Open problems left by the Cooper–Liberman–Oja (1979) threshold-passive-modification theory of neuron specificity in visual cortex: the labile/non-labile synapse substrate and the uncompleted population-level mathematics of acquisition, loss, and recovery

In 1979 Cooper, Liberman and Oja proposed that visual cortical neurons acquire, lose, and regain orientation specificity through threshold passive modification of labile lateral-geniculate-to-cortex synapses, while non-labile synapses carry permanent information. The theory's theorems were linearized mean-convergence statements, and its deprivation predictions hinged on unmeasured correlations between the noise impinging on the two synapse classes. The successor BCM (Bienenstock–Cooper–Munro) theory and its experimental program (sliding-threshold bidirectional plasticity, binocular interaction, and measurements of lateral geniculate activity under deprivation) resolved the binocular-interaction issue, the biophysical basis of threshold-dependent modification, and the input-statistics question within the BCM formulation, so these items fall outside the open core considered here. What no later work has addressed survives as the open core: (1) the biological substrate of the 1979-specific commitments — the dichotomy between labile and non-labile geniculocortical synapses and the intracellular signaling of integrated subthreshold postsynaptic state to labile synapses; (2) the mathematical completion of the 1979 dynamics beyond linearization, iid presentation, and small learning gains — the population distribution of asymptotic tuning, finite-sample probabilities of deprivation-induced loss or switching of the preferred pattern, and the robustness of the correlated-deprivation prediction $m\to -z$; and (3) the empirical measurement of the input-channel correlations needed to test that prediction. The core is open in the literature's own terms but bound to a superseded framework, so its significance is chiefly model-completion and historical, with indirect relevance to synapse-class-specific plasticity.

Background

Primary visual cortex (V1) contains neurons that respond selectively to the orientation of visual stimuli; its main feed-forward input arrives from the lateral geniculate nucleus (LGN), the thalamic relay of retinal signals. Classical kitten experiments showed that this orientation specificity is experience-dependent: dark rearing (raising the animal with no patterned visual input) or eyelid suture (closing one eyelid, which removes patterned vision from that eye while admitting only diffuse light) during the critical period (the early developmental window in which visual experience lastingly shapes cortical responses) degrades orientation selectivity, and specificity can be regained when patterned vision returns (Buisseret and Imbert, 1976; Frégnac and Imbert, 1978). Hebbian learning rules (synapses strengthen when pre- and postsynaptic activity are correlated) could explain how patterned input sharpens selectivity, but early self-organizing models had no mechanism for the loss and recovery of specificity under deprivation.

Cooper, Liberman and Oja (1979) addressed this with a model in which each cortical neuron receives two parallel LGN inputs: pattern components $d^{k}$ through labile (modifiable) synapses with strength vector $m(t)$, and components $b^{k}$ through non-labile (fixed) synapses with strength vector $z$ that carry permanent innate information and provide the initial tuning bias that enables recovery. The pre-threshold (subthreshold) postsynaptic potential for pattern class $k$ is $c(t)=(m(t),d^{k}(t))+(z,b^{k}(t))+x(t)$, where $x(t)$ is channel noise and $(\cdot,\cdot)$ denotes an inner product; the inputs are noisy, $d^{k}(t)=d^{k}+r(t)$ and $b^{k}(t)=b^{k}+s(t)$, with $r(t)$ and $s(t)$ the noise on the labile and non-labile channels.

The synapses modify by threshold passive modification, a piecewise rule on the labile vector: when the pre-threshold potential lies between the modification threshold $\theta_M$ and the saturation level $\mu$, the synapse strengthens, $\Delta m=\eta^{+}(\mu-c)\,d$; when it lies below $\theta_M$, the synapse weakens, $\Delta m=-\eta^{-}c\,d$; above $\mu$ only passive forgetting by the factor $\gamma$ acts. The distinctive feature is the weakening branch: integrated subthreshold postsynaptic state — input that drives the cell without producing output firing — suffices to weaken labile synapses, while the fixed synapses provide a permanent scaffold. The paper's theorems are mean-convergence and bounded-variance statements obtained under strong approximations: the modification branch is assumed fixed per pattern class (linearized branch selection), pattern classes are presented independently and uniformly at random (independent, identically distributed — i.i.d. — sampling), learning gains $\eta^{\pm}$ are small, and $\gamma$ is near 1. The main analytic deprivation result is that under noise-only input the mean labile vector drifts toward $m\to -z$ when the channel noises $r$ and $s$ are strongly correlated (the labile synapses learn to cancel the innate contribution, collapsing the response), whereas for uncorrelated noise the tuning curve (the neuron's response amplitude as a function of stimulus orientation) broadens while remaining dominated by $z$. Single-run simulations illustrated acquisition, loss, and recovery of a sharp tuning curve.

