Alexander Ororbia ~ Active Inference Insights 008 ~ Mortal Computation, Cybernetics, AI
Feb 29, 2024
auto_awesome
Delving into mortal computation and cybernetics, the podcast explores biomimetic intelligence and the benefits of in-memory computing for neural simulations. It discusses minimizing variational free energy in biological and artificial systems, as well as the concept of cybernetic variety in problem-solving. The episode also touches on active inference, Bayesian surprise, and the limitations of current robotics technology in achieving adaptive intelligence.
Mortal computation advocates for building adaptive AI systems inspired by natural intelligence.
Cybernetics principles include ultra-stability, growth, and noise for effective problem-solving.
Regulation in cybernetics relies on the law of requisite variety and the good regulator theorem.
Mortal computation offers a promising alternative to current AI models like neural networks.
Active inference provides a neurobiologically inspired approach for adaptive AI decision-making.
Cybernetics frameworks like blind selection and retention enhance understanding and functionality in dynamic systems.
Deep dives
Professor O'Rorbia's Background and Research Focus
Professor Alexander O'Rorbia, an assistant professor of computer science at the Rochester Institute of Technology, directs the Neural Adaptive Computing Laboratory. His research entails developing new learning procedures and computational architectures based on theories of mind and brain functionality. He co-authored the paper 'Mortal Computation: A Foundation for Biomimetic Intelligence' with Karl Fristen, advocating for Active Inference's potential in guiding computational system construction.
Journey Into Active Inference and Predictive Coding
Professor O'Rorbia's involvement in Active Inference stemmed from his work with colleagues in eye tracking and cognitive science research, where they explored reinforcement learning control models. Given his background in predictive coding, Active Inference's extension naturally aligned with his research interests, offering a solution to handling exploration challenges in their computational models.
Application of Active Inference in Real-world Tasks
Building on his understanding of predictive coding circuitry, Professor O'Rorbia delved into integrating Active Inference at a raw, low-level model, deviating from conventional deep learning approaches. His collaboration led to a model integrating Active Inference principles in gaze tracking tasks, emphasizing the importance of predictive coding in developing biological models.
Mortal Computation Thesis and Biomimetic Intelligence
The Mortal Computation thesis posits a paradigm shift in computational frameworks where software is intertwined with hardware, promoting energy efficiency and reducing environmental impact. Biomimetic Intelligence draws inspiration from the adaptability and intelligence of natural organisms, striving to emulate these qualities in technological artifacts, paving the way for environmentally friendly artificial intelligence systems.
Understanding Variety in Cybernetics
In Cybernetics, variety refers to the number of states in a system, indicating the cardinality of the state space. By considering the example of debugging a light not turning on, variety is key to problem-solving. For instance, if analyzing the light switch with two states (on or off) has internal variety of one, and adding another component with two states results in an internal variety of two. However, if an unseen third component with three states causes failure, violating the law of requisite variety.
Key Principles in Cybernetics: Stability, Growth, and Noise
In Cybernetics, key principles include ultra-stability, growth, and noise. Ultra-stability focuses on maintaining balance between internal and external variety for homeostasis. Growth involves auto-catalytic growth and recursive systems, describing the formation of complex structures. Noise, essential for exploring space, leads to order through the signal to order principle.
Regulation and Control in Cybernetics
Regulation in Cybernetics is guided by the law of requisite variety and the good regulator theorem. To control and manipulate the environment effectively, having an accurate internal generative model, representing a model of the external environment, is crucial. The Model of Internal Model Controller (IMC) principle stresses stability in a closed-loop feedback system for an effective synthesis of control and environment.
Exploring Further Frameworks and Implications in Cybernetics
Beyond the basics of cybernetics, additional frameworks like blind selection and retention and order through noise connect to dealing with chaotic and random dynamic systems. These principles underscore the significance of variety, regulation, growth, and noise in enhancing understanding and functionality within cybernetics.
Philosophical Concepts and Problem-Solving Perspectives
Delving into cybernetic variety, the law of requisite variety, and other key principles underscores their applicability in problem-solving contexts. By analyzing the interplay of principles such as stability, growth, noise, and regulation, cybernetics offers a comprehensive framework for understanding and navigating complex systems for effective problem resolution.
Importance of Mortal Computation in Machine Intelligence
Mortal computation, as discussed in the podcast, offers a valuable perspective on the future trajectory of artificial general intelligence (AGI). Despite the current focus on models like neural networks and reinforcement learning, the podcast emphasizes mortal computation as a promising alternative for building intelligent systems. Mortal computation highlights the significance of adaptive systems that can navigate uncertain environments, self-repair, and evolve. This approach provides a more flexible and robust framework for developing AI systems that can effectively interact with and learn from complex real-world scenarios.
The Role of Active Inference in AI Development
Active inference emerges as a crucial concept in the podcast, representing a shift towards more adaptive and exploratory approaches in artificial intelligence. Compared to traditional reinforcement learning methods, active inference offers a more nuanced and neurobiologically inspired framework for learning and decision-making. By emphasizing the importance of prior preferences and intelligent exploration, active inference addresses key limitations of current AI models, such as sample inefficiency and brittleness. The podcast underscores the potential of active inference to drive advancements in AI research and application, particularly in navigating noisy and uncertain environments.
With all the current excitement around the prospect of artificial general intelligence, it is easy to forget that there already exist exemplars of problem-solving and creativity all around us - namely, ourselves. With this in mind, join Darius and computer scientist Alexander Ororbia for a discussion of biomimicry and mortal computation, which centres on the viability of building artificial systems modelled on the evolved, intelligent life forms on this planet.
Remember Everything You Learn from Podcasts
Save insights instantly, chat with episodes, and build lasting knowledge - all powered by AI.
