In the quest to grasp our surroundings, the mind often relies on concepts, frequently conveyed through language or mathematics. However, this quest faces a fundamental dilemma: where do the concepts that shape our understanding originate? How can we devise a new theory using established terminology without falling prey to the assumptions embedded in that language?

The influence of 18th-century empiricists like Locke, Hume, and Berkeley continues to permeate modern scientific thinking. Empiricism posits that the world is an objective reality awaiting discovery. Our sensory experiences purportedly provide accurate descriptions of this reality, which we articulate through various concepts and abstractions. Yet, does this truly reflect the functioning of our brains?

Philosophy is the battle against the bewitchment of our intelligence by means of language.

—Ludwig Wittgenstein

Cognitive Neuroscience, while relatively new, utilizes terminology that can trace its roots back to ancient Greece. The 19th-century practice of phrenology exemplifies the potential pitfalls of oversimplified approaches to understanding the brain. Phrenology suggested that distinct brain regions corresponded to specific tasks—implying that one area manages knitting while another handles poetry.

Though this notion appears absurd today, some argue that contemporary cognitive neuroscience engages in a form of "neo-phrenology," relying on concepts such as attention, memory, and reasoning that were pioneered by 19th-century psychologist William James, among others. We continue to search for specific functions and algorithmic representations of concepts in the brain, which were defined long before we understood its complexities.

This situation is encapsulated by David Marr's framework, which proposes that we first conceptualize a function in terms of information processing, followed by identifying an algorithm, and ultimately determining the neural implementation of that algorithm. However, it’s important to note that Marr, a foundational figure in computational neuroscience, advocated for a reciprocal relationship among these components, beginning his career at the anatomical level.

Despite this, the prevailing view in the field remains one of understanding the brain from an outside-in perspective, akin to von Neumann's analogy of the brain as an information processor. We envision the brain as a computer, processing sensory data and responding based on derived insights.

Yet, as neuroscientist György Buzsáki argues in his insightful work The Brain From Inside Out, this perspective is severely limited. The subsequent sections will delve into Buzsáki's ideas, exploring their implications for neuroscience and AI, highlighting the shortcomings of our current understanding of the brain, and discussing how a shift in perspective might enhance our grasp of its functions.

Things won are done; joy’s soul lies in the doing.

—Troilus And Cressida, Act 1, Scene 1

How do sensory inputs gain significance for us? How do we ascertain what matters to our perception?

Buzsáki offers a straightforward response: the only method to connect inputs in a meaningful manner is through action. Actions ground our senses in meaning. The development of sensory systems in organisms is fundamentally tied to their ability to act. The brain prioritizes information that aids in fulfilling its needs over abstract truths. It doesn't merely process and evaluate evidence; rather, it actively engages with the environment. Perception is inseparable from action, as sensing often involves active engagement—be it through sniffing, head movements, echolocation, or microsaccades.

Children initially explore the physical properties of their bodies through random movements, and learn motor skills in speech by producing a range of syllables, eventually receiving feedback from parents or through experiences that shape their understanding.

Even scientific inquiry arguably advanced with the experimenter taking an active role in manipulating the world, exemplified by Galileo's legendary experiments with gravity from the Leaning Tower of Pisa.

Consequently, the speed at which an organism acts defines the relevant timescales for perception: understanding becomes futile if actions cannot be executed swiftly enough, a principle observable in nature. For instance, trees, unable to move, have no need for rapidly saccadic eyes to perceive their surroundings. The velocity of our muscular responses closely correlates with cognitive speed, suggesting that muscle speed might have historically limited cognitive development.

Buzsáki contends that we must reevaluate the representation-centric paradigm of neuroscience: instead of questioning what a neuron or neural assembly computes, we should inquire what it actively accomplishes.

This perspective is pragmatically significant, as the current neuroscientific paradigm influences experimental design and execution. In many studies, such as memory research, participants are typically exposed to stimuli while their neural responses are recorded. However, from an inside-out perspective, solely observing neural activity is insufficient because it lacks grounding.

It resembles documenting the words of an extinct language without a Rosetta Stone.

Buzsáki suggests that the brain’s vocabulary consists of internally generated dynamic sequences. Words can be viewed as sequences at the neuronal assembly level. Instead of forming new sequences during exposure to external stimuli, learning involves selecting pre-existing internal sequences that best correspond to new experiences. The brain is not a blank slate; it is equipped with stable dynamics from the outset. These dynamic sequences create a dictionary filled with initially meaningless terms, and learning entails establishing context for these terms to acquire meaning.

Without delving deeply into neuroscientific specifics, consider the formation of new memories involving the hippocampus and prefrontal cortex. The hippocampus is crucial for memory formation, as evidenced by patients with hippocampal lesions who struggle to create new memories. The hippocampus acts as a sequence generator, while the neocortex learns to identify relevant associations from those sequences, converting short-term sequences into long-term memories.

