Excitatory-Inhibitory Recurrent Dynamics Produce Robust Visual Grids and Stable Attractors

Are Grid-Like Representations a Component of All Perception and Cognition?

“…A fundamental building block of the cognitive map is the grid-like representation of objects in a high-dimensional multi-sensory space. In terms of the computational mechanism, emergent grid-like representations may arise from reinforcement learning (Stachenfeld et al, 2017 ), or from training recurrent neural networks (RNNs) on navigation or multiple normative tasks from supervised learning (Banino et al, 2018 ; Cueva and Wei, 2018 ; Sorscher et al, 2020 ; Zhang et al, 2022 ). Specifically, the grid-like representation is the eigenvector of the state-space transition matrix derived by the successor representation (SR) algorithm, and it is explained as a low-dimension sparse representation of the cognitive map.…”

“…Specifically, the grid-like representation is the eigenvector of the state-space transition matrix derived by the successor representation (SR) algorithm, and it is explained as a low-dimension sparse representation of the cognitive map. Whereas in RNNs, recurrent dynamics can generate stable ring or torus-like attractors that are associated with the grid patterns (Sorscher et al, 2020 ; Zhang et al, 2022 ). Although the existing computational theories of grid cells have focused on spatial navigation and path integration, conceptual analogies can be made between navigating in a Cartesian space and navigating in other physical or non-physical spaces.…”

“…In this theory, minicolumns would enable computers to represent and detect multidimensional features of objects, performing coincident voting and view-invariant recognition (Hinton, 2021 ). Future developments of theories and biologically realistic computational models that implement grid-like location-based computations across all sensory cortices or higher-order cortices would provide deeper insight into the mechanism (Shilnikov and Maurer, 2016 ; Cueva and Wei, 2018 ; Yu et al, 2021 ; Zhang et al, 2022 ).…”

“…Recent theories based on RNNs based on supervised learning also suggested alternative computational mechanisms (Cueva and Wei, 2018 ; Sorscher et al, 2020 ). Excitation-inhibition (E/I) balance and Dale's principle in synaptic connections may also provide additional biological constraints to understand the recurrent circuits for grid codes (Zhang et al, 2022 ).…”

“…Furthermore, theta oscillations have been found in both rat S1 and V2 areas, providing a source of speed and acceleration inputs. Head-direction signals reported in the rat S1 and V2 may receive indirect directional input from the RSC and postsubiculum (Taube, 2007 ), or may be derived from the sensorimotor input, such as the visual optical flow (Zhang et al, 2022 ). Additionally, the visual cortex also provides an important source of self-motion information to mEC grid cells; such multimodal signals may play a vital role in spatial perception (Campbell and Giocomo, 2018 ).…”