructed and original images may not accurately reflect human perception.
• This can lead to privacy leakage if the wrong candidates are selected based on the metrics.

8. Proposed method for evaluating information leakage in images.

🥈85 29:39

The proposed method involves training a model using a triplate loss and comparing the feature vectors of original and reconstructed images.

• The model is trained using anchor, positive, and negative samples.
• The generated scores align well with human observers' judgments.

9. Importance of using a universal metric for evaluating different datasets and attack methods.

🥇92 31:22

Different metrics perform well for different datasets and attack methods, highlighting the need for a universal metric.

• The proposed method consistently performs well across different datasets and attack methods.
• Changing the training setup or loss function also improves the results.

10. Future work on scaling the information leakage evaluation.

🥉78 32:53

The authors plan to work on scaling the evaluation to a larger scale, from 1 to 10, to capture different levels of information leakage.

• This is an area for further research and improvement.
• The dataset, labels, and codes will be made accessible for others to use.