Leadership in AI Research: 9 papers accepted to NeurIPS 2024

Cisco Research and Outshift by Cisco have scored big with nine papers, including one workshop paper, accepted at NeurIPS 2024, the Neural Information Processing Systems conference.  

Our strong presence at NeurIPS demonstrates our commitment to pushing the boundaries of technology and contributing valuable knowledge to the global AI/ML community. These contributions are driving innovation, setting new standards, and inspiring future research in the broader AI space. 

Yearly, groundbreaking ideas in AI come to light at NeurIPS, such as the technology behind deep learning that has revolutionized image recognition and the concept of GANs that’s powering today’s generative AI (GenAI). Simply put, NeurIPS is like the World Cup of AI/ML research.  

Vijoy Pandey, Senior Vice President at Outshift, recounts how the research presented at the conference left a lasting impression on him. “Ever since I walked through Jeff Dean’s seminal presentation on “ML for Systema and Systems for ML” at NIPS 2017 (it was NIPS before it became NeurIPS), I have been fascinated by how many ground-breaking research papers are presented at this conference.” 

The 38th annual conference will gather top researchers, practitioners, and industry leaders from around the world in Vancouver, Canada, from December 10-15 in Vancouver, Canada.    

Outshift and Cisco Research provide leaders with research insights to solve challenges in AI/ML  

The nine papers accepted from the Outshift and Cisco Research teams cover several domains, such as AI privacy, AI security, data compliance in large language models (LLMs), AI problem solving, securing AI-generated content, and AI in cloud infrastructure.  

Each of these topics are at the forefront of current challenges and opportunities facing enterprises, governments, and industries. The papers summarized below provide leaders with frameworks and solutions to some of the most pressing issues in the AI space.  

AI privacy and security

UnlearnCanvas: Redefining Privacy in AI 

Authors: Gaowen Liu, Ramana Rao Kompella, et al. 

Summary: In this paper, Head of Cisco Research and Cisco Fellow, Ramana Rao Kompella, and Research Engineering Technical Leader, Gaowen Liu, introduced UnlearnCanvas, a new dataset and evaluation framework for "machine unlearning" in AI image generation models.  

It aims to develop better ways to make AI forget specific data or styles, crucial for privacy and data rights. Kompella and Liu plan to integrate their contributions to this paper into ModelSmith, an open-source toolbox for machine learning model optimization. 

From Trojan Horses to Castle Walls: Revolutionizing AI Security 

Authors: Gaowen Liu, Ramana Rao Kompella, et al. 

Summary: This explores how manipulated data “Trojan Horses,” can affect AI image generation models. The ‘Castle Walls’ insights highlighted the defensive potential of diffusion models (DMs) when used in data poisoning detection and robust image classification against attacks. Furthermore, Liu and Kompella unveiled a connection between data poisoning and data replication. The paper raises awareness for businesses relying on diffusion models for AI-powered applications, to be cautious about data security risks. 

Data compliance and language models

Reversing the Forget-Retain Objectives: An Efficient LLM Unlearning Framework from Logit Difference  

Authors:  Gaowen Liu, Ramana Rao Kompella, et al. 

Summary: This paper presents a new method for making LLMs, such as ChatGPT, "unlearn" specific information more efficiently, essential for maintaining user privacy and complying with data regulations in AI systems. This framework proposed by Kompella and Liu has a broad impact on improving privacy and data leakage issues in LLM usage, making LLMs safer and more reliable in practical application.  

AI problem-solving capabilities

MACM: Utilizing a Multi-Agent System for Condition Mining in Solving Complex Mathematical Problems  

Authors: Ali Payani, et al. 

Summary: MACM (Multi Agent system for Condition Mining) introduces a new approach to solving complex math problems using AI. By employing multiple AI "agents" working together, it achieves state-of-the-art performance in mathematical reasoning. This has potential applications in fields requiring advanced problem-solving capabilities. Currently, the MACM algorithm holds a top spot on the leaderboard for the complex Math dataset. 

