Memory without feedback in a neural network

Abstract

Memory storage on short timescales is thought to be maintained by neuronal activity that persists after the remembered stimulus is removed. Although previous work suggested that positive feedback is necessary to maintain persistent activity, here it is demonstrated how neuronal responses can instead be maintained by a purely feedforward mechanism in which activity is passed sequentially through a chain of network states. This feedforward form of memory storage is shown to occur both in architecturally feedforward networks and in recurrent networks that nevertheless function in a feedforward manner. The networks can be tuned to be perfect integrators of their inputs or to reproduce the time-varying firing patterns observed during some working memory tasks but not easily reproduced by feedback-based attractor models. This work illustrates a mechanism for maintaining short-term memory in which both feedforward and feedback processes interact to govern network behavior.

Figure 1

Integration by a feedforward network. (A) Left, Feedforward network consisting of stages that linearly filter their inputs with time constant τ. Right, rearranged network with same output as the left network. (B) Integration of a pulse into a step. The network effectively decomposes the output into temporal basis functions (middle) reflecting the number of times the input was linearly filtered (n=1,8,15,22,… shown). When appropriately summed, the output is a unit-amplitude step for a duration ∼Nτ (N=100, τ=100 ms chosen here for illustration). (C) First three basis functions g_n and their sum. (D) Response to a pulse of input varies linearly with the input amplitude. (E) Integration of a step into a ramp by the same network.

Figure 2

Variants on the feedforward network architecture. (A) Top, Network in which the time scale τ of each stage reflects a mixture of intrinsic neuronal dynamics and positive feedback between locally connected clusters of neurons with shorter intrinsic time constant. Each neuron projected to all neurons in its own cluster and in the cluster ahead of it. The neurons in each stage of this network produced identical responses (middle) to those in a simplified network (effective network) consisting of a linear chain of neurons with time constant τ. (B) Network in which integration gets successively prolonged at each successive stage. Final output (thick black trace) is identical to the summed output of the network in panel A (black trace). Stages color-coded using legend of panel A. (C) Schematic of a more complicated feedforward network. Due to the multitude of pathways through the network, units exhibit a diversity of temporal responses (colored traces, 4 examples) that are composed of the same temporal basis functions as in the simpler networks and can be summed to yield perfect integration (bottom). Each stage (columns) consisted of 20 units. For all panels: N=20 stages, τ=500 ms.

Figure 3

Feedforward processing of inputs by a recurrent integrator network. (A) Processing by an architecturally feedforward network (top). Firing of each individual neuron can be alternatively viewed as firing of the activity pattern in which this neuron is active and others are silent (middle). In this view, connections between neurons are replaced by connections between patterns so that earlier patterns of activity trigger subsequent patterns of activity. Each pattern can be represented by a point in N dimensions (N=3 here) whose component along the i^th axis represents the i^th neuron's activity level in this pattern. (B) Amount of activity in each pattern, plotted across time, when a pulse of input stimulates the first activity pattern. Because the activity patterns correspond to a single neuron being active, this graph reproduces the progression of firing of the individual neurons. (C) Recurrent network that behaves analogously to the feedforward network of panel A. Activity propagates in a feedforward manner between orthogonal patterns (middle, bottom). (D) Amount of activity in each pattern is identical to that of panel A, reflecting the feedforward propagation through patterns.

Figure 4

Schur, but not eigenvector, decomposition reveals feedforward interactions between patterns of activity. (A-C) Top, Architecture of a network that generates persistent activity through positive feedback (A), a functionally feedforward network (B), and a network with a mixture of functionally feedforward and feedback interactions (C). Bottom, Neuronal responses when a pulse of input is given to neuron 1 (the green neuron). In B, the intrinsic neuronal decay time (black dashed) is shown for comparison. (D-F) Eigenvectors and eigenvalues of the corresponding networks. Top, the eigenvectors are indicated by colored dots and corresponding axes extending through these dots. The corresponding eigenvalues are indicated by the strength of a feedback loop of the eigenvector onto itself (no loop indicates an eigenvalue of zero). In E, the two eigenvectors perfectly overlap so there is only one distinct eigenvector. Bottom, Decomposition of the neuronal responses from panels A-C into their eigenvector components (see Fig. 5D). The responses cannot be decomposed in panel E because there is only a single eigenvector. In F, the neuronal responses from panel C are shown (dashed) for comparison. (G-I) Schur decomposition of the network activity. Top, Schur modes are indicated by colored dots and corresponding axes. The strengths of self-feedback (loops) or feedforward interactions are indicated next to the connections. Bottom, Activity in each Schur mode over time. Neuronal responses equal the sums and differences of these activities. See main text for details.

