Palmistry

Peter DayanPLoS 2004 · DOI

Abstract

In this review, computational neuroscientist Peter Dayan evaluates the provocative theories presented in On Intelligence by Jeff Hawkins and Sandra Blakeslee. Hawkins, an entrepreneur turned neuroscientist, proposes that intelligence is fundamentally a hierarchical system of prediction within the neocortex rather than a set of behaviors. While Dayan praises the book's infectious optimism and its potential to inspire a new generation of researchers, he critiques its "cortical chauvinism"—the tendency to overlook the essential subcortical machinery that governs emotion, value, and motor precision.

Introduction

Theoretical neuroscience faces a grand challenge: creating a unified, functional model of how the biological brain produces intelligence. In On Intelligence, Jeff Hawkins (founder of Palm and Handspring) argues that the Artificial Intelligence (AI) community has been "fighting yesterday’s battles" by focusing on behavior and the Turing Test approach. Hawkins proposes a shift in perspective, suggesting that the core of intelligence is the neocortex’s ability to extract statistical structures from sensory input to predict future events.

The cover of "On Intelligence" by Jeff Hawkins with Sandra Blakeslee.

Dayan’s review serves as a bridge between Hawkins’ engineering-led intuition and the rigorous, often messy, realities of biological neuroscience. He examines whether Hawkins' elegant, simplified model can withstand the complexities of actual brain anatomy and function.

The Theoretical Framework

Hawkins’ proposal rests on several key pillars of hierarchical unsupervised learning. He views the neocortex as a general-purpose engine that extracts relationships from raw data—such as pixels in a movie or the pressure on skin—without the need for an external "teacher." This system utilizes predictive auto-association, storing sequences of patterns and retrieving them based on similarity to past experiences to forecast upcoming stimuli.

Central to this theory is the concept of invariant representations. Hawkins posits that the brain stores information in a form independent of "deformations," allowing us to recognize a face regardless of lighting, pose, or distance. He attempts to map these functions onto specific cortical layers, suggesting that different types of neurons are dedicated to either "boosting" lower-level activity (priors) or "suppressing" it (predictive coding) to save bandwidth.

Critical Insights and the "Lumper" Philosophy

Hawkins identifies as a "lumper," a term in biology for those who ignore the immense complexity of individual parts in favor of finding universal principles. He focuses on the "crystalline regularity" of the cortical structure, arguing that the same basic algorithm is used for vision, hearing, and even high-level thought. Dayan notes that while this abstraction often irks experimental biologists, it is a necessary step for finding functional principles in complex systems.

However, Dayan introduces a sharp critique known as the "carpet problem." If the brain’s primary goal is to model all statistical structure, it risks wasting representational power on irrelevant details—like the intricate, meaningless patterns in a carpet. Dayan argues that true intelligence requires knowing what not to represent, a filter that Hawkins’ theory lacks.

Limitations: Cortical Chauvinism

The most significant critique Dayan levels against the book is its "cortical chauvinism." Hawkins focuses almost exclusively on the neocortex, treating it as the seat of all that is intelligent. Dayan points out that "graceful behavior," such as a child hopping across rocks, relies heavily on the cerebellum and other subcortical systems that Hawkins largely ignores.

Furthermore, the model neglects the basal ganglia and neuromodulatory systems. These are essential for assigning "value" to outcomes and making decisions based on rewards or threats. Without these systems, a predictive model is merely a passive observer, unable to translate predictions into meaningful action. Dayan also cites "change blindness" experiments to argue that the brain represents far less of the world than Hawkins suggests; we don't build a perfect predictive model of everything, but rather focus on what is salient to our survival.

Conclusion

Despite these technical and philosophical disagreements, Dayan concludes that On Intelligence is a successful and appealing "call to arms." It provides a neat solution for the dilemma of hierarchical models—explaining how higher brain areas can simultaneously provide "prior" information to boost lower areas while suppressing redundant, predicted information.

While the book’s treatment of the philosophy of mind and the history of AI may be somewhat "gossamer," its central thesis remains a stimulating challenge to the field. Future work must reconcile Hawkins’ elegant cortical hierarchies with the messy subcortical reality of biological organisms to truly understand the nature of intelligence.

Summary updated

Abstract

A review of "On Intelligence", a book exploring brain function by entrepreneur Jeff Hawkins and science writer Sandra Blakeslee
BibTeX
@inproceedings{dayan2004palmistry,
  author = {Peter Dayan},
  title = {{Palmistry}},
  booktitle = {PLoS},
  year = {2004},
  doi = {10.1371/journal.pbio.0020394},
}
Figure