エピソード

  • 21: AI Unleashes Doom with Stable Diffusion

    21: AI Unleashes Doom with Stable Diffusion
    2024/10/29
    In the latest AI Paper Club podcast, hosts Rafael Herrera and Sonia Marques are joined by João Costa, Senior Machine Learning Software Engineer at Deeper Insights. Together, they explore the paper “Diffusion Models are Real-Time Game Engines,” produced by researchers at Google. This episode delves into the intriguing evolution of AI as it replicates the iconic game Doom using stable diffusion—an AI model typically associated with image generation.

    The team discusses the paper’s innovative methodology, detailing how stable diffusion models were adapted to generate frame-by-frame gameplay, capturing Doom’s game logic through AI. João unpacks the technical nuances behind the real-time generation of 20 frames per second using powerful TPU processors and explores the research’s practical applications and limitations.

    We also extend a special thank you to the Google DeepMind team for developing this month's paper. If you are interested in reading the paper yourself, please visit this link: https://gamengen.github.io.

    For more information on all things artificial intelligence, machine learning, and engineering for your business, please visit www.deeperinsights.com or reach out to us at thepaperclub@deeperinsights.com.
    続きを読む 一部表示
    26 分
  • 20: Technical Debt and Its Hidden Costs in Machine Learning Development

    20: Technical Debt and Its Hidden Costs in Machine Learning Development
    2024/09/26
    In this episode of the AI Paper Club Podcast, hosts Rafael Herrera and Sonia Marques sit down with senior machine learning engineer Bernardo Ramos from Deeper Insights. Together, they explore the classic 2015 paper "Hidden Technical Debt in Machine Learning Systems". The paper highlights the often-overlooked issue of technical debt in machine learning projects and how it silently accumulates over time, much like financial debt.

    The discussion delves into the nuances of technical debt, particularly how data dependencies differ from code dependencies and why they are harder to detect. The podcast also covers unstable data signals, feedback loops, and the unique challenges faced by large language models (LLMs) in today's data-driven world. Bernardo shares potential mitigation strategies to help manage these technical debts effectively.

    A special thank you to the authors D. Sculley, G. Holt, D. Golovin, and their team for developing this month's paper. If you are interested in reading the paper yourself, please visit this link: https://dl.acm.org/doi/10.5555/2969442.2969519.

    For more information on artificial intelligence, machine learning, and engineering solutions for your business, please visit www.deeperinsights.com or contact us at thepaperclub@deeperinsights.com.
    続きを読む 一部表示
    23 分
  • 19: Unlocking Explainable Machine Learning in Manufacturing

    19: Unlocking Explainable Machine Learning in Manufacturing
    2024/08/22
    This month’s episode of the AI Paper Club Podcast welcomes Dr. Diogo Ribeiro, a senior machine learning engineer at Deeper Insights. Diogo presents a research paper he co-developed, focusing on the industrial application of AI, titled "Isolation Forest and Deep Autoencoders for Industrial Screw Tightening Anomaly Detection." The podcast explores the intricacies of combining traditional machine learning models with deep learning techniques to address a critical problem in industrial manufacturing: detecting anomalies in screw tightening processes.

    The conversation highlights the importance of explainability in AI, particularly in industrial settings where safety and cost are paramount. The episode also touches on the broader implications of machine learning in AI, contrasting it with the current excitement surrounding generative AI models.

    We also extend a special thank you to Diogo and his team of researchers for developing this month's paper. If you are interested in reading the paper yourself, please visit this link: https://www.mdpi.com/2073-431X/11/4/54.

    For more information on all things artificial intelligence, machine learning, and engineering for your business, please visit www.deeperinsights.com or reach out to us at thepaperclub@deeperinsights.com.
    続きを読む 一部表示
    26 分
  • 18: The Future of Sport: AI-Generated Soccer Commentary

    18: The Future of Sport: AI-Generated Soccer Commentary
    2024/07/23
    In the latest AI Paper Club Podcast episode, hosts Rafael Herrera and Sonia Marques welcome Dr. Catarina Carvalho, Senior Data Scientist and Computer Vision SME from Deeper Insights, to discuss, "Match Time: Towards Automatic Soccer Game Commentary Generation". This paper introduces a method for generating engaging soccer commentary from video sequences using advanced AI techniques.

    The episode explores the importance of data quality, the innovative pipeline for dataset curation, and the perceiver-like architecture ensuring temporal coherence. It also covers broader applications, such as in cooking shows or assisting the hearing impaired. Tune in to discover how AI is revolutionising sports commentary and how you can try these techniques at home.

