Dynamics and robustness of familiarity memory

When one is presented with an item or a face, one can sometimes have a sense of recognition without being able to recall where or when one has encountered it before. This sense of recognition is known as familiarity. Following previous computational models of familiarity memory we investigate the dynamical properties of familiarity discrimination, and contrast two different familiarity discriminators: one based on the energy of the neural network, and the other based on the time derivative of the energy. We show how the familiarity signal decays after a stimulus is presented, and examine the robustness of the familiarity discriminator in the presence of random fluctuations in neural activity. For both discriminators we establish, via a combined method of signal-to-noise ratio and mean field analysis, how the maximum number of successfully discriminated stimuli depends on the noise level.

💡 Research Summary

The paper investigates how the brain can generate a sense of familiarity—recognizing that an item has been encountered before without recalling the specific context. Building on earlier computational models of familiarity memory, the authors examine two distinct discriminators within a Hopfield‑type recurrent neural network. The first discriminator uses the global energy of the network (E) as a familiarity signal: when a stored pattern is presented, the network’s energy drops toward a lower attractor, and the magnitude of this drop distinguishes familiar from novel stimuli. The second discriminator relies on the time derivative of the energy (dE/dt), capturing the rapid initial change in energy that occurs immediately after stimulus presentation. This derivative‑based signal peaks within a few update steps and then quickly vanishes, offering a fast but potentially noise‑sensitive cue.

To study the temporal dynamics, the authors simulate the network’s response to both familiar and novel inputs. They find that the energy signal decays gradually over tens of iterations, whereas the dE/dt signal exhibits a sharp transient within the first few iterations. This difference mirrors empirical observations that familiarity judgments can be both rapid (a “gut feeling”) and more deliberative.

The robustness of each discriminator is evaluated under stochastic perturbations that model random neural activity. Noise is introduced as independent Gaussian fluctuations added to each neuron’s input, with standard deviation σ. Using mean‑field theory, the authors derive analytical expressions for the expected energy ⟨E⟩ and its variance, as well as for ⟨dE/dt⟩ and its variance, in the presence of noise. They then define signal‑to‑noise ratios (SNRs):

• SNR_E = ΔE / √Var(E)
• SNR_Δ = Δ(dE/dt) / √Var(dE/dt)

where ΔE and Δ(dE/dt) denote the mean differences between familiar and novel stimuli. A discriminator is considered successful when its SNR exceeds one. By varying σ, the authors compute the maximal number of patterns that can be reliably discriminated (capacity) for each discriminator.

The results reveal complementary regimes. At low noise levels (σ below a critical threshold σ_c), the derivative‑based discriminator achieves a higher SNR because the initial energy drop is large and relatively unaffected by small fluctuations; consequently, its capacity exceeds that of the energy‑based discriminator. As noise increases beyond σ_c, the transient peak becomes obscured, the SNR_Δ falls sharply, and the energy‑based discriminator, whose signal integrates over many iterations, retains a more stable SNR_E. In this high‑noise regime, the energy‑based capacity surpasses the derivative‑based one.

Recognizing the complementary strengths, the authors propose a hybrid strategy: use the dE/dt signal for rapid decisions when its SNR is above a confidence threshold, and fall back on the slower energy‑based assessment when the derivative signal is ambiguous. Simulations of the hybrid model demonstrate a broader noise tolerance and higher overall capacity than either method alone.

The paper concludes by linking these theoretical findings to behavioral data. Human familiarity judgments often combine fast, low‑confidence impressions with slower, more accurate assessments—a pattern consistent with the dual‑discriminator framework. Moreover, conditions that increase neural noise, such as aging or neurological disorders, may shift reliance from the fast derivative cue to the more robust energy cue, explaining observed declines in familiarity performance.

Overall, the study provides a rigorous dynamical and statistical analysis of familiarity memory, quantifies how noise impacts two plausible neural read‑outs, and offers predictions that can guide future neurophysiological experiments and the design of artificial systems that emulate human‑like familiarity detection.