Questioning the Computational Nature of Cognition:
Rodney Brooks challenges the widely accepted notion that cognition is computational in nature, stating he doesn’t believe cognition to be computational, while admitting it’s a question, not a concrete assertion.

He uses various real-world examples to illustrate instances where we wouldn’t naturally attribute computational processes, such as stones being scattered by cars or the physics involved in launching rockets.

Analogy of Non-Computational Processes:
Brooks provides the example of how rocks, leaves, and sticks scatter off the road due to car movement, and argues this isn’t a computational process, but a random scattering due to the car’s motion.

Similarly, he argues that while launching a rocket involves computations for simulations and controls, getting the rocket into space is fundamentally a physical problem, not a computational one.

Differentiating between Brains and Computers:
Brooks questions the common analogy that brains are essentially biological computers. He points out that we don’t consider non-neuronal entities like rocks, rockets, or trees to be computational.

He argues that if non-neuronal entities can accomplish complex tasks, like trees searching for nutrients, then these tasks might not be computational at all.

The Definition of Computation and its Implication for AI:
He objects to redefining computation to encompass whatever the brain does as it creates a tautology that prevents us from understanding whether our current computational systems (computers) can truly mimic the brain or produce intelligence.

By leaving this definition vague, it creates ambiguity around the goal of building intelligent AI programs with conventional computers.

Limitations of Computation in Predicting Complex Systems:
Brooks points out that despite significant computational resources, we still struggle to accurately predict the weather, a complex system. This analogy suggests that intelligence or cognition, which are arguably more complex, might also prove to be beyond the reach of computational modeling.

Social Construction of Computation:
Brooks proposes that computation, unlike properties such as mass or energy, is a socially constructed concept rather than a fundamental property of the universe.

This socially constructed understanding of computation has been tied to the concept of computers as we understand them, based on the Turing definition.

However, Brooks questions the accuracy of this definition as he believes many things in our world, like rockets, aren’t fundamentally computational but can merely be described metaphorically using computations.

The Birth of Artificial Intelligence:
Brooks reflects on the birth of artificial intelligence (AI) in 1956, when the term was coined by John McCarthy, and the belief that any aspect of intelligence could be simulated by a machine.

The inherent assumption here was that the machines of the time, referred to as automatic calculators, were universal machines capable of simulating any other sort of machine.

Questioning Computation:
Brooks has long questioned the pervasive belief in computation as the foundation of all machine intelligence.

In a 2001 paper, he discussed the idea that more than computation might be necessary for truly replicating intelligence and life, suggesting that we might need a new, non-threatening concept that includes the situatedness and embodiment of intelligent entities in the world.

Cybernetics vs. Computation:
Brooks discusses the rivalry between computationalism and cybernetics in the 1950s, with the latter losing despite its focus on understanding all possible behaviors of an individual machine.

Cybernetics focused on continuous functions and lacked the concept of abstract representation, which was a cornerstone of AI. Cybernetics was also about the coupling of creatures to their environment, a feature largely ignored by AI.

The Misinterpretation of Computation:
Brooks criticizes the loose and often imprecise usage of known results about computation to support various arguments in AI.

He argues that merely invoking the term ‘computation’ doesn’t add precision to these discussions and suggests that a more nuanced understanding of the nature of computation is needed.

Four Key Fields of Modern Computation:
Brooks identifies four fields pivotal to the understanding of modern computation: Artificial Intelligence (AI), Artificial Life, Neuroscience, and Abiogenesis (the origin of life from non-living matter).

These fields emerged between 1945 to 1965, with significant overlapping contributions from many of the same researchers.

The fields of AI, Artificial Life, and Neuroscience have explicitly recognized their connection to computation, potentially limiting broader conceptual understanding.

Historical Evolution of Algorithms and Computation:
Algorithms have been around for at least 2000 years, with early instances such as Heron’s method for calculating the square root of a number.

The use of algorithms was often inspired by practical needs, such as land adjudication.

The industrialization of algorithms began around 1614 with John Napier’s hand-computed logarithmic tables, which proved useful for other computational tasks.

The Development of Computation Devices:
The demand for tables of functions during the Industrial Revolution led to attempts to mechanize computation, although early efforts, like those of Leibnitz, weren’t successful.

Charles Babbage’s designs for a “calculating engine” and an “analytical engine” laid the groundwork for modern computational devices.

The analytical engine, described in detail by Ada Lovelace, was equipped with an arithmetic logic unit, a sequencer (or program) with possible loops, storage for variables, and input/output facilities.