The 1979 paper acknowledged its open issues: the cellular mechanisms realizing threshold passive modification (the intracellular signal conveying subthreshold state to synapses, and the substrate separating labile from non-labile junctions) were unspecified; binocular interactions and realistic lateral circuitry were omitted; the statistical correlations among $r(t)$, $s(t)$ and the channel noise, on which the deprivation predictions critically depend, were unknown; the theorems did not describe the distribution of tuning across a population of neurons or give finite-sample probabilities of which cells lose or switch their preferred pattern after deprivation; and the linearized, iid, small-gain approximations were unquantified.

The theory was immediately superseded by the authors' own BCM theory (Bienenstock, Cooper and Munro, 1982), which restructured it: the fixed threshold became a sliding modification threshold depending on a time-averaged postsynaptic activity (metaplasticity — the plasticity rule itself changes with activity history), the labile/non-labile synapse dichotomy was dropped, and binocular interaction and ocular dominance (the relative weighting of the two eyes' inputs in a cortical neuron's response) were treated directly, with a convergence result to states of maximum selectivity. The modification-threshold concept then acquired an experimental and biophysical basis: bidirectional long-term potentiation and depression (LTP and LTD — persistent strengthening and weakening of synaptic efficacy) dependent on N-methyl-D-aspartate (NMDA) receptors (a subtype of glutamate receptor whose calcium entry triggers these modifications) were demonstrated in visual cortex with the predicted frequency dependence (Bear, Cooper and Ebner, 1987; Kirkwood et al., 1993; Kirkwood and Bear, 1994); monocular deprivation (depriving one eye of patterned vision, for example by eyelid suture) was shown to induce homosynaptic LTD (weakening confined to the synapses that were themselves active) in vivo (Rittenhouse et al., 1999); and a calcium-based model unified the induction protocols for the modification function and its threshold (Shouval, Bear and Cooper, 2002). Cooper and Bear (2012) review this program as a successful theory–experiment interaction.

What the successor lineage did not carry over defines the remaining open problem: no later work identified or ruled out the 1979 labile/non-labile synapse-class substrate or the subthreshold-integrated signaling mechanism; no later work analyzed the 1979 recursion itself beyond its linearized mean results (population distributions, finite-sample switching probabilities, correlated non-iid inputs); and the input-correlation question was answered only in the BCM framing, through measurements of LGN activity under lid suture versus retinal inactivation (Linden et al., 2009) and analyses of how correlated LGN firing interacts with different plasticity rules (Blais, Cooper and Shouval, 2009; Blais et al., 2008).

Problem Statement

Complete and make testable the specific commitments of the 1979 Cooper–Liberman–Oja theory that its successor lineage left untouched, so that the theory's schematic account of acquisition, loss, and recovery of visual cortical neuron specificity becomes a completed and testable account. The requested results serve this one objective and are scientifically linked: the substrate identification makes the model's constructs measurable, the mathematical completion makes its predictions statistical rather than anecdotal, and the input-statistics measurement decides between its deprivation outcomes. Item (1)(a) and item (1)(b) are independent sub-claims, so that resolving one does not settle the other; item (2) comprises the four results (a)–(d); and item (3) is conditional on item (1) as stated. Concretely:

(1) Biological substrate. Identify, or rule out, (a) a biophysical basis for the assumed dichotomy between labile and non-labile geniculocortical synapses on the same visual cortical neuron — the class whose strengths follow the modification rule versus the class carrying permanent information — and (b) an intracellular mechanism that conveys integrated subthreshold postsynaptic state to the labile synapses, implementing weakening when the pre-threshold potential is below the modification threshold $\theta_M$ even in the absence of output firing.

