August 2024

This month’s newsletter features insights into Amazon's research at CAV 2024 and ACL 2024, advances like cross-region inference on Amazon Bedrock, insights into the future of generative AI with AWS CEO Matt Garman, updates on academic collaborations, and more.

Deep dives

How the Lean language brings math to coding and coding to math: Lean, an open-source functional programming language and interactive theorem prover, was spearheaded by Amazon senior applied scientist Leonardo de Moura. In his CAV 2024 keynote, de Moura explores Lean's diverse applications, including formal mathematics, software and hardware verification, and AI for math and code synthesis and its role in math and computer science education.

Formal verification makes RSA faster — and faster to deploy: Amazon's Automated Reasoning group has boosted RSA signature throughput on Amazon's Graviton2 chip by 33% to 94%, depending on the key size. They achieved this while ensuring the functional correctness of optimizations through formal verification processes.

A quick guide to Amazon’s papers at ACL 2024: Amazon researchers presented over 30 papers at the conference, covering a wide array of LLM-related topics like code synthesis and automatic speech recognition, as well as crucial issues in LLM training and deployment such as continual pretraining and hallucination mitigation. Two papers received awards: “MultiPICo”, a multilingual perspectivist corpus of ironic short conversations in various languages and linguistic varieties, co-developed with l'Università degli Studi di Torino, received the Outstanding Paper Award; and “CaMML", a novel methodology employing a context-aware multimodal learner for fine-tuning large multimodal models, received the Area Chair Award.

News and updates 

Empowering everyone with GenAI to rapidly build, customize, and deploy apps securely: At the AWS New York Summit, Amazon unveiled a suite of capabilities to customize generative-AI applications swiftly and securely. These include the ability to create generative-AI applications with Amazon Q Apps using simple natural-language prompts, streamline deployment with new features in Amazon Bedrock, and collaborate with partners like Scale AI for efficient customization of applications.

AWS CEO Matt Garman on the future of generative AI in the enterprise: In a conversation with Silicon Angle, Matt Garman shares insights into the cloud's trajectory, the evolution of generative AI, and why AWS views inference as a significant future investment.

Cross-region inference is now available for customers on Amazon Bedrock: The latest feature enables dynamic traffic routing across multiple regions, ensuring optimal availability and smoother performance during peak usage. With this update, developers can scale generative AI workloads by tapping into capacity from various AWS regions, streamlining complex resiliency structures with pre-configured AWS region sets.

Academic collaborations

Max Planck–Amazon Science Hub 2024 Research Symposium: On September 16-17, the Max Planck and Amazon Science Hub will co-host a fall research symposium, uniting research communities to showcase the latest innovations and advances. The two-day event will feature keynote and lightning talks, breakout discussions, and networking opportunities.

UCLA and Amazon research team uncover new understanding of how LLMs reason: A collaboration between UCLA and Amazon researchers has shed light on how LLMs approach deductive and inductive reasoning, offering valuable insights for future generative-AI applications that necessitate advanced reasoning capabilities.

Amazon and UCLA’s Science Hub for Humanity and Artificial Intelligence select 2024 fellows: The fellowship cohort for this year will receive funding for independent research projects and have the chance to intern at Amazon. Meet the 14 fellows and explore their research initiatives alongside their UCLA faculty advisors.

Upcoming conferences

• ECCV 2024, September 29 - October 4

New publications

• A flexible forecasting stack

• A passive back posture monitoring system using orthogonal RFID design

• AffordanceLLM: Grounding affordance from vision language models

• AlphaRouter: Quantum circuit routing with reinforcement learning and tree search

• Boosting entity recognition by leveraging cross-task domain models for weak supervision

• Building natural language interface for product search

• Controlled automatic task-specific synthetic data generation for hallucination detection

• Cooperative multi-agent deep reinforcement learning in content ranking optimization

• DetoxBench: Benchmarking large language models for multitask fraud & abuse detection

• Diffusion Soup: Model merging for text-to-image diffusion models

• DiNADO: Norm-disentangled neurally-decomposed oracles for controlling language models

• DOSSIER: Fact checking in electronic health records while preserving patient privacy

• Empowering shoppers with event-focused search

• Energy estimation of last mile electric vehicle routes

• FANTAstic SEquences and where to find them: Faithful and efficient API call generation through state-tracked constrained decoding and reranking

• FlexEControl: Flexible and efficient multimodal control for text-to-image generation

• Formal verification of cryptographic software at AWS - Current practices and future trends

• Frustrated random walks: A fast method to compute node distances on hypergraphs

• Graph chain-of-thought: Augmenting large language models by reasoning on graphs

• Improving multilingual ASR robustness to errors in language input

• Inductive or deductive? Rethinking the fundamental reasoning abilities of LLMs

• Membrane – Safe and performant data access controls in Apache Spark in the presence of imperative code

• MERLIN: Multimodal & multilingual embedding for recommendations at large-scale via item associations

• MERLIN: Multiple enhanced representations with LLM generated indices

• Meta knowledge for retrieval augmented large language models

• On early detection of hallucinations in factual question answering

• Pattern template manifest for live video streaming

• Planes, trains and automobiles: Leverage multimodal in-mission signals for shopping journeys

• REAPER: Reasoning based retrieval planning for complex RAG systems

• Reconciling methodological paradigms: Employing large language models as novice qualitative research assistants in talent management research

• Robustness preserving fine-tuning using neuron importance

• SERP interference network and its applications in search advertising

• Sub-Resolution mmWave FMCW radar-based touch localization using deep learning

• Text-guided video masked autoencoder

• User-assisted code query customization and optimization

• VERA: Validation and evaluation of retrieval-augmented systems

• Very low complexity speech synthesis using framewise autoregressive GAN (FARGAN) with pitch prediction

• ViGoR: Improving visual grounding of large vision language models with fine-grained reward modeling

• Zero-shot controllable image-to-video animation via motion decomposition

LinkedIn | X/Twitter | Facebook | Instagram | GitHub | RSS

© 1996-2024 Amazon.com, Inc. or its affiliates | Privacy | Conditions of Use