A Neural Network Model of Retrieval-Induced Forgetting
Kenneth A. Norman, Ehren L. Newman and Greg J. Detre
Princeton Technical Report #06-01

A Neural Network Model of Retrieval-Induced Forgetting (PDF)

Retrieval-induced forgetting (RIF) refers to the finding that retrieving a memory can impair subsequent recall of similar memories. Here, we present a new model of how the brain gives rise to RIF. The core of the model is a recently developed neural network learning algorithm (based on neural theta oscillations) that leverages regular oscillations in feedback inhibition to strengthen weak parts of target memories and to weaken competing memories. We use the model to address several puzzling findings from the RIF literature, including: why retrieval practice leads to more forgetting than simply presenting the target item; how RIF is affected by the strength of competing memories and the strength of the target (to-be-retrieved) memory; and why RIF sometimes generalizes to "independent cues", and sometimes does not.We also use the model to address non-monotonic effects of retrieval practice, whereby repeated practice first helps, then hurts recall of competing memories. For all of these questions, we show that the model can account for existing results, and we generate novel predictions regarding boundary conditions on these results.

How inhibitory oscillations can train neural networks and punish competitors
Kenneth A. Norman, Ehren Newman, Greg Detre
and Sean Polyn
Princeton Technical Report #05-01

This supersedes Technical Report #04-03(now removed from website). This manuscript is currently in press at Neural Computation.

How inhibitory oscillations can train neural networks and punish competitors (PDF)

We present a new learning algorithm that leverages oscillations in the strength of neural inhibition to train neural networks. Raising inhibition can be used to identify weak parts of target memories, which are then strengthened. Conversely, lowering inhibition can be used to identify competitors, which are then weakened. To update weights, we apply the Contrastive Hebbian Learning equation to successive time steps of the network. The sign of the weight change equation varies as a function of the phase of the inhibitory oscillation. We show that the learning algorithm can memorize large numbers of correlated input patterns without collapsing, and that it shows good generalization to test patterns that do not exactly match studied patterns.

The locus coeruleus, adaptive gain, and the optimization of simple decision tasks
Eric T. Brown, Mark S. Gilzenrat, and Jonathan D. Cohen
Princeton Technical Report #04-02

The locus coeruleus, adaptive gain, and the optimization of simple decision tasks (PDF)

We show that adaptive gain changes, by hypothesis mediated by the locus coeruleus (LC), can help optimize performance on simulated sensory discrimination tasks, even when no knowledge of stimulus timing is assumed. The metric of performance used here is the rate of correct responses (or 'reward rate') achieved by the simulated decision network. The primary model that we study has two layers: the first integrates sensory input directly and the second accumulates this filtered input (as well as noise from other brain areas) and translates it into motor responses via threshold crossing events. Gain transients occur after a physiologically-motivated delay following threshold crossings in the first layer - in this sense, gain schedules adapt according to accumulated sensory information. We adopt a linearization and reduction of the two layer model that allows a clear understanding of parameter effects and which is used to obtain a simpler set of optimization problems without sacrificing the generality of their solutions. By comparing optimal model reward rates in the presence of simulated LC-mediated gain changes with the (separately) optimized reward rates in the absence of such gain changes, the extent to which these gain changes contribute to enhanced task performance is determined, all for a 'standard parameter set' derived from fits to experiments. The results indicate that a significant improvement in reward (12-24%) is attributable to the LC-mediated gain mechanism. Additionally, the statistical variations in the optimal model gain transients from trial-to-trial agree with trends reported in recent experimental studies involving direct recordings from the LC. This provides converging evidence for the hypothesis that the LC plays a part in optimizing the dynamics of simple decision tasks.

Optimizing reward rate in two alternative choice tasks: Mathematical formalism
Jeff Moehlis, Eric Brown, Rafal Bogacz, Philip Holmes,
and Jonathan D. Cohen
Princeton Technical Report #04-01

Optimizing reward rate in two alternative choice tasks: Mathematical formalism (PDF)

In this report we collect and summarize mathematical results and formalism appropriate to describing one-dimensional drift-diffusion processes (stochastic ordinary differential equations) and related first passage and probability density evolution problems, governed by the backward and forward Kolmogorov (Fokker-Planck) equations respectively. We start by reviewing the Neyman-Pearson and Sequential Probability Ratio tests as optimal strategies for choosing between two alternative hypotheses in the presence of accumulating, noisy data. The continuum analog of both of these tests is a constant drift-diffusion process, and we give direct proofs of optimality with respect to reward rate of such a process in the broader class of Ornstein-Uhlenbeck processes, both in terms of first passages and density evolution. These correspond to the free response and interrogation protocols used in psychological testing. We end by considering the effects of variable gain on selected inputs to drift-diffusion and Ornstein-Uhlenbeck processes, and deriving optimal gain schedules for time-varying signal-to-noise ratios.