(2) Mathematical completion of the 1979 dynamics. For the threshold-passive-modification recursion — labile vector $m(t)$, fixed vector $z$, noisy inputs $d^{k}(t)=d^{k}+r(t)$ and $b^{k}(t)=b^{k}+s(t)$, channel noise $x(t)$, pre-threshold potential $c(t)=(m(t),d^{k}(t))+(z,b^{k}(t))+x(t)$, strengthening $\Delta m=\eta^{+}(\mu-c)\,d$ for $\theta_M\le c<\mu$, weakening $\Delta m=-\eta^{-}c\,d$ for $c<\theta_M$, decay-only above $\mu$, and forgetting factor $\gamma$ — characterize the recursion's long-run behavior over the regime class defined by relaxing the paper's approximations: branch selection not fixed per pattern class, pattern presentation correlated rather than independent and identically distributed, gains not small, and $\gamma$ not necessarily near 1. That list delimits the class of regimes over which the behavior is to be characterized; it is not a set of conditions that must all be handled within a single result. Supplying the required result types (a)–(d) below:

$$\Delta m=\begin{cases}\eta^{+}\big(\mu-c(t)\big)\,d(t), & \theta_M\le c(t)<\mu\\ -\eta^{-}\,c(t)\,d(t), & c(t)<\theta_M\\ 0 \ \text{(decay only)}, & c(t)\ge\mu\end{cases}.$$

In this display, $d(t)$ and $c(t)$ denote the noisy input $d^{k}(t)=d^{k}+r(t)$ and the resulting pre-threshold potential of the pattern class presented at time $t$, and the third branch reflects the Background description that above $\mu$ only passive forgetting by the factor $\gamma$ acts. The authoritative recursion is that of Cooper–Liberman–Oja (1979), in which the forgetting factor $\gamma$ multiplies the labile vector at every step, $m(t)=\gamma\,m(t-1)+\Delta m(t)$, so that above $\mu$ the vector decays by $\gamma$ while no increment is added; 'arbitrary $\gamma$ and $\eta^{\pm}$' in (c) ranges over the forgetting and gain parameters of this rule.

The requested results are: (a) the population distribution of asymptotic tuning — where a neuron's asymptotic tuning is its asymptotic mean pre-threshold response as a function of the presented pattern class, its preferred pattern class being the maximizing class, and the distribution is taken over a specified population ensemble of innate vectors $z$, initial conditions, and input wiring — not only single-neuron mean convergence; (b) finite-sample probability bounds for the loss or switching of the preferred pattern class after a deprivation epoch and for its recovery — which cells lose specificity, which switch preference, and with what probabilities, replacing the published single-run simulations; (c) the behavior under correlated, non-iid stimulus sequences and arbitrary $\gamma$ and $\eta^{\pm}$; and (d) a determination of whether the correlated-deprivation prediction — $m\to -z$ with collapse of the total response when the channel noises $r$ and $s$ are strongly correlated, versus broadened tuning dominated by $z$ when they are not — is a robust property of the rule or an artifact of the simulated parameter regime (including the phase-dependent weakening gain of the paper's own simulations, which is larger during the deprivation phase than in the patterned acquisition and recovery phases).

(3) Empirical test. Measure the statistical structure (means, correlations, spatial and temporal structure) of LGN-to-cortex afferent (incoming input) activity under deprivation conditions relative to the model's two input channels, sufficient to decide between the theory's deprivation outcomes. This item is conditional on (1): if (1) identifies the substrate, fix the mapping from the model constructs (labile and non-labile channels, channel noise) to measurable afferent and synapse classes and specify the measurement of their statistical structure; if (1) rules the dichotomy out, state instead the measurement that would be needed to test the correlated-deprivation prediction in the resulting model.

Scope note: binocular interaction and ocular dominance, the biophysical basis of sliding-threshold (BCM-type) modification, and LGN correlation measurements in the BCM framework are excluded from this problem because the successor lineage resolved them (Bienenstock, Cooper and Munro, 1982; Bear, Cooper and Ebner, 1987; Shouval, Bear and Cooper, 2002; Linden et al., 2009); no other restriction of the theory's stated open problems is intended.

