Sony AI at ICLR 2024: Pioneering the Future of AI with Cutting-Edge Research

Welcome to our latest research update where we dive into our work that has been accepted at the International Conference on Learning Representations (ICLR) 2024. ICLR stands at the forefront of artificial intelligence research, fostering an environment where the latest developments in deep learning and machine learning are shared and celebrated from some of the brightest researchers from around the world. Sony AI has continuously shared work at this conference, and this year is no exception.

This year, ICLR 2024 has shown impressive statistics in terms of paper submissions and acceptances.

• The conference saw a total of 7,404 submissions.
• Out of these, 2,260 papers were accepted, which equates to an acceptance rate of approximately 30.52%. (Sony AI had 8 papers accepted.)
• The specific numbers for each type of presentation were 1,807 for poster presentations (24.41% of total submissions), 367 for spotlight presentations (4.96%), and 86 for oral presentations (1.16%).
• The rejection rate stood at 46.95%, with 3,476 papers not making it through the selection process.

This year, in the vibrant city of Vienna, Austria, Sony AI has emerged as a leader in pioneering research, contributing an impressive lineup of papers and workshops that push the boundaries of what's possible in AI.

Sony AI: ICLR 2024 Spotlights

At ICLR 2024, Sony AI is proud to announce its substantial contribution with ten accepted papers and active participation in key workshops.

Our involvement underscores a commitment to advancing the knowledge of AI with research that spans preserving guided diffusion models, the intricacies of federated learning, and the crucial aspects of privacy in AI.

Here's a closer look at some of the groundbreaking work we're presenting:

Music Foundation Model Team

1. Consistency Trajectory Models: This paper introduces a novel approach to understanding probability flow in diffusion models, marking a significant step forward in generative model research.
   
2. Manifold Preserving Guided Diffusion: A testament to our team's innovation, this research showcases state-of-the-art (SOTA) achievements, enhancing the robustness and quality of generated content.
   
3. SAN: Inducing Metrizability of GAN with Discriminative Normalized Linear Layer: By redefining the structural foundations of GANs, this paper opens new avenues for generative models, ensuring greater stability and performan
4. Understanding multimodal contrastive learning through pointwise mutual information: This workshop paper presents the importance of integrating different modalities, such as text, vision and audio for real-time applications.
   
   

Privacy-Preserving Machine Learning (PPML) Team—part of our Imaging and Sensing Flagship

1. Detecting, Explaining, and Mitigating Memorization in Diffusion Models: Selected for an oral presentation, this work addresses the critical issue of data memorization in AI models, proposing novel solutions for enhancing data privacy.
   
2. FedWon: Triumphing Multi-domain Federated Learning Without Normalization: Explores advancements in federated learning, paving the way for more inclusive and diversified AI training environments.
3. DIAGNOSIS: Detecting Unauthorized Data Usages in Text-to-image Diffusion Models. This work proposes a new method for detecting unauthorized data usage by planting the injected memorization into the text-to-image diffusion models trained on the protected dataset.
4. FedP3: Federated Personalized and Privacy-Friendly Network Pruning under Model Heterogeneity: This paper focuses on making federated learning more accessible and efficient, demonstrating Sony AI's leadership in privacy-aware AI development.

New Sony AI Team Members With Work at ICLR 2024

Before joining Sony AI, Researchers Akanksha Saran, and Swami Sankaranarayanan, each authored two unique separate pieces of work while post-doc at Microsoft and MIT, respectively. Both authors were accepted in the Top 5% Spotlight session.

• Views Can Be Deceiving: Improved SSL Through Feature Space Augmentation: This paper delves into the intricacies of Self-Supervised Learning (SSL) by demonstrating how common augmentation techniques can unintentionally emphasize spurious features. Swami Sankaranarayanan and collaborators propose a novel method, “Late-layer Transformation-based View Generation” (LATE-TVG), which mitigates these effects by pruning later layers of the encoder, enhancing representation quality for more reliable and unbiased AI decision-making processes​​.
  
  
• Towards Principled Representation Learning from Videos for Reinforcement Learning: Akanksha Saran and collaborators present a principled study of video-based representation learning methods for reliable reinforcement learning on downstream tasks. They showcase theoretical analyses for the success and failure cases of several pre-existing representation learning methods, which are validated empirically in three visual domains. Their work builds a rigorous understanding of the building blocks of foundation models for decision-making.
  

Workshops Led by the AI Ethics Team

At ICLR 2024, we're not just presenting papers; we're also fostering discussions that shape the future of AI.

ICLR 2024 Workshop on Data-centric Machine Learning Research (DMLR): Spearheaded by Jerone Andrews from our AI Ethics team, this workshop delves into the intricacies of data-centric AI, emphasizing the importance of high-quality training data, robust data management, and ethical considerations in AI projects. It aims to highlight AI’s intersection with global sustainable development goals, fostering a dialogue on AI for good.

Navigating and Addressing Data Problems for Foundation Models (DPFM): This workshop focuses on the data-centric challenges and opportunities in the development and application of Foundation Models. It aims to explore how curated training data can address critical issues such as ethics, privacy, and security, offering a platform for collaboration and innovation.

Join Us in Vienna

ICLR 2024 is not only a conference — more importantly, it's a gathering of minds eager to explore the uncharted territories of AI. We are so grateful to share our research that not only advances technology but also addresses the ethical and practical implications of AI in our world. Stay tuned for more updates, including blog posts detailing our research, social media highlights, and exclusive content from the conference floor.