• AI Research Insights
  • Posts
  • 🔥 What is Trending in AI Research?: AutoGen + DYnet++ + Why Don’t Language Models Understand ‘A is B’ Equals ‘B is A’?.....

🔥 What is Trending in AI Research?: AutoGen + DYnet++ + Why Don’t Language Models Understand ‘A is B’ Equals ‘B is A’?.....

This newsletter brings AI research news that is much more technical than most resources but still digestible and applicable

Hey Folks!

This newsletter will discuss some cool AI research papers and AI tools. Happy learning!

👉 What is Trending in AI/ML Research? 

How can developers simplify and optimize workflows when leveraging large language models (LLMs), given their growing complexity? Addressing this challenge, this paper from Microsoft introduces AutoGen, a framework designed to streamline the orchestration, optimization, and automation of LLM-based workflows. AutoGen features customizable conversational agents that tap into the capabilities of advanced LLMs such as GPT-4. Notably, these agents can also counterbalance the limitations of LLMs by interacting with humans, tools, and even other agents through automated chats, ensuring a more seamless and effective workflow management.

How can one measure complex surfaces with low reflectivity using single-shot deflectometry when obtaining the phase from a poor-quality single complex pattern is challenging? Addressing this problem, this paper introduces a novel approach leveraging deep learning for single-shot deflectometry to measure low-reflectivity complex surfaces. To train the deep learning model, a deformable mirror with nine actuators was designed to produce extensive data for different surface shapes. The researchers developed a model called DYnet++ to extract the phase from single composite patterns, inclusive of both closed and opened loops. The effectiveness of this deep learning-based method was confirmed by contrasting its results with the 16-step phase-shifting technique.

How do large language models (LLMs) fare when it comes to generalizing from one statement to its logical reverse? This study unveils the "Reversal Curse" in auto-regressive LLMs: a model trained on the statement "A is B" struggles to deduce "B is A". For example, training on "Olaf Scholz was the ninth Chancellor of Germany" doesn't aid the model in answering "Who was the ninth Chancellor of Germany?". Even with fictitious data, models like GPT-3 and Llama-1 do not generalize this prevalent pattern. This phenomenon persists across different model sizes and families. Notably, GPT-4 performs well when asked about real-world celebrities in one direction, but its performance drops significantly when the question is reversed, suggesting a fundamental flaw in logical deduction.

How can robots be trained to acquire diverse and generalizable skills, especially beyond simple tasks that solely rely on visual guidance? Addressing this, this paper introduces an approach to harness multi-modal perception for robotic manipulation. The authors have compiled an expansive dataset containing over 110,000 contact-rich robot manipulation sequences, representing a broad array of skills and scenarios. Collected in real-world settings, this dataset uniquely offers visual, force, audio, and action data for each sequence, enriched with a corresponding human demonstration video and a descriptive language annotation. Their method emphasizes the potential of robots to master hundreds of complex skills using both visual and tactile information.

How can the efficiency and effectiveness of vision transformers (ViTs) be improved while addressing their challenges of computational intensity and lack of inductive biases? To address this, this paper introduces a novel vision transformer model, DualToken-ViT. This model fuses tokens from both local, convolution-based structures and global, self-attention-based mechanisms, thereby creating an efficient attention framework. The unique feature of this model is its use of position-aware global tokens across all stages, enriching global context and enhancing position information within images. When tested on the ImageNet-1K dataset, DualToken-ViT showed remarkable performance, achieving accuracy rates of up to 79.4% with reduced computational overhead, even outshining models like LightViT-T.

🚀 Here is another free AI webinar on 'How to Use Kafka & Vectors for Real-Time Anomaly Detection' with a live demo. [Register Now]

👉 What is Trending in AI Tools? 

  • CSM: Converts images and text into immersive 3D assets and game scripts.

  • Pickaxe: Pickaxe is a no-code platform that lets you create and embed GPT-4 apps on your website in minutes.

  • Adcreative AI: Boost your advertising and social media game with AdCreative.ai - the ultimate Artificial Intelligence solution. [Marketing and Sales]

  • Notion: A feature-rich note-taking and project management tool that serves as an all-in-one workspace for teams and individuals alike. [Project Management]

  • Decktopus: The ultimate online presentation tool that harnesses the power of AI to help you craft captivating presentations effortlessly. [Presentation]

  • Aragon: Get stunning professional headshots effortlessly with Aragon. [Profile]