Autonomous confusion-gated learning in asymmetric neural networks: the unfinished verification of Parisi's 1986 self-organizing memory

In 1986 G. Parisi proposed an associative-memory network with strongly asymmetric synaptic couplings whose learning rule applies Hebbian reinforcement with exponential decay only when time-averaged neuronal correlations exceed a threshold, relying on the chaotic, time-dependent character of asymmetric networks' 'confused' (non-retrieval) states to switch learning off. The paper verified only the premise that random asymmetric networks wander chaotically, and explicitly left the central question open: whether the gated architecture actually self-organizes a sustained autonomous cycle of learning, retrieval, forgetting and re-learning — including the automatic transition from a retrieval state back into confusion — even from a blank memory, and for which choices of threshold, averaging window and forgetting rate, and which biologically plausible realizations of the asymmetry. No work known to date establishes or refutes the fully self-gated autonomous cycle for Parisi's rule, and no criteria for its threshold, averaging window and forgetting rate exist.

Background

An attractor (associative-memory) network is a system of interacting neuron-like units in which stored patterns appear as stable activity states: presenting a noisy version of a stored pattern makes the dynamics flow to the clean pattern (content-addressable retrieval). The classic example is the Hopfield model (J. J. Hopfield, 1982): $N$ binary neurons $S_i \in \{\pm 1\}$, updated synchronously or sequentially by $S_i \leftarrow \mathrm{sgn}(\sum_k J_{i,k}S_k)$, with symmetric synaptic couplings $J_{i,k}=J_{k,i}$ set by the Hebbian prescription (strengthen a synapse in proportion to the correlation of the pre- and post-synaptic activities across the stored patterns). Symmetry gives a Lyapunov (energy) function, so the dynamics always relaxes to a fixed point. The model stores up to $\alpha = p/N \approx 0.14$ patterns ($p$ = number of stored patterns); beyond that the network collapses into a 'confused' state uncorrelated with any pattern, and even below capacity spurious mixture states (combinations of several stored patterns, in which e.g. only part of the network retrieves one pattern) compete with retrieval.

Two limitations motivate this problem. First, overloading is catastrophic: once the capacity is exceeded, no new pattern can be learned without destroying the old ones. Second, and more subtly, a symmetric network provides no internal signal telling an observer — or the network itself — whether it is currently retrieving a memory or is confused, because both kinds of state are static fixed points. The first limitation was addressed by 'palimpsest' schemes (G. Parisi, 'A memory which forgets', 1986; J.-P. Nadal, G. Toulouse, J.-P. Changeux, S. Dehaene, 1986; reviewed by Nadal, Toulouse, Mézard, Changeux and Dehaene, 1988), in which bounded or decaying synaptic strengths erase old patterns so that new ones can always be stored, at the price of forgetting — like a palimpsest, a manuscript scraped clean for rewriting.

Parisi's 1986 source paper ('Asymmetric neural networks and the process of learning', J. Phys. A 19, L675, 1986) attacked the second limitation in order to solve the first autonomously. If learning (a generalized Hebb rule with a forgetting term) were applied at all times, the network would memorize its own confused states and destroy memory; some intrinsic signal must indicate that the network is not confused. Parisi's observation was that a strongly asymmetric synaptic matrix $J_{i,k} \neq J_{k,i}$ has no energy function, and that in a random asymmetric network the time-windowed order parameter

$$q(t) = \frac{1}{N}\sum_{i=1}^{N}\left[\bar{S}_i(t)\right]^2, \qquad \bar{S}_i(t) = \frac{1}{T}\int_0^T S_i(t-\tau)\,\mathrm{d}\tau$$

vanishes even at low temperature: the network wanders chaotically rather than settling, so confused states are time-dependent while retrieval states (which echo a stored pattern) remain time-independent. This temporal signature can gate learning. Parisi proposed the rule

$$\frac{\mathrm{d}J_{i,k}}{\mathrm{d}t} = \left[-\lambda J_{i,k} + \bar{S}_i\bar{S}_k\right]\, f\!\left(\bar{S}_i\bar{S}_k\right),$$

where $\lambda$ is a forgetting rate and $f(x)$ is a threshold function that is essentially zero below a threshold $x_t$ and near one above it: synapses change only when the relevant neurons are quiescent (time-stable) over the window $T$, i.e. during retrieval-like activity, while chaotic confusion freezes learning. The intended outcome is a fully autonomous memory: no external 'learn now' signal and no externally scheduled pattern presentations are needed; the network's own dynamics decide when to imprint.