A compelling connection can be made between these ideas and concepts in artificial intelligence.

In the early AI field, the focus was largely on a symbol-based approach, where abstract representations of the world were hard-coded into systems, aiming to cultivate common sense and abstract reasoning. This approach exemplified the outside-in methodology, which ultimately proved inadequate, as it was challenging to translate abstract reasoning into genuine behavior.

’Tis written: “In the beginning was the Word!”

Here now I’m balked! Who’ll put me in accord?

It is impossible, the Word so high to prize,

I must translate it otherwise (…)

The Spirit’s helping me! I see now what I need

And write assured: In the beginning was the Deed!

—Goethe, Faust

Similarly, beginning with abstraction and moving towards action may fall short in elucidating brain function.

Another intriguing connection between AI and the inside-out perspective can be seen in a neural network architecture known as reservoir computing. Reservoir computers consist of an input layer, a hidden layer, and an output layer, similar to standard recurrent neural networks. However, recurrent neural networks face significant challenges with training due to the vanishing and exploding gradient problem.

An innovative solution to these gradient challenges is to keep the hidden layer fixed after random initialization. The reservoir then serves as a stable, non-linear dynamical system, rich in pre-configured dynamic sequences (analogous to neurons communicating with each other). The objective of learning is not to develop new dynamics from scratch but to train the output layer to align the existing reservoir dynamics with relevant data sequences.

This parallels Buzsáki’s ideas: the brain arrives with a vast array of complex non-linear dynamics due to its intricate wiring and the immense combinatorial potential of billions of interconnected neurons. When we learn to act, we do not begin from a blank state; rather, we derive meaning by associating these dynamics with beneficial outcomes in the real world.

This approach also addresses the issue of catastrophic forgetting that often plagues neural networks. They tend to erase previously acquired skills when learning new tasks, even if the new tasks are relatively similar to the old ones. From an inside-out perspective, if learning new information involves acquiring entirely new dynamics, it could destabilize existing dynamics. However, if it means integrating new knowledge into established patterns, catastrophic forgetting is less likely to occur.

Another noteworthy recent advancement in deep learning that aligns with this perspective is the edge-popup algorithm. This algorithm trains a sufficiently large, randomly initialized network not by adjusting weights but by removing connections that do not contribute meaningfully to the network's performance, effectively revealing a preconfigured subnetwork that accomplishes the task without requiring any training.

It should be noted that reservoir computing has lost favor in recent years, with architectures like LSTMs and GRUs becoming more popular for modeling sequential data. Nevertheless, insights from the inside-out approach could still inspire the AI and machine learning community, possibly leading to a resurgence of similar architectures in the future.

“Strong reasons make strong actions.”

—King John, Act 3, Scene 4

Buzsáki advances the idea further, proposing that many higher reasoning abilities, such as thought and cognition, can be viewed as internalized actions.

Memories may form when events in the world are mapped to existing dynamic patterns that best align with their content. Future intentions might even be seen as episodic memories of actions played in reverse.

The ability to internalize actions provides humans with a significant evolutionary advantage—the capacity to envision hypothetical futures. Imagining plans allows complex brains to choose actions based on predictions regarding their consequences, integrating past experiences with sensory input. As the complexity of potential outcomes increases, the brain adds layers of sophistication, but the underlying goal remains consistent.

The emergence of abstract thought may have stemmed from spatial reasoning, as explored in greater detail in my article on The Geometry of Thought, and as articulated by Barbara Tversky in her book Mind in Motion. This notion is further supported by the observation that the prefrontal cortex, often associated with human cognition, shares a neural architecture with the motor cortex. While the primary motor cortex connects directly to muscles, the prefrontal cortex influences higher-order brain regions related to emotional regulation, such as the hypothalamus, amygdala, and hippocampus.

“I have always thought the actions of men the best interpreters of their thoughts.”

—John Locke

The brain excels at interpreting reality and formulating thoughts. However, brains would not exist solely to compute integrals or flaunt knowledge.

This perspective strengthens Buzsáki’s argument: if we regard brains as a vast reservoir of dynamics aimed at aligning with meaningful experiences through action, it renders them practically valuable, firmly rooting his theory in evolutionary principles.

Our separation of action and intelligence may stem from the long-standing Cartesian divide between body and mind. It's time to bridge this conceptual gap, recognizing that action could be central to intelligence. It remains intriguing to consider whether forthcoming breakthroughs in neuroscience and AI will be action-driven, incorporating a more robust inside-out approach into their foundational concepts.

Finally, I must acknowledge that I have omitted numerous fascinating details in this article. I strongly recommend exploring Buzsáki’s book further or watching one of his presentations on YouTube (for instance, here or here). I have also recently discussed many of his ideas on my podcast, providing a deeper dive into his work and the neuroscience involved.