Machine Learning Street Talk (MLST) avatar

Dexa/Machine Learning Street Talk (MLST)

Learn more
Published Jul 8, 2020

Robert Lange on NN Pruning and Collective Intelligence

Robert Lange delves into the world of AI, uncovering the significance of intrinsic motivation and system-level thinking, the promise of multi-agent reinforcement learning for collective decision-making, and the intriguing Lottery Ticket Hypothesis in neural network pruning, offering insightful connections between cognitive science and technology.
Episode Highlights
Machine Learning Street Talk (MLST) logo

Topics covered

Popular Clips

Episode Highlights

  • Decentralized Control

    In the realm of multi-agent reinforcement learning, emphasizes the importance of decentralized control for real-world applications. He explains that centralized training with decentralized control allows agents to access shared information during training but operate independently during deployment, which is crucial for managing computational demands and real-time operations 1. This approach is mirrored in his research, where he studies fish behavior to distill intelligent exploration mechanisms into computational models 2.

    If we actually want to deploy something in the real world, having decentralized control is really, really critical.

    ---

    believes that understanding these mechanisms can lead to significant advancements in how autonomous systems interact with their environments.

       

    Collective Behavior

    explores the intersection of economics and multi-agent systems, highlighting the potential of reinforcement learning to address societal issues. He draws parallels between economic theories and collective behavior in multi-agent systems, suggesting that concepts like Markov decision processes are applicable across both fields 3. His journey from economics to cognitive neuroscience and machine learning reflects his commitment to understanding intelligence from a multidisciplinary perspective 4.

    Economics is really good at describing social phenomena and problems, but not necessarily very good at tackling them.

    ---

    believes that integrating these disciplines can lead to innovative solutions for complex challenges.

       

    Ethical Dilemmas

    The ethical implications of deploying multi-agent reinforcement learning systems are a significant concern for . He acknowledges the moral complexities that arise when these systems are applied to societal contexts, such as taxation policies, as highlighted by the controversial Salesforce paper 5. Despite these challenges, he sees value in using simulations to explore public policy and strategy, provided that the limitations of these models are recognized 6.

    Formalizing this as a reinforcement learning problem is inherently nothing bad. It gives a different perspective on the problem.

    ---

    stresses the need for conscious decision-making in the implementation of these technologies.

Related Episodes