Alan Turing’s Impact on Computation:
Turing, motivated by Hilbert’s problems, made significant contributions to the theory of computation.

In his paper “On Computable Numbers,” Turing explored the application of computation to decision problems, thereby advancing the understanding of computation’s potential scope and capabilities.

Historical Context of Computational Theory:
Brooks examines the development of computational theory, from Minsky’s 1967 book on theoretical computer science to contemporary perspectives. He highlights how the definition of computation has been refined and how machines’ capacities have evolved. He touches on the notion of Turing machines and the shift to the finite memory model of the RAM, showing how our conceptualization of computational capacity has expanded with technological progress.

The Nature of Algorithms:
Donald Knuth’s definition of algorithms is explored. An algorithm, according to Knuth, should have finiteness, definiteness, have input/output, and be effective. Effectiveness, in this context, means the operations should be simple enough that they could be done by a person using a pencil and paper in a finite length of time.

Limitations of Computational Models:
Brooks points out that computational models don’t capture all aspects of cognition. He highlights how computations may not give the same output when repeated, a trait that differs from many natural systems, including our brain. He suggests this could indicate a fundamental difference between digital computation and cognitive processes.

Human Spatial Understanding:
He discusses the importance of spatial understanding as foundational to abstract thought, referencing psychologist Barbara Tversky’s laws of cognition. He suggests that our ability to think in spatial terms significantly influences how we conceptualize and explain phenomena, including computation.

Computation as a Social Construct:
Brooks posits that computation is not purely a mechanical or mathematical construct, but rather a social agreement influenced by our ability to think and reason about it. He acknowledges that a great deal of theory has been developed around computation, but still points to the role of human intuition in shaping it.

Computation as a Metaphor in Various Disciplines:
He describes how computation has become a primary metaphor or analogy for several scientific and engineering disciplines, including neuroscience and artificial intelligence. However, he highlights that some fields, like abiogenesis (the study of how life arises naturally from non-living matter), still primarily rely on other models such as chemistry.

Historical Definition of Computers:
Brooks notes the historical definition of computers, pointing out that they were originally defined as people, not machines. This historical insight challenges our current understanding of computers as purely mechanical entities.

Significance of Claude Shannon’s Work:
Claude Shannon, a co-author of the Artificial and Intelligence proposal in 1956, is significant for showing that Boolean algebra could describe relay and switching circuits.

He invented AND gates, OR gates, and NOT gates, fundamentally changing how we think about and understand circuit diagrams.

Shannon’s work is noted for introducing abstract thinking into understanding circuits.

The McCulloch and Pitts Paper:
The 1943 paper, “A logical calculus of the ideas imminent in nervous activities” by McCulloch and Pitts, attempted to describe the brain using a logic similar to Shannon’s gates.

Their description of neurons as two-input machines, using inhibitory and excitatory signals, and likening this behavior to AND and OR gates was revolutionary but later proved to be overly simplistic.

Despite this, their work greatly influenced others, including John von Neumann and Marvin Minsky.

Von Neumann and Artificial Life:
John von Neumann’s design of the EDVAC (Electronic Discrete Variable Automatic Computer) was inspired by McCulloch and Pitts’ work, describing the computer in terms similar to nerve cells.

Von Neumann also conceptualized self-reproducing machines and anticipated the discovery of DNA’s role in biogenesis.

This lineage of thinking eventually led to the development of genetic algorithms, the basis of artificial life, by John Holland, a student of a student of Von Neumann’s.

Consequences of McCulloch’s Influence:
McCulloch’s influence led to a culture of speculating about brain function within the AI community, leading to assertions that weren’t necessarily grounded in scientific evidence. This is exemplified by Marvin Minsky’s K-Lines theory.

Despite the controversy, McCulloch and his colleagues, often referred to as his “intellectual posse,” had a significant influence across fields, pushing for a reductionist approach in science.

Notably, this group, which included pioneers like John McCarthy and Manuel Blum, argued for computation as the primary metaphor for artificial intelligence and neuroscience.

Computation as a Metaphor for Brain Function:
Brooks reflects on the adoption of computation as a metaphor for brain function and neuroscience, noting how this idea started as a hypothesis but has become widely accepted.

He cautions against the overuse of this metaphor, especially when people equate systems capable of Turing computation with a universal model of the brain or the universe.

