Prediction: multi-scale pattern completion of the future

The notion of the brain as a prediction machine has been extremely influential and productive in cognitive sciences.
One prominent framework is of a “Bayesian brain” that explicitly generates predictions and uses resultant errors to guide adaptation.
The prediction-generation component of this framework may involve little more than a pattern completion process.
Brain-like systems can get prediction “for free,” without the need to posit formal logical representations with Bayesian probabilities or an inference machine that holds them in working memory.

Several fields in the cognitive and neural sciences have recently been promoting the idea that prediction of sensory input is a fundamental aspect of many cognitive and perceptual processes. For example, work in real-time language comprehension suggests that readers and listeners tend to develop anticipations for certain categories of upcoming words. In fact,  the recent prediction-based language learning model, GPT-2, predicts the next word in 720 Wikipedia articles with impressive accuracy (e.g., very low surprisal values). It also mimics human data of neural activation patterns quite accurately for words that are less surprising or more surprising. Comparison of various language models suggests that, since predictive models (such as GPT-2) fit neural data better than non-predictive models, prediction may indeed be a fundamental mechanism of language processing and language learning.

Visual pattern completion process entails additional time to achieve recognition (perhaps an additional 90 ms). This additional time is not treated as a prediction of the future but instead simply as a temporally drawn-out process of “filling-in” the gaps in the original visual stimulus.
The perceptual effects of filling-in those gaps that are present in a retinal image have often been treated as the result of a pattern completion process whereby the neurons that form the population code for a particular stimulus (e.g., a distributed cell ensemble) pass recurrent activation back and forth among each other to eventually activate even those neurons that were not initially responding because they code for aspects of the stimulus that are currently absent (or occluded) in the visual input. This filling-in process is strikingly apparent in the Kanizsa Triangle illusion.

The activity of the network is not literally an encoding of predictions or prediction errors. Rather than exhibiting “weak anticipation” by generating predictions of future input that are labeled as such, this homeostatic reservoir network does “strong anticipation” by coupling with its environment instead of representationally modeling it. This illustrates the fact that systems can be implicitly anticipatory, without forming explicit predictions.

After training on sequences of inputs generated from a simple, probabilistic grammar, this network learns to balance its internal propagation of activity with external perturbations, and anticipation of future input naturally emerges. After training, when the network receives only a partial sequence as input, the internal activity of the network nonetheless produces a pattern that is quite similar to the one we would see if the network had received an appropriate input – as if the network “knew” what was coming next.
The network also exhibits an increase in activity in response to lower-probability sequences, and sequences that were never presented in training, which is often taken as a signature pattern of predictive coding.
Importantly, neither prediction nor memory were explicit features of this network. The only mechanism driving this emergent anticipation is a few basic rules for maintaining homeostasis.

The predictiveness of pattern completion is generalizable across a range of cognitive and perceptual processes. The pattern completion paradigm is easily applicable to many spatiotemporal scales of analysis, whether at the scale of phonemes or words or sentences, or at the scale of visual features or objects or scenes.
The notion of multi-scale pattern completion extends our understanding of prediction by showing that prediction is understanding, extended through time.