“Working at Cisco and Outshift has been instrumental in providing the resources and collaborative environment necessary to drive this innovation forward,” says Ali Payani, Senior Researcher at Cisco Research. 

Securing AI-generated content

RAW: A Robust and Agile Plug-and-Play Watermark Framework for AI-Generated Images with Provable Guarantees  

Authors: Jayanth Srinivasa, Ashish Kundu, et al. 

Summary: RAW is a new method for quickly and securely adding invisible watermarks to AI-generated images. This is important for protecting intellectual property and verifying the source of AI-created content, which is becoming increasingly important in the age of widespread AI image generation.  

The RAW framework that Cisco Research’s Ashish Kundu, Head of Cybersecurity Research, and Jayanth Srinivasa, Senior Researcher, helped create provides a low-latency and robust way of watermarking AI-generated images. 

AI for infrastructure optimization

IaC-Eval: A Code Generation Benchmark for Infrastructure-as-Code Programs  

Authors: Myungjin Lee, et al. 

Summary: This paper introduces a new benchmark for testing AI models' ability to generate Infrastructure-as-Code (IaC) programs from plain language instructions. It's designed to push the boundaries of AI in cloud computing and DevOps, addressing a crucial gap in automated infrastructure management. By providing a standardized way to evaluate AI performance in this domain, the research aims to accelerate the development of more reliable and efficient tools for cloud infrastructure automation. 

“As a whole, the community can leverage our dataset to accelerate their IaC-related research,” says Myungjin Lee, Senior Researcher at Cisco Research. 

A Structure-Aware Framework for Learning Device Placements on Computation Graphs  

Authors: Nesreen Ahmed, et al. 

Summary: This paper addresses the optimization of neural networks, represented as computation graphs (Directed Acyclic Graphs or DAGs) by distributing them efficiently across multiple devices with varying capabilities, a process known as the device placement problem. The proposed approach learns from smaller computation graphs to make more effective decisions about device placement, which is important for enhancing the performance of neural networks on hardware such as CPUs, GPUs, and specialized accelerators.  

“This work has the potential to make neural networks run much faster, which could be important in fields like AI research, where time and efficiency are critical,” says Nesreen Ahmed, Principal Scientist at Outshift. 

High-performance computing and code optimization

CodeRosetta: Pushing the Boundaries of Unsupervised Code Translation for Parallel Programming 

Authors: Nesreen Ahmed, et al. 

Summary: The core problem the paper addresses is the challenge of translating between general-purpose programming languages, like C++, and their high-performance computing (HPC) extensions, such as CUDA and Fortran. CodeRosetta, a new model specifically designed to handle complex translations, making it easier to convert between languages like C++ and its HPC extensions (CUDA, Fortran). With more accurate translations for HPC code, the performance of AI/ML models that depend on high-performance computing can be significantly enhanced, accelerating the development of more complex AI systems or models. 

Workshop paper, OMPar: Automatic Parallelization with AI-Driven Source-to-Source Compilation 

Authors: Nesreen Ahmed, et al. 

Summary: Most computers today have multiple cores, allowing them to perform many tasks simultaneously. However, making software that takes full advantage of all these cores (i.e., parallelizing the code) is still difficult and often requires manual work by experts, especially when dealing with complex code. This process is error-prone and time-consuming. The paper introduces a tool called OMPar, which uses Large Language Models (LLMs) to automate this parallelization process for C/C++ code. 

“OmPar automates a traditionally manual and error-prone task, making it easier to optimize code for modern hardware without needing expert-level knowledge,” Ahmed says. “This could lead to more efficient, scalable computing systems as the tool helps developers produce faster code with less effort.” 

Join Outshift at NeurIPS 2024 

If you’re attending NeurIPS 2024, keep an eye out for the groundbreaking research conducted by the Outshift and Cisco Research teams. Stay up to date with the latest findings from Cisco Research.