    We also extend a special thank you to the research teams from Shanghai University and Shanghai AI Laboratory for developing this month’s paper. If you are interested in reading the paper for yourself, please check this link: https://arxiv.org/abs/2406.18530

    For more information on all things artificial intelligence, machine learning, and engineering for your business, please visit www.deeperinsights.com or reach out to us at thepaperclub@deeperinsights.com.
    続きを読む 一部表示
    24 分
  • 17: Meta’s Chameleon: Redefining Data Integration with Mixed-Modal AI

    17: Meta’s Chameleon: Redefining Data Integration with Mixed-Modal AI
    2024/06/27
    In this episode of the AI Paper Club Podcast, hosts Rafael Herrera and Sonia Marques are joined by Andrew Eaton, an AI Solutions Consultant from Deeper Insights, to explore Meta’s latest paper, “Chameleon: Mixed Modal Early Fusion Foundation Models.” This paper marks Meta’s first steps into the mixed modal AI space, combining text, images, and other data types from the start for a more integrated understanding.

    The podcast explores how, unlike traditional models that process text and images separately before combining them, Chameleon integrates these modalities right from the beginning. This early fusion method promises enhanced performance in tasks like image captioning and interleaved text-image outputs, setting new benchmarks in the field.

    We also extend a special thank you to the research team at Meta for developing this month’s paper. If you are interested in reading the paper for yourself, please check this link: https://arxiv.org/abs/2405.09818.

    For more information on all things artificial intelligence, machine learning, and engineering for your business, please visit www.deeperinsights.com or reach out to us at thepaperclub@deeperinsights.com.
    続きを読む 一部表示
    28 分
  • 16: Infini-Attention: Google's Solution for Infinite Memory in LLMs

    16: Infini-Attention: Google's Solution for Infinite Memory in LLMs
    2024/05/22
    In this episode of the AI Paper Club Podcast, hosts Rafael Herrera and Sonia Marques welcome Leticia Fernandes, a Deeper Insights Senior Data Scientist and Generative AI Ambassador. Together, they explore the groundbreaking "Leave No Context Behind: Efficient Infinite Context Transformers with Infini-attention" paper from Google. This paper addresses the challenge of fitting infinite context into large language models, introducing the Infini-attention method. The trio discusses how this approach works, including how it uses linear attention and employs compressive memory to store key-value pairs, enabling models to handle extensive contexts.

    We also extend a special thank you to the research team Google for developing this month’s paper. If you are interested in reading the paper for yourself, please check this link: https://arxiv.org/pdf/2404.07143.pdf

    For more information on all things artificial intelligence, machine learning, and engineering for your business, please visit www.deeperinsights.com or reach out to us at thepaperclub@deeperinsights.com.
    続きを読む 一部表示
    23 分
  • 15: The AI Composer: Meta AI’s Innovations in Music and Melody

    15: The AI Composer: Meta AI’s Innovations in Music and Melody
    2024/04/23
    This month's episode of the AI Paper Club Podcast covers AI Music! Hosts Rafael and Sonia welcome Deeper Insights Data Scientist Joan Rosello to discuss the paper "Simple and Controllable Music Generation" from Meta AI. He introduces us to Meta AI's model that not only generates audio from text prompts but also introduces a novel feature—melody conditioning. This allows the creation of music that adheres to a provided melody, pushing the boundaries of AI in music generation. We also explore the technical concepts involved in creating AI music today, including residual vector quantization used in this model.

    We also extend a special thank you to the research teams at Meta AI for developing this month’s paper. If you are interested in reading the paper for yourself, please check this link: https://arxiv.org/pdf/2306.05284.pdf

    For more information on all things artificial intelligence, machine learning, and engineering for your business, please visit www.deeperinsights.com or reach out to us at thepaperclub@deeperinsights.com.
    続きを読む 一部表示
    24 分
  • 14: Using Gemini for Enterprise with Googler Jupiter Angulo

    14: Using Gemini for Enterprise with Googler Jupiter Angulo
    2024/03/18
    In this month’s episode of the AI Paper Club Podcast, hosts Rafael Herrera and Sonia Marques welcome special guest Jupiter Angulo from Google. He helps us explore the innovations behind Gemini, its applications in enterprise settings, and Google's latest advancements in multimodal models.

    Developed by Google's think tank, DeepMind, Gemini represents a significant leap forward, offering efficient, scalable, and cost-effective solutions for both enterprise and consumer applications. With Jupiter's insights, we get a behind-the-scenes look at the challenges and opportunities presented by AI integration across various enterprise sectors, exploring the model's versatility and Google's commitment to responsible deployment.

    We also extend a special thank you to the research teams at Google’s DeepMind for developing this month's paper. If you are interested in reading the paper for yourself, please check this link: https://arxiv.org/abs/2312.11805.

    For more information on all things artificial intelligence, machine learning, and engineering for your business, please visit www.deeperinsights.com or reach out to us at thepaperclub@deeperinsights.com.
    続きを読む 一部表示
    28 分