Page 16 - The 'X' Chronicles Newspaper - September 2023
P. 16
16 Brains May Predict The Future
To Make Sense of the
Present, Brains May
Predict The Future
A controversial theory suggests
that perception, motor control,
memory and other brain
functions all depend on
comparisons between ongoing
actual experiences and the
brain’s modeled expectations.
by Jordana Cepelewicz
Some neuroscientists favor a predictive
coding explanation for how the brain
works, in which perception may be
thought of as a “controlled
hallucination.” This theory emphasizes situation, it will be less surprised.” College London, a renowned
the brain’s expectations and predictions neuroscientist and one of the pioneers of
about reality rather than the direct Neuroscientists have long suspected that the predictive coding hypothesis.
sensory evidence that the brain receives. a similar mechanism drives how the brain
works. (Indeed, those speculations are Over the past decade, cognitive scientists,
In mid-2018, the artificial intelligence part of what inspired the GQN team to philosophers and psychologists have
company DeepMind introduced new pursue this approach.) According to this taken up predictive coding as a
software that can take a single image of a “predictive coding” theory, at each level compelling idea, especially for
few objects in a virtual room and, without of a cognitive process, the brain describing how perception works, but
human guidance, infer what the three- generates models, or beliefs, about what also as a more ambitious, all-
dimensional scene looks like from information it should be receiving from encompassing theory about what the
entirely new vantage points. Given just a the level below it. These beliefs get entire brain is doing. Experimental tools
handful of such pictures, the system, translated into predictions about what have only recently made it possible to
dubbed the Generative Query Network, should be experienced in a given start directly testing specific mechanisms
or GQN, can successfully model the situation, providing the best explanation of the hypothesis, and some papers
layout of a simple, video game-style of what’s out there so that the experience published in the past two years have
maze. will make sense. The predictions then get provided striking evidence for the theory.
sent down as feedback to lower-level Even so, it remains controversial, as is
There are obvious technological sensory regions of the brain. The brain perhaps best evidenced by a recent debate
applications for GQN, but it has also compares its predictions with the actual over whether some landmark results were
caught the eye of neuroscientists, who are sensory input it receives, “explaining replicable.
particularly interested in the training away” whatever differences, or
algorithm it uses to learn how to perform prediction errors, it can by using its Coffee, Cream and Dogs
its tasks. From the presented image, GQN internal models to determine likely
generates predictions about what a scene causes for the discrepancies. (For “I take coffee with cream and ____.” It
should look like — where objects should instance, we might have an internal seems only natural to fill in the blank
be located, how shadows should fall model of a table as a flat surface with “sugar.” That’s the instinct cognitive
against surfaces, which areas should be supported by four legs, but we can still scientists Marta Kutas and Steven
visible or hidden based on certain identify an object as a table even if Hillyard of the University of California,
perspectives — and uses the differences something else blocks half of it from San Diego, were banking on in 1980
between those predictions and its actual view.) when they performed a series of
observations to improve the accuracy of experiments in which they presented the
the predictions it will make in the future. The prediction errors that can’t be sentence to people, one word at a time on
“It was the difference between reality and explained away get passed up through a screen, and recorded their brain activity.
the prediction that enabled the updating connections to higher levels (as Only, instead of ending with “sugar,”
of the model,” said Ali Eslami, one of the “feedforward” signals, rather than when the last word popped into place, the
project’s leaders. feedback), where they’re considered sentence read: “I take coffee with cream
newsworthy, something for the system to and dog.”
According to Danilo Rezende, Eslami’s pay attention to and deal with
co-author and DeepMind colleague, “the accordingly. “The game is now about (Continued on Page 17)
algorithm changes the parameters of its adjusting the internal models, the brain
[predictive] model in such a way that dynamics, so as to suppress prediction CHECKOUT ALL THE GREAT XZBN
next time, when it encounters the same error,” said Karl Friston of University SHOWS AT WWW.XZBN.NET