The paper verified only the premise (the vanishing of $q(t)$ for random asymmetric $J$ in a finite system) and, in its closing discussion, listed what remains unresolved: whether the network automatically transits into the confused state after a successful retrieval (the 'most delicate requirement' — without it, a retrieved pattern would be reinforced forever and nothing new could be learned); whether learning can start from a blank memory (no patterns stored, $\alpha = 0$), where pre-existing connectivity of the kind discussed in the Changeux selective-stabilization program might be essential; which biologically plausible realization of the asymmetry (random asymmetric couplings, asymmetric dilution, or a separation of excitatory and inhibitory synapses, the last reflecting the neurobiological Dale principle that each neuron's outputs are all of one sign) could support the mechanism, given that purely random connections are probably unrealistic; how the parameters $x_t$, $T$ and $\lambda$ must be tuned to avoid spurious states in which only half the network retrieves a pattern; and the missing theoretical framework for imprinting — the initial formation of memory traces.

Subsequent literature developed the ingredients but not the integration. Chaos in strongly asymmetric random networks was confirmed and quantified (H. Sompolinsky, A. Crisanti, H.-J. Sommers, 1988; M. Bauer and W. Martienssen, 1989, 1991; U. Bastolla and G. Parisi, 1997, 1998, who showed a sharp transition between short oscillatory attractors and exponentially long chaotic ones as the degree of symmetry is varied). Palimpsest schemes were solved exactly in asymmetric diluted networks (B. Derrida and J.-P. Nadal, 1987) and analyzed broadly (S. Shinomoto, 1987; Nadal et al., 1988). Dynamical signatures that discriminate retrieval from non-retrieval were constructed (J. F. Fontanari and R. Köberle, 1988; S. Nara and P. Davis, 1992; I. Tsuda, 1992; J. A. Horas and E. A. Bea, 2002). The storage capability of asymmetric and sign-constrained couplings was mapped out (E. Gardner, H. Gutfreund, I. Yekutieli, 1989; T. Fukai, 1990; H. Zhao, 2004; F. Aguirre-López, M. Pastore, S. Franz, 2022). Most recently, dynamical mean-field theory (DMFT — a self-consistent treatment of the fluctuating local fields in a large random network, replacing a single finite system by an effective stochastic single-neuron process) was developed for online learning with forgetting on directed sparse graphs (U. Pereira-Obilinovic, J. Aljadeff, N. Brunel, 2023), and the theory of non-reciprocal (asymmetrically coupled) many-body systems matured (M. Fruchart, R. Hanai, P. B. Littlewood, V. Vitelli, 2021; S. J. Fournier and P. Urbani, 2026).

Problem Statement

Following Parisi (J. Phys. A 19, L675, 1986), consider a network of $N$ binary neurons $S_i(t) \in \{\pm 1\}$ with strongly asymmetric synaptic strengths $J_{i,k} \neq J_{k,i}$, updated by $S_i(t+1) = \mathrm{sgn}\big(\sum_k J_{i,k}(t) S_k(t)\big)$, whose synapses evolve under the gated learning rule

$$\frac{\mathrm{d}J_{i,k}}{\mathrm{d}t} = \left[-\lambda J_{i,k} + \bar{S}_i \bar{S}_k\right]\, f\!\left(\bar{S}_i \bar{S}_k\right), \qquad \bar{S}_i(t) = \frac{1}{T}\int_0^T S_i(t-\tau)\,\mathrm{d}\tau,$$

with $f(x)$ a threshold function that is essentially zero below $x_t$ and near one above it, and $\lambda$ a forgetting rate. The question left open by the source and still unanswered is whether this architecture — with no external 'learn/stop' signal and no externally scheduled pattern presentations — self-organizes a sustained autonomous memory cycle. All of the following requested results serve this single objective:

(i) Demonstrate or refute that, after a successful retrieval and once the stored pattern has eroded under the $-\lambda J$ forgetting term, the network automatically exits into a chaotic confused state that switches the learning gate off and permits new patterns to be imprinted; or show instead that the network locks into permanent self-reinforcement of one pattern, or into spurious states in which only part of the network (e.g. half) retrieves a pattern, so that additional mechanisms are required.

(ii) Determine whether the cycle can start from a blank memory (no stored patterns, $\alpha = p/N = 0$) and establish what role pre-existing connectivity would have to play for the first memory traces to be imprinted at all — this is the source's missing 'framework for imprinting (the initial formation of memory traces)'.

(iii) Determine for which realizations of the asymmetry the cycle can operate: fully random asymmetric couplings, asymmetric dilution, or couplings constrained by a separation of excitatory and inhibitory synapses (the Dale principle), noting the source's remark that purely random connections are probably not biologically realistic.

(iv) Derive criteria for the threshold $x_t$, the averaging window $T$ and the forgetting rate $\lambda$ under which retrieval remains reliable, spurious half-retrieval states are avoided, and the autonomous cycle is sustained — numerical phase boundaries are required.

The autonomous-cycle question must be resolved without any external gating of learning or externally scheduled pattern presentations.

References

1. G. Parisi, 'Asymmetric neural networks and the process of learning', J. Phys. A: Math. Gen. 19 (1986) L675-L680. DOI: 10.1088/0305-4470/19/11/005. https://doi.org/10.1088/0305-4470/19/11/005
2. G. Parisi, 'A memory which forgets', J. Phys. A: Math. Gen. 19 (1986) L617-L620. DOI: 10.1088/0305-4470/19/10/011. https://doi.org/10.1088/0305-4470/19/10/011
3. J. J. Hopfield, 'Neural networks and physical systems with emergent collective computational abilities', Proc. Natl. Acad. Sci. USA 79 (1982) 2554-2558. DOI: 10.1073/pnas.79.8.2554. https://doi.org/10.1073/pnas.79.8.2554
4. J.-P. Nadal, G. Toulouse, J.-P. Changeux, S. Dehaene, 'Networks of formal neurons and memory palimpsests', Europhys. Lett. 1 (1986) 535-542. DOI: 10.1209/0295-5075/1/10/008. https://doi.org/10.1209/0295-5075/1/10/008
5. J.-P. Nadal, G. Toulouse, M. Mézard, J.-P. Changeux, S. Dehaene, 'Neural networks: learning and forgetting', in Computer Simulation in Brain Science (Cambridge University Press, 1988) pp. 221-231. DOI: 10.1017/cbo9780511983467.014. https://doi.org/10.1017/cbo9780511983467.014
6. B. Derrida, J. P. Nadal, 'Learning and forgetting on asymmetric, diluted neural networks', J. Stat. Phys. 49 (1987) 993-1009. DOI: 10.1007/BF01017556. https://doi.org/10.1007/BF01017556
7. S. Shinomoto, 'Memory maintenance in neural networks', J. Phys. A: Math. Gen. 20 (1987) L1305-L1309. DOI: 10.1088/0305-4470/20/18/015. https://doi.org/10.1088/0305-4470/20/18/015
8. A. Treves, D. J. Amit, 'Metastable states in asymmetrically diluted Hopfield networks', J. Phys. A: Math. Gen. 21 (1988) 3155-3169. DOI: 10.1088/0305-4470/21/14/016. https://doi.org/10.1088/0305-4470/21/14/016
9. J. F. Fontanari, R. Köberle, 'Neural networks with transparent memory', J. Phys. A: Math. Gen. 21 (1988) L259-L262. DOI: 10.1