Limits and Dangers of Scientific Metaphors:
Rodney Brooks discusses the limits and potential pitfalls of metaphors in scientific understanding. Metaphors can be wrong or incomplete, leading to centuries-long misunderstandings, such as the “phlogiston theory” that proposed there was a common material in everything that could burn, while the truth was that the common material (oxygen) was in the air, not in the burning objects.

He mentions how metaphors are used to explain complex phenomena like quantum mechanics. However, these metaphors, such as particle-wave duality, often fall short in explaining all aspects of the phenomena.

Quantum Mechanics and Metaphors:
Metaphors are often relied upon to explain quantum mechanics, leading to different interpretations such as the Copenhagen and many-worlds interpretations. But these are still metaphors, not direct descriptions of the underlying phenomena. Real quantum mechanicians use abstract algebra, not relying on intuitive models.

Metaphors in Cosmology:
Metaphors are also used in cosmology, such as dark matter, to explain the abundance of gravity in the universe. The idea of dark matter, invisible mass that explains gravitational effects, is essentially a metaphor that we use to explain things we don’t fully understand yet.

Scientific Tools as “Black Boxes”:
Brooks refers to scientific tools such as calculus, probability theory, group theory, and numerical methods as “black boxes”. These are well-understood and broadly applicable tools, but sometimes their underlying assumptions about the world can lead us to gloss over important details or misuse them.

Brains as Computational Engines:
Brooks questions the metaphor of the brain as a computational engine. He suggests that as our understanding advances, we may eventually discard this metaphor, just as we have discarded the idea of trees as computational engines despite their complex behaviors.

New Tools for Understanding:
Brooks suggests that we may need new tools or “black boxes” to better understand complex systems such as the brain. However, he admits he does not know what these new tools might be.

Alternative Approaches:
He references alternative approaches to understanding complex systems, such as homeostasis and autopoiesis. Homeostasis is about self-regulating systems, while autopoiesis concerns the way a cell regenerates itself, maintaining its identity through a network of processes, not through the identity of individual atoms. These theories propose different ways of looking at natural systems and could inform our understanding of artificial intelligence and robotics.

Perando’s Paradox:
Perando’s Paradox, devised by an Argentine mathematician, presents a paradoxical situation with two hypothetical slot machines, A and B.

Machine A operates on a simple probability mechanism of winning and losing (with a slightly higher probability of losing), while Machine B has its win/lose probabilities dependent on the amount of money you currently have.

Running numerous trials reveals that both machines A and B lose more often than not. However, the paradoxical element arises when one alternates randomly between the two machines, which somehow results in a winning strategy.

Brooks highlights that our intuitions and tools are currently inadequate to fully understand these complex, interacting processes linked by structure.

Introduction of Structure:
Brooks introduces a different scenario where bits in a grid (mimicking a 4×4 chessboard) are flipped randomly until the entire board is either black or white.

By introducing a geometric structure (e.g., making the board a torus), and flipping a square to match its neighbor’s color, Brooks finds that this does not accelerate the board reaching all one color as expected.

However, further tweaking the rules (choosing a square to change a neighbor), the behavior changes even though the statistical point of view seems identical.

This experiment demonstrates that a little bit of geometric structure can emerge from random events.

Internal Structure from Random Events:
With the tweak that a square has to choose a change that affects a neighbor, Brooks notes that the resulting grid shows protection levels against change for each square.

Each square’s “protection level” is the number of steps it would take before that square could potentially change color.

Through this, patterns begin to emerge showing how many steps of “protection” a configuration of squares has from changing color.

This shows how an internal structure can come from random events, a concept that may be counterintuitive but proves important for understanding complex systems.

Perception and Existence of Unrecognized Tools:
The questioner asked whether we’re missing something in our cognitive computation models or simply not recognizing what’s already there.

This reflects a notion of unrecognized tools or methods in our current approach to cognitive computation.

Temporal and Ensemble Averages in Dynamics:
The questioner introduced the idea of ensemble and temporal averages, a concept rooted in ergodic theory, being potentially unequal in dynamic systems.

This perspective highlights the existence of historical dependence in dynamics, which could explain the perceived shortcomings in the current approaches.

Questioning Fundamental Assumptions:
The questioner suggested that people have started questioning fundamental assumptions of ergodic theory across different disciplines, such as economics.

This shows a trend towards critically examining the foundational principles that drive current cognitive computation models.

Need for New Tools and Concepts:
Brooks responded by acknowledging that we may not have the necessary tools to conveniently handle these issues yet.

