hierarchical autoencoders in time (hat) model  (Cortical Dynamics)

Journal: Neuron

Article Title: Constructing and Forgetting Temporal Context in the Human Cerebral Cortex

doi: 10.1016/j.neuron.2020.02.013

Figure Lengend Snippet: (A) HLI model schematic: the new state of each unit is a linear weighted sum of its old state and its new input. (B) HAT model schematic: each region maintains a representation of temporal context, which is combined with new input to form a simplified joint representation. (C) An AT unit, in which local context CNTX is updated via hidden representation HID and current input IN, modulated by time constant τ and “surprise” α. α is computed via auto-associative error Δ and a scaling parameter k. (D) In HAT, the input to level i is gated by surprise α from level (i-1). (E) HLI simulation of rSIDE:CE predicts longer alignment time at higher stages of processing. (F) HAT simulation of rSIDE:CE predicts longer alignment time at higher stages of processing. (G) Empirical rSIDE:CE results grouped by alignment time, consistent with predictions of both HLI and HAT. (H) HLI simulation of rSICD:CE predicts that regions that construct context slowly will also forget context slowly. (I) HAT simulations predict that the timescale of context separation (rSICD:CE) need not be slower in levels of the model with longer alignment times (rSIDE:CE). (J) Empirical rSICD:CE results grouped by alignment time. HLI = hierarchical linear integrator, HAT = hierarchical autoencoders in time, AT = autoencoder in time, rSI = intact-scramble ISPC.

Article Snippet: Finally, to account for the decoupling between alignment times and separation times in cortical dynamics, we proposed a hierarchical nonlinear integrator model – the hierarchical autoencoders in time (HAT) model. By combining non-linear integration and context gating mechanisms, HAT generated learning-dependent representations that account for the existing empirical phenomena (P1 and P2, above), while also exhibiting hierarchical alignment times, and a distinct set of timescales for alignment and separation.

Techniques: Construct

Journal: Neuron

Article Title: Constructing and Forgetting Temporal Context in the Human Cerebral Cortex

doi: 10.1016/j.neuron.2020.02.013

Figure Lengend Snippet: (A) HLI model schematic: the new state of each unit is a linear weighted sum of its old state and its new input. (B) HAT model schematic: each region maintains a representation of temporal context, which is combined with new input to form a simplified joint representation. (C) An AT unit, in which local context CNTX is updated via hidden representation HID and current input IN, modulated by time constant τ and “surprise” α. α is computed via auto-associative error Δ and a scaling parameter k. (D) In HAT, the input to level i is gated by surprise α from level (i-1). (E) HLI simulation of rSIDE:CE predicts longer alignment time at higher stages of processing. (F) HAT simulation of rSIDE:CE predicts longer alignment time at higher stages of processing. (G) Empirical rSIDE:CE results grouped by alignment time, consistent with predictions of both HLI and HAT. (H) HLI simulation of rSICD:CE predicts that regions that construct context slowly will also forget context slowly. (I) HAT simulations predict that the timescale of context separation (rSICD:CE) need not be slower in levels of the model with longer alignment times (rSIDE:CE). (J) Empirical rSICD:CE results grouped by alignment time. HLI = hierarchical linear integrator, HAT = hierarchical autoencoders in time, AT = autoencoder in time, rSI = intact-scramble ISPC.

Article Snippet: Gated integration using hierarchical autoencoders in time (HAT) The mismatch of alignment and separation times in cortical dynamics indicates that the integration rate is variable, consistent with the notion that temporal sequences are grouped into events, and that prior context is more rapidly forgotten at event boundaries ( Reynolds et al., 2007 ; Zacks and Tversky, 2001 ).

Techniques: Construct