References

1. Leon N. Cooper, Fishel Liberman, and Erkki Oja (1979). A theory for the acquisition and loss of neuron specificity in visual cortex. Biological Cybernetics 33, 9–28. DOI 10.1007/BF00337414. https://doi.org/10.1007/BF00337414
2. P. Buisseret and M. Imbert (1976). Visual cortical cells: their developmental properties in normal and dark reared kittens. The Journal of Physiology 255, 511–525. DOI 10.1113/jphysiol.1976.sp011293. https://doi.org/10.1113/jphysiol.1976.sp011293
3. Y. Frégnac and M. Imbert (1978). Early development of visual cortical cells in normal and dark-reared kittens: relationship between orientation selectivity and ocular dominance. The Journal of Physiology 278, 27–44. DOI 10.1113/jphysiol.1978.sp012290. https://doi.org/10.1113/jphysiol.1978.sp012290
4. Elie L. Bienenstock, Leon N. Cooper, and Paul W. Munro (1982). Theory for the development of neuron selectivity: orientation specificity and binocular interaction in visual cortex. The Journal of Neuroscience 2, 32–48. DOI 10.1523/JNEUROSCI.02-01-00032.1982. https://doi.org/10.1523/JNEUROSCI.02-01-00032.1982
5. Mark F. Bear, Leon N. Cooper, and Ford F. Ebner (1987). A physiological basis for a theory of synapse modification. Science 237, 42–48. DOI 10.1126/science.3037696. https://doi.org/10.1126/science.3037696
6. A. Kirkwood, S. M. Dudek, J. T. Gold, C. D. Aizenman, and M. F. Bear (1993). Common forms of synaptic plasticity in the hippocampus and neocortex in vitro. Science 260, 1518–1521. DOI 10.1126/science.8502997. https://doi.org/10.1126/science.8502997
7. A. Kirkwood and M. F. Bear (1994). Homosynaptic long-term depression in the visual cortex. The Journal of Neuroscience 14, 3404–3412. DOI 10.1523/JNEUROSCI.14-05-03404.1994. https://doi.org/10.1523/JNEUROSCI.14-05-03404.1994
8. Nathan Intrator and Leon N. Cooper (1992). Objective function formulation of the BCM theory of visual cortical plasticity: statistical connections, stability conditions. Neural Networks 5, 3–17. DOI 10.1016/S0893-6080(05)80003-6. https://doi.org/10.1016/S0893-6080(05)80003-6
9. E. E. Clothiaux, M. F. Bear, and L. N. Cooper (1991). Synaptic plasticity in visual cortex: comparison of theory with experiment. Journal of Neurophysiology 66, 1785–1804. DOI 10.1152/jn.1991.66.5.1785. https://doi.org/10.1152/jn.1991.66.5.1785
10. Cynthia D. Rittenhouse, Harel Z. Shouval, Michael A. Paradiso, and Mark F. Bear (1999). Monocular deprivation induces homosynaptic long-term depression in visual cortex. Nature 397, 347–350. DOI 10.1038/16922. https://doi.org/10.1038/16922
11. Harel Z. Shouval, Mark F. Bear, and Leon N. Cooper (2002). A unified model of NMDA receptor-dependent bidirectional synaptic plasticity. Proceedings of the National Academy of Sciences USA 99, 10831–10836. DOI 10.1073/pnas.152343099. https://doi.org/10.1073/pnas.152343099
12. B. S. Blais, M. Y. Frenkel, S. R. Kuindersma, R. Muhammad, H. Z. Shouval, L. N. Cooper, and M. F. Bear (2008). Recovery from monocular deprivation using binocular deprivation. Journal of Neurophysiology 100, 2217–2224. DOI 10.1152/jn.90411.2008. https://doi.org/10.1152/jn.90411.2008
13. M. L. Linden, A. J. Heynen, R. H. Haslinger, and M. F. Bear (2009). Thalamic activity that drives visual cortical plasticity. Nature Neuroscience 12, 390–392. DOI 10.1038/nn.2284. https://doi.org/10.1038/nn.2284
14. B. S. Blais, L. N. Cooper, and H. Z. Shouval (2009). Effect of correlated lateral geniculate nucleus firing rates on predictions for monocular eye closure versus monocular retinal inactivation. Physical Review E 80, 061915. DOI 10.1103/PhysRevE.80.061915. https://doi.org/10.1103/PhysRevE.80.061915
15. Leon N. Cooper and Mark F. Bear (2012). The BCM theory of synapse modification at 30: interaction of theory with experiment. Nature Reviews Neuroscience 13, 798–810. DOI 10.1038/nrn3353. https://doi.org/10.1038/nrn3353
16. Robert E. Soodak (1994). Simulation of visual cortex development under lid-suture conditions: enhancement of response specificity by a reverse-Hebb rule in the absence of spatially patterned input. Biological Cybernetics 70, 303–309. DOI 10.1007/BF00200328. https://doi.org/10.1007/BF00200328
17. Julian Bartram, Martin C. Kahn, Simon Tuohy, Ole Paulsen, Tony Wilson, and Edward O. Mann (2017). Cortical Up states induce the selective weakening of subthreshold synaptic inputs. Nature Communications 8, 665. DOI 10.1038/s41467-017-00748-5. https://doi.org/10.1038/s41467-017-00748-5