He emphasised the ongoing nature of science and tool development, suggesting that advancements in understanding ergodic theory could be an example of this process.

The Challenge of Hidden Geometric Structures:
Brooks pointed out that hidden geometric structures are hard to reason about with current tools, acknowledging an area for future exploration and understanding.

Interplay of Dynamics and Space:
The questioner discussed the non-intuitive intertwining of dynamics and space in non-ergodic systems.

This shows an acknowledgement of complex phenomena that aren’t easily understood with our existing conceptual toolkit.

Call for Tool Development:
Brooks expressed a desire for the development of new, intuitive tools to capture these complex dynamics.

He used the example of Perundo’s paradox to illustrate the complexity of some problems, noting the extensive work needed to make probabilities work out in just that one case.

This underscores the need for more generalized tools in the field of cognitive computation.

Scientific Perspectives Post-Talk:
The questioner, identifying as a neuroscientist, asked Brooks for insights on how scientists should approach their work differently after his talk.

This indicates an interest in practical implications of his discussion on cognitive computation for scientific research.

The Bayesian Brain Hypothesis:
The questioner highlighted the existence of diverse views in neuroscience and computational science, including the ‘Bayesian brain’ hypothesis, which posits that the brain functions according to Bayesian principles.

This shows a diversity of theories and approaches to understanding brain function.

Abstraction vs. Building Blocks:
The questioner suggested a distinction between the abstract principles some believe the brain operates on and the tangible ‘building blocks’ and interactions between them, which they personally believe are key to understanding cognition.

This difference in focus reflects contrasting approaches to understanding the brain and cognition: a more top-down, abstraction-based approach versus a bottom-up, building blocks-based approach.

Abstraction Misleads Understanding:
The questioner argued that while abstraction layers can sometimes assist in understanding complex systems, they often lead scientists astray.

Brooks agreed with this sentiment, citing the McCulloch-Pitts paper as an example of how certain concepts can mislead researchers, especially in the AI field.

Questioning Existing Abstractions:
Brooks stressed the need for scientists to critically evaluate their abstractions and to be wary of disregarding data that contradicts their models.

He advised against focusing solely on ideal cases that fit within existing abstractions, indicating the need for a more open and adaptive approach to data and theory.

Invention of New Reasoning Tools:
Brooks underscored the probable need to invent new reasoning tools, noting that we don’t yet know what these will look like.

This highlights the ongoing and evolving nature of our understanding of complex systems like the brain.

Skepticism Towards Artificial General Intelligence:
Brooks expressed skepticism towards the imminent development of artificial general intelligence, based on current approaches in neuroscience and AI.

This reflects a viewpoint that we still have significant gaps in our understanding of cognition and brain function.

Hope for Future Innovation:
Despite his skepticism, Brooks conveyed hope that his talk might inspire innovative thinking in future generations.

He expressed a desire for the development of new ways to discuss and understand structure, which could prove valuable across a range of complex systems, such as economics.

Acknowledging the Unanswered:
Brooks noted that his talk didn’t provide specific answers or prescriptions, but posed questions for further contemplation.

This captures his emphasis on sparking new ideas and promoting continuous exploration in the field of cognitive computation.

Reconsidering the utility of computational abstractions:
Brooks highlights the potential dangers of relying heavily on computational abstractions, or ‘black boxes,’ when studying complex systems. He cautions that these abstractions can sometimes limit understanding or misdirect investigation, particularly when they are uncritically accepted or inaccurately applied.

The risk of reductionism in neuroscience:
He further points out that reductionist methods, like studying individual cells to understand the function of a whole computer, may not yield comprehensive understanding. The tools used in investigation are influenced by preconceived notions of how something works. This can bias the process and results of research, hence, Brooks advises caution.

Struggle with the prevalent “computation” narrative in AI:
Brooks critiques the prevailing notion that computation is the core of intelligence, which can lead AI researchers to oversimplify the process of cognition. He argues that this misinterpretation stems from labeling a wide array of processes as “computation,” which can confuse different fields’ unique meanings and theories.

The need for a nuanced understanding of “computation”:
According to Brooks, the word “computation” carries certain connotations (e.g., Turing computability) that might limit the conceptualization of unconventional computing methods, which may perform tasks beyond the scope of traditional Turing machines. He suggests a more nuanced approach, avoiding the term “computation” when describing certain processes to prevent potential misunderstandings.

