We have tried this approach to analyzing the collective activity of 1000 + neurons in the hippocampus, revealing scaling behaviors that are reproducible across animals, sometimes to the second decimal place. Renormalization group ideas from statistical physics offer other paths to simplification, reducing the dimensionality of models rather than the data itself. Īs an example, the scale invariant structure of natural images means that one can achieve a thousand-fold reduction of dimensionality with only a two-fold loss of variance, but this would miss important aspects of the natural image ensemble. It is proposed that differences between the performance of natural and artificial systems, e.g., the deep learning networks, are mainly due to the fact that recurrent processing permits exploitation of the temporal dimension for computation. These computations complement those realized in multilayer feed forward architectures and allow for the coexistence of rate and temporal codes. In addition, the network dynamics allow for the generation of stimulus specific response sequences (temporal codes) and the superposition of information provided by sequentially presented stimuli. Results obtained with parallel recordings of neuronal responses in cat and monkey visual cortex suggest that the cerebral cortex exploits this high dimensional dynamic space for the flexible encoding of relations among features (feature binding), for the acquisition and storage of information about statistical contingencies of features in the environment (priors), for the ultra-fast matching of priors with sensory evidence (predictive coding) and the classification of stimulus specific activity vectors by segregation in high dimensional space.

Such networks exhibit high dimensional non-linear dynamics that can be exploited for computations. The supra-granular layers of the cerebral cortex can be considered as a delay coupled recurrent network whose nodes are feature selective and have a propensity to oscillate.

Grounded mossy key