Figure 5

Comparison of the eigenvector and Schur decompositions. (A) Interactions between neurons (or, equivalently, patterns in which only a single neuron is active, top) are characterized by a connectivity matrix (bottom) whose entries w_ij describe how strongly the activity x_j of neuron j influences the activity x_i of neuron i. (B) The eigenvector decomposition finds activity patterns (top) that only interact with themselves. These self-feedback interactions are described by a diagonal matrix (bottom) whose elements along the diagonal (pink) are the eigenvalues. (C) The Schur decomposition finds orthogonal activity patterns (top) that can interact both with themselves (pink, with same self-feedback terms as in B) or in a feedforward manner (green, lower triangular entries give connection strengths from Schur pattern j to i where j<i). (D) Decomposition of an activity pattern consisting of only neuron 1 firing (black arrow) into single neuron states (left), non-orthogonal eigenmodes (middle), or orthogonal Schur modes (right). The decomposition shown corresponds to the network of Figure 4C. States are shown as solid dots that define the directions of the coordinate axes drawn through them.

Figure 6

Activity in a functionally feedforward network can outlast the slowest decaying eigenmode by a factor equal to the network size. (A) The slowest decaying eigenmode of a functionally feedforward network equals the intrinsic neuronal time constant τ=100 ms. (B) Neuronal firing rates of the same network exhibit activity for a time ∼Nτ=10 seconds that (C) can be summed to yield persistent neural activity over the same time scale.

Figure 7

Pulse-to-step integration for a line attractor versus a feedforward network. (A-D) In response to a brief pulse of input (top panels), neurons in a line attractor rapidly converge to a single temporal pattern of activity (middle panels) and therefore cannot be summed to give 2 seconds of constant-rate activity to within ±5% (dashed lines) unless the exponential decay is tuned to a much longer time scale (panel B, 20 sec decay). Small mistunings lead to rapid decay (A) or exponential growth (D) of both neuronal activity and summed output. (E-H) Performance of a feedforward network consisting of a chain of neurons. Due to the diversity of neuronal responses, a readout mechanism that sums these responses can produce activity that is maintained for 2 seconds even when the feedforward chain has larger mistunings of weights than was found for the line attractor networks. In all panels, mistuning is relative to a network (C, G) that performs perfect integration (up to a time ∼Nτ = 10 seconds for the feedforward network).

Figure 8

Generation of time-varying persistent neural activity by functionally feedforward versus positive-feedback models. (A) Trial-averaged, time-varying persistent neural activity of 5 neurons representing common classes of response observed in prefrontal cortical neurons during a visual working memory task (percentages of each class shown to the right of the plots). Adapted from Batuev (1994). (B) Fits to data in (A) attained by linearly summing the activities of neurons in networks with various connectivity patterns. Traces for the feedforward (cyan), functionally feedforward (purple), cyclic (mustard), and orthogonal (brown) networks overlap nearly exactly. For the functionally feedforward, Gaussian random (red), and random orthogonal networks, traces show the performance of the median of 100 randomly generated networks obeying the connectivity pattern. (C) Root mean squared error in the fits achieved by the median, 25^th and 75^th best performing network. (D) Fits to the persistent activity exhibited in the bottom panel of (A) when the networks were required to not only fit the delay period activity but also return to baseline activity at the end of the delay period. The functionally feedforward networks can stop responding at any time up to their maximal response duration by assigning zero readout weight to the later-occurring basis functions; the other networks cannot both fit the delay-period activity and return to baseline because they contain response modes that long outlast the delay period. (E) Examples of activity of individual neurons in the feedforward versus feedback-based networks when weights were increased by 2%. Left, feedforward versus cyclic networks. Right, functionally feedforward versus random attractor and random orthogonal networks. Exponential growth of the feedforward, but not feedback-based, networks is limited by the time taken for activity to propagate through the feedforward chain of activity patterns.

Figure 9

Activities of individual neurons for the networks tested in Figure 8. Left column, schematic of network connectivity for an unconnected network and the networks tested in Figure 8. Right column, example responses of 6 of the 100 neurons in each network. The unconnected network is shown to illustrate the intrinsic neuronal decay time constant τ=100 ms.