Black boxes as external tools, not system components:
Brooks clarifies that his use of ‘black boxes’ refers to external tools for investigation, not components within the system being studied. He doesn’t advocate for a universal approach (like ‘pan computationalism’ or ‘pan black boxes’), but for the development and application of multiple, diverse investigative tools, each suited to the specifics of what’s being studied.

Non-Computational Approaches to Cognition:
The audience member raises the topic of non-computational approaches to cognition, such as enactivism, which Rodney Brooks acknowledges. However, these approaches are considered marginal and not mainstream.

Reasons for Marginality:
In response to why these approaches have not become dominant, Brooks suggests that these models might not be accurate or powerful enough. He also acknowledges the herd mentality in science, where the majority gravitates towards an approach that seems to be making progress, in this case, computation.

Consequences of Dominant Approaches:
Brooks uses the example of Alzheimer’s disease to illustrate the possible negative impact of a dominant approach. He explains that early mechanisms chosen to understand Alzheimer’s disease might have been incorrect, starving out other potential approaches. This is a cautionary tale about the dangers of an overly dominant approach, in this context, the computation approach in AI.

Role of Moore’s Law:
The computational approach has been greatly aided by Moore’s law, which predicts the doubling of computing power approximately every two years. This has allowed AI programs to improve without significant theoretical breakthroughs, just by waiting for more powerful machines.

Challenges in Breaking Established Thinking:
Brooks discusses the difficulty in shifting away from established modes of thought, especially when the alternative theories or models are weaker or less effective. He states that it may take a long time for a new approach to emerge that can challenge the dominance of the computational perspective.

Clarification of Computation and Cognition:
Brooks clarifies his statements about cognition and computation. He believes cognition is not necessarily computational but asserts we utilize computation to understand cognition. He highlights that computation is merely a tool, not the definitive essence of cognition.

Metaphor and Misunderstanding:
He argues that using computation as a primary metaphor for cognition may be misleading. The metaphor may help our understanding, but it’s not necessarily reflective of the actual process or mechanics. As an analogy, he cites how SpaceX uses computational simulation for rocket design, but that doesn’t physically launch rockets into space.

Misuse of Metaphors in Science:
Brooks points out the danger of clinging to metaphors in scientific discussions. He suggests that over-reliance on metaphors can lead to confusion and misconceptions. Using quantum physics as an example, he explains that while metaphors can be useful, if pushed too far, they can lead to erroneous understanding.

Critique of Neuroscience and Computation:
Brooks criticizes the field of neuroscience for what he sees as a too rigid association between cognition and computation. He believes this association came from a limited group of researchers tackling a set of problems and then became broadly accepted. He fears this might misguide the field into believing computation is a fundamental component of cognition rather than just a tool for understanding.

The Exponential Growth of Computation:
Brooks discusses the incredible rise in computational power during his lifetime, a factor of 10 to the 15th increase. He points out this trend may have further fueled the attachment of the metaphor of computation to cognition due to its ubiquity and impact on our lives.

Computation as a Metaphor for Modern Computers:
He reveals that even modern computers are not purely computational in the way we traditionally think. For instance, an Intel processor can have up to 180 instructions operating concurrently, which breaks from our standard serial understanding of computation. Therefore, he says that computation has become a metaphor even for the machines designed to carry it out.

Non-reduction of cognition to computation:
Rodney Brooks acknowledges that cognition and behavior are not completely reducible to computational processes. He finds it intuitive to consider the influence of embodiment and complex spatial geometry in the process.

Brooks questions why evolution chose the path of creating complex bodies if cognition and computation could be achieved without such bodies.

Clash with mainstream technological perspective:
Brooks is arguing against the mainstream technology-driven people, namely computationalists, who have influenced the fields of neuroscience and others.

He shares his experiences of receiving vehement backlash from such individuals when expressing his views on different platforms, including social media like Facebook.

Debate on the metaphor of computation:
Brooks’ counter viewpoint on computation as cognition is compared to a virus that has the potential to spread, challenging the widely accepted perspective.

His viewpoint has provoked reflection and questioning, leading some audience members to reconsider their understanding of computation.

Reception and Impact of the Discussion:
Despite Brooks’ perspective being seen as radical in many places, the audience here appears appreciative and understanding of his ideas.

The discussion has led to a shift in thinking for some audience members, indicating the infectious nature of his ideas.

Conclusion and Appreciation:
Brooks’ presentation is appreciated for its thought-provoking content and for inspiring rigorous questioning and discussions about the nature of cognition and computation.