Prediction is a powerful principle in neuroscience, and it is not a new one. It has been central to interpretation of brain function since the influential work of E. N. Sokolov (Reference Sokolov and Brazier1960) (see target article, Note 28). He found that cortical responses depend not on the amplitude of an incoming signal, but on its information value. An expected stimulus caused hardly a ripple, while an unexpected one triggered what Sokolov termed an orienting response. The key experiment was to repeat a stimulus until its cortical signal nearly disappeared (habituation of the orienting response, or Clark's “repetition suppression”). Then Sokolov decreased the stimulus amplitude or its duration. Sokolov reasoned that if the cortex were merely echoing stimulus properties the response should have decreased, but instead it increased. With a qualitative change, no amount of fussing with nonlinearities and thresholds could explain the result. The cortex was coding not stimulus properties but stimulus information, the difference between signal and expectation. In this context it is no wonder that we ignore and fail to remember most of the vast streams of signals emanating from our millions of sensory receptors. So Clark's prediction thesis has been the dominant interpretation of cortical sensory coding for more than a half-century.
Another insight that shaped neuroscience is that the brain is not about representing the stimulus; it is about organizing action. The evidence begins with an anatomical paradox that the precentral “motor” cortex is innervated by the dorsal thalamus, a region homologous to the dorsal spinal cord that processes sensory information (Pribram Reference Pribram1971, p. 241). Pribram asks why the motor cortex should be closely tied to an otherwise sensory structure. His answer is that the motor cortex is really a sensory cortex for an image of achievement, analogous to the images in sensory regions and organized similarly. Motor cortex codes environmental contingencies, not literal muscle movements, and continuously compares progress in execution of an act with its goal.
Similarly, it has long been known that receptive fields in sensory cortex are shaped not only by anatomy but also by experience, so that they encode best what is predicted to be present in the environment. I was privileged to witness the first evidence that sensory experience could tune the receptive field properties of the primary visual cortex (V1). Helmut Hirsch, then a Stanford graduate student, was studying kittens that he raised wearing masks that exposed one eye to vertical stripes and the other to horizontal stripes. Together with Nico Spinelli and Robert Phelps we began recording from single cells in V1 of the mask-reared kittens, using the first automated receptive-field mapping apparatus. We prepared our first kitten and dipped our microelectrode into its cortex.
The first cells we recorded had large, poorly defined receptive fields of the sort to expect in a visually deprived cat. Then around 10 p.m. we found a cell with a huge, vertically oriented receptive field. Perhaps it was an artifact, the bursting discharges of an injured cell as the mapping stimulus swept vertically across our screen. So we changed to a horizontal scan. The field remained, five times bigger than any oriented receptive field ever recorded from a cat. Our jaws dropped as we looked at each other, a moment of discovery – this wasn't a normal cortex, but something completely different. It was the magical moment in science when you know something about nature that no one else knows. We covered one eye, then the other; the receptive field disappeared and reappeared. Later that night we recorded several other similar fields, all vertical or horizontal, all monocular, and all huge. It turned out later that the receptive field orientations matched the mask orientations for the corresponding eye (Hirsch & Spinelli Reference Hirsch and Spinelli1970). Plasticity in this cat's cortex extended beyond any mere selection of normal receptive fields, beyond anything that anyone had suspected. The cat had reorganized its cortex from visual experience alone. Clearly the cortex, by the structure of its receptive fields, was predicting future input.
This would be an interesting curiosity if not for its under-appreciated implication that the same process must be occurring in normal cats, and, by extension, in humans as well. Sensory receptive fields are tuned to the structure of the world that the animal encounters in its early experience. The receptive fields of normal animals have a 1/f statistical structure, as does the natural world.
It is even possible that the dominance of the foveal projection onto V1, a quarter of the entire surface in humans, is a consequence of the huge number of projections coming up from the periphery. The small size of V1 receptive fields representing the fovea might originate from the better optics and smaller convergence of the foveal anatomy. The distribution of receptive field orientations and spatial frequencies reflects the properties of the normal visual environment (Switkes et al. Reference Switkes, Mayer and Sloan1978); the cortex is predicting its own input by its very structure. This is precisely what Clark realizes when he concludes, “dig a little deeper and what we discover is a model of key aspects of neural functioning that makes structuring our worlds genuinely continuous with structuring our brains” (sect. 3.4, para. 1). But the evidence has been there all along.
You have
Access
Prediction is a powerful principle in neuroscience, and it is not a new one. It has been central to interpretation of brain function since the influential work of E. N. Sokolov (Reference Sokolov and Brazier1960) (see target article, Note 28). He found that cortical responses depend not on the amplitude of an incoming signal, but on its information value. An expected stimulus caused hardly a ripple, while an unexpected one triggered what Sokolov termed an orienting response. The key experiment was to repeat a stimulus until its cortical signal nearly disappeared (habituation of the orienting response, or Clark's “repetition suppression”). Then Sokolov decreased the stimulus amplitude or its duration. Sokolov reasoned that if the cortex were merely echoing stimulus properties the response should have decreased, but instead it increased. With a qualitative change, no amount of fussing with nonlinearities and thresholds could explain the result. The cortex was coding not stimulus properties but stimulus information, the difference between signal and expectation. In this context it is no wonder that we ignore and fail to remember most of the vast streams of signals emanating from our millions of sensory receptors. So Clark's prediction thesis has been the dominant interpretation of cortical sensory coding for more than a half-century.
Another insight that shaped neuroscience is that the brain is not about representing the stimulus; it is about organizing action. The evidence begins with an anatomical paradox that the precentral “motor” cortex is innervated by the dorsal thalamus, a region homologous to the dorsal spinal cord that processes sensory information (Pribram Reference Pribram1971, p. 241). Pribram asks why the motor cortex should be closely tied to an otherwise sensory structure. His answer is that the motor cortex is really a sensory cortex for an image of achievement, analogous to the images in sensory regions and organized similarly. Motor cortex codes environmental contingencies, not literal muscle movements, and continuously compares progress in execution of an act with its goal.
Similarly, it has long been known that receptive fields in sensory cortex are shaped not only by anatomy but also by experience, so that they encode best what is predicted to be present in the environment. I was privileged to witness the first evidence that sensory experience could tune the receptive field properties of the primary visual cortex (V1). Helmut Hirsch, then a Stanford graduate student, was studying kittens that he raised wearing masks that exposed one eye to vertical stripes and the other to horizontal stripes. Together with Nico Spinelli and Robert Phelps we began recording from single cells in V1 of the mask-reared kittens, using the first automated receptive-field mapping apparatus. We prepared our first kitten and dipped our microelectrode into its cortex.
The first cells we recorded had large, poorly defined receptive fields of the sort to expect in a visually deprived cat. Then around 10 p.m. we found a cell with a huge, vertically oriented receptive field. Perhaps it was an artifact, the bursting discharges of an injured cell as the mapping stimulus swept vertically across our screen. So we changed to a horizontal scan. The field remained, five times bigger than any oriented receptive field ever recorded from a cat. Our jaws dropped as we looked at each other, a moment of discovery – this wasn't a normal cortex, but something completely different. It was the magical moment in science when you know something about nature that no one else knows. We covered one eye, then the other; the receptive field disappeared and reappeared. Later that night we recorded several other similar fields, all vertical or horizontal, all monocular, and all huge. It turned out later that the receptive field orientations matched the mask orientations for the corresponding eye (Hirsch & Spinelli Reference Hirsch and Spinelli1970). Plasticity in this cat's cortex extended beyond any mere selection of normal receptive fields, beyond anything that anyone had suspected. The cat had reorganized its cortex from visual experience alone. Clearly the cortex, by the structure of its receptive fields, was predicting future input.
This would be an interesting curiosity if not for its under-appreciated implication that the same process must be occurring in normal cats, and, by extension, in humans as well. Sensory receptive fields are tuned to the structure of the world that the animal encounters in its early experience. The receptive fields of normal animals have a 1/f statistical structure, as does the natural world.
It is even possible that the dominance of the foveal projection onto V1, a quarter of the entire surface in humans, is a consequence of the huge number of projections coming up from the periphery. The small size of V1 receptive fields representing the fovea might originate from the better optics and smaller convergence of the foveal anatomy. The distribution of receptive field orientations and spatial frequencies reflects the properties of the normal visual environment (Switkes et al. Reference Switkes, Mayer and Sloan1978); the cortex is predicting its own input by its very structure. This is precisely what Clark realizes when he concludes, “dig a little deeper and what we discover is a model of key aspects of neural functioning that makes structuring our worlds genuinely continuous with structuring our brains” (sect. 3.4, para. 1). But the evidence has been there all along.