On 'On Intelligence'
On Intelligence by Jeff Hawkins
My rating: 5 of 5 stars
On Intelligence is a telling reminder that there is as much role for theoretical ideation as is for experimental inquiry in the field of neuroscience. At a time when any concrete ‘universal’ theories of the brain are elusive, books like these bring fresh perspectives through the sheer boldness of their hypothesis.
So this then is intelligence. Intelligence, as understood in context of the brain, stems from an ability to glean information from the inherent statistics of the natural world at various levels of abstraction in order to make better forecasts about future events. In plain English, nature has certain hidden regularities and the job of any brain is to unearth those regularities in the interest of survival. Hawkins identifies the Neocortex as the spring from which intelligence gushes forth.
The tantalising take home message of the book is that, essentially, the cortex is running a common algorithm across all sensory modalities. That is to say that there is no fundamental difference in the cortical processing that goes into visual perception and other forms of perception like auditory or somatosensory(touch) perception. All sensory modalities are fed by the same back-end algorithm and if I were to tweak the appropriate inputs, I could use the visual cortex to “perceive”sound. Much of these insights are built on the pioneering work of Vernon Mountcastle. It’s easy to fathom that the implications of this assertion on AI are humongous.
At the centre of this framework is the cortical column, which, in Hawkins’s understanding, is the fundamental unit of computation. The book goes at length to lay down the blueprint for such an algorithm. Hawkins dedicates a substantial portion of the book proffering testable hypothesis aimed at, in the Popperian tradition, disproving his theory.
In line with his memory-prediction framework, Hawkins identifies certain key aspects of memory organization. Firstly, as experience will testify, memory comes in sequences of patterns. The memories of an event, specially those events that have an inherent temporal structure, say for example a song, are never recalled at random but are always organized in the same order as the original stimulus. In that sense, there is no such thing as a random recall. Secondly, memory has the curious property of invariant representation. The memory of a face, for example, is independent of the exact low-level details present at the time of initial stimulus. The memory of the face of your mother is invariant if I change its orientation by a few degrees or present it to you in different lighting conditions. Similarly the memory of a song remains the same if instead of the studio recording, I present the live version to you. Thirdly, memory has the interesting property that presenting even a small part of it unleashes a cascade of activity that results in the completion of the entire memory. For example, even a brief presentation of a song snippet causes you to retrieve the entire song involuntarily. Hawkins devotes much real estate to describe possible models of cortical architecture that account for the salient features of memory though a thorough reading might be required for understanding the nitty-gritty.
Along the way, Hawkins dispels naive notions of what truly intelligent AI would mean. Think prediction systems for weather, stock markets etc. and driver less cars instead of the emotionally labile humanoids as envisioned by popular science fiction.
A new wave of neural network research, buoyed by concurrent progress in neuroscience, is currently sweeping academia and industry with renewed promises of “truly intelligent” systems. These calls are all too familiar to historians of AI who will point to a chequered past, replete with instances of promising avenues that lead to dead ends. Its as if the AI revolution is destined to occur in fits and starts.