Hardware Podcasts

326: Wrong in the Right Way

Embedded.fm · 2 Apr 2020 · 1h 03m · Episode 326

What this episode covers

Erin Talvitie, an associate professor at Harvey Mudd College, discusses the challenges of model-based reinforcement learning, using the Atari game Pong to illustrate how small prediction errors can lead to catastrophic planning failures. She explains her research on 'hallucinated replay' and designing models that are 'wrong in the right way' to mitigate these issues. The conversation also covers the distinction between machine learning and artificial intelligence, recommended textbooks for learning the field, and the immediate ethical concerns of AI surveillance and bias.

Most relevant to

makers and hobbyists

Topics

Mentioned

Products
Pong, Atari 2600, Zoom
Also
Harvey Mudd College, Franklin and Marshall College, University of Michigan, Reinforcement Learning: An Introduction, Hands-On Machine Learning with Scikit-Learn, Keras, and TensorFlow, Deep Learning, David Silver, Richard S. Sutton, Andrew G. Barto, Aurélien Géron, Ian Goodfellow, Patrick Pularski

Related episodes from other shows

Other shows on the same ground, closest first.

  • Robot Talk · 33m

    Episode 121: Adaptable robots for the home - Lerrel Pinto

    Lerrel Pinto, an Assistant Professor at New York University, discusses his lab's work on general-purpose robots that can adapt to messy, unpredictable home environments using machine learning. He details the 'Okay Robot' system, which uses Vision Language Models (VLMs) and Large Language Models (LLMs) to interpret natural language commands and retrieve objects from cluttered real-world homes. Pinto explains the significant failure modes in robotics, such as hardware limitations and sensing errors, and contrasts the ease of AI tasks like chess with the difficulty of physical manipulation for robots. He also explores the challenges of imitation learning, noting that robots struggle to translate third-person observations into their own motor actions without extensive data or specific morphological compatibility.

    AI & machine learningRobotics & autonomy
    8 chapters

  • The Next Byte Podcast · 28m

    104. Recapping 2022

    Hosts Daniel and Forbode recap their top four episodes from 2022, covering biomimicry inspired by dandelion seeds, AI curiosity models like DALL-E, wearable sleep apnea detection, and sustainable hydropower turbines. They discuss the ethical implications of generative AI, the challenges of explaining complex topics like reinforcement learning to a lay audience, and their podcast's growth metrics, including a shift in global listenership away from the US. The hosts also outline goals for 2023, such as improving community engagement, producing better swag, and rebranding the show.

    AI & machine learningConnected products & IoT
    6 chapters

  • Robot Talk · 24m

    Episode Forty-Eight: Georgia Chalvatzaki

    Georgia Chalvatzaki, Professor of Robot Perception and Learning at the Technical University of Darmstadt, discusses her work with the Intelligent Robotic Systems for Assistance (IROSA) lab on mobile assistive robots. She explains the challenge of integrating classical planning with machine learning to handle complex household tasks like table setting while guaranteeing safety and stability. Chalvatzaki highlights the importance of continual learning to adapt to new environments without forgetting previous skills, and the need for multimodal human feedback to adjust robot behavior. She also addresses the integration of large language models like ChatGPT, warning about the risks of hallucinations and the need for robust uncertainty quantification before such robots can be deployed in semi-structured real-world settings like nursing homes.

    Robotics & autonomyAI & machine learning
    7 chapters

  • The Next Byte Podcast · 25m

    99. AI Learns How To Be Curious

    Hosts Daniel and Forbode discuss a research paper from MIT on 'Extrinsic Intrinsic Policy Optimization' (EIPO), an AI model that balances exploration and exploitation to learn faster. They explain how this approach, validated on 61 Atari games, allows AI to fluctuate between being curious and optimizing known strategies, potentially accelerating development in fields like drug discovery and generative hardware design. The hosts also touch on Microsoft Copilot and GPT-3 as examples of AI assisting developers, and mention Elon Musk's hiring strategy regarding video game engineers.

    AI & machine learningEmbedded firmware
    6 chapters

  • Oxide and Friends · 1h 23m

    Adversarial Machine Learning

    Nicholas Carlini, a security researcher specializing in adversarial machine learning, joins Bryan Cantrill and Adam Leventhal to discuss how sequences of seemingly random tokens can cause Large Language Models (LLMs) to ignore safety restrictions. Carlini explains that these attacks, which include 'poisoning' training data by exploiting expired domain names and using adversarial suffixes to bypass jailbreaks, demonstrate that LLMs are vulnerable to inputs that look like noise but trigger specific internal features. The conversation highlights the transferability of these attacks across different models, the importance of open-source models for security research, and the practical risks of deploying LLMs without understanding their failure modes.

    Security & privacyAI & machine learning
    8 chapters

The summary, chapters and topics are generated from the episode’s audio by this directory, not written by Embedded.fm. The episode itself belongs to its publisher.