AI Lab itinai.com

  • Reshaping the Model’s Memory without the Need for Retraining

    Large language models (LLMs) have become widely used, but they also pose ethical and legal risks due to the potentially problematic data they have been trained on. Researchers are exploring ways to make LLMs forget specific information or data. One method involves fine-tuning the model with the text to be forgotten, penalizing the model when…

  • Oh, you meant “manage change”?

    This text explores different perspectives on change in a data organization. Alex, the CDO, focuses on driving business value and staying ahead of market shifts, while Jamie, a data engineer, is more concerned with day-to-day challenges and keeping things running smoothly. The article emphasizes the importance of transparency, collaboration, and standardization in managing change effectively.…

  • KAIST Researchers Propose SyncDiffusion: A Plug-and-Play Module that Synchronizes Multiple Diffusions through Gradient Descent from a Perceptual Similarity Loss

    Researchers from KAIST have introduced SYNCDIFFUSION, a module that aims to improve the generation of panoramic images using pretrained diffusion models. The module addresses the problem of visible seams when stitching together multiple images. It synchronizes multiple diffusions using gradient descent based on a perceptual similarity loss. Experimental results show that SYNCDIFFUSION outperforms previous techniques,…

  • Common-Knowledge Effect: A Harmful Bias in Team Decision Making

    Teams often make worse decisions than individuals because they rely too heavily on widely understood data and ignore information possessed by only a few team members. Research has consistently shown that teams spend too much time discussing information they all already know, leading to poor decision-making.

  • The 4 Degrees of Anthropomorphism of Generative AI

    Chatbots and AI are often seen as human-like, with users treating them as companions. This anthropomorphism has a functional role, as users believe AI will perform better, and a connection role, to enhance the user experience. A usability study of ChatGPT identified two new behaviors for managing length and detail: accordion editing and apple picking.

  • Meet ScaleCrafter: Unlocking Ultra-High-Resolution Image Synthesis with Pre-trained Diffusion Models

    Researchers have developed ScaleCrafter, a method that enables the generation of ultra-high-resolution images using pre-trained diffusion models. By dynamically adjusting the convolutional receptive field, ScaleCrafter addresses issues like object repetition and incorrect object topologies. It also introduces innovative strategies like dispersed convolution and noise-damped classifier-free guidance. The method has been successfully applied to a text-to-video…

  • 6 Magic Commands for Jupyter Notebooks in Python Data Science

    Jupyter Notebooks are widely used in Python-based Data Science projects. Several magic commands enhance the notebook experience. These commands include “%%ai” for conversing with machine learning models, “%%latex” for rendering mathematical expressions, “%%sql” for executing SQL queries, “%run” for running external Python files, “%%writefile” for quick file creation, and “%history -n” for retrieving previous commands.…

  • Dimensionality Reduction with Scikit-Learn: PCA Theory and Implementation

    The Curse of Dimensionality refers to the challenges that arise in machine learning when dealing with problems that involve thousands or millions of dimensions. This can lead to skewed interpretations of data and inaccurate predictions. Dimensionality reduction techniques, such as Principal Component Analysis (PCA), can help mitigate these challenges by reducing the number of features…

  • How Meesho built a generalized feed ranker using Amazon SageMaker inference

    Meesho, an ecommerce company in India, has developed a generalized feed ranker (GFR) using AWS machine learning services to personalize product recommendations for users. The GFR considers browsing patterns, interests, and other factors to optimize the user experience. Meesho used Amazon EMR with Apache Spark for model training and SageMaker for model deployment. The implementation…

  • Meta announces the AI-robot training platform Habitat 3.0

    Facebook AI Research (FAIR) introduces Habitat 3.0, a virtual training ground for building AI agents that understand their environment and collaborate with humans. Habitat 3.0 allows robots and virtual humans to complete tasks in a digital environment, providing a safer and faster alternative to real-world training. FAIR also released the Habitat Synthetic Scenes Dataset (HSSD-200)…