Upside-Down Reinforcement Learning with Jürgen Schmidhuber - #357

Topics covered
Popular Clips
Episode Highlights
Simulated Success
AI's prowess in simulated environments is evident in its success with video games like Dota and Starcraft, where models like LSTM networks excel. explains how these models learn through policy gradient methods and reward optimization, improving over billions of simulations 1. This approach has been extended to real-world applications, such as OpenAI's robotic hand, which transitions learned behaviors from virtual to physical settings. However, Schmidhuber notes that real-world processes often lack the precise simulations available in virtual environments, posing challenges for AI adaptation 1.
You don't have that in the real world. You have all kinds of slippage, and it's not well modeled in any simulation.
---
In industrial settings, AI is being used to improve processes like quality inspection and manufacturing. Schmidhuber highlights projects with companies like Audi, where reinforcement learning enables toy cars to learn parking strategies through trial and error, showcasing AI's potential to enhance real-world operations 2.
Real-World AI
In real-world applications, AI faces challenges due to the lack of precise models, unlike in controlled environments. Schmidhuber emphasizes the need for AI to learn from experience, akin to a baby learning through interaction, to build models of industrial processes 3. This approach is crucial for tasks like assembling smartphones, where robots must quickly adapt to new tasks through a combination of curiosity and imitation 4.
The future of manufacturing is all the messy things, all the messy stuff, like when you are assembling a smartphone.
---
Reinforcement learning plays a significant role in these advancements, as seen in projects like OpenAI's robotic hand, which uses LSTM networks to perform complex tasks such as solving a Rubik's cube 5. Despite the challenges, AI continues to make strides in bridging the gap between virtual simulations and real-world applications.
Related Episodes


Reinforcement Learning Deep Dive with Pieter Abbeel - #28
Answers 383 questions

Trends in Reinforcement Learning with Simon Osindero - TWiML Talk #217
Answers 383 questions

Advancements in Machine Learning with Sergey Levine - #355
Answers 383 questions

Robotics at OpenAI with Jonas Schneider - #76
Answers 383 questions

Deep Robotic Learning with Sergey Levine - #37
Answers 383 questions

LSTMs, Plus a Deep Learning History Lesson with Jürgen Schmidhuber - #44
Answers 383 questions

Deep Reinforcement Learning for Game Testing at EA with Konrad Tollmar - #517
Answers 383 questions

Hierarchical and Continual RL with Doina Precup - #567
Answers 383 questions

Deep Reinforcement Learning for Logistics at Instadeep with Karim Beguir - #302
Answers 383 questions

Trends in Reinforcement Learning with Chelsea Finn - #335
Answers 383 questions

Inverse Reinforcement Learning Without RL with Gokul Swamy - 643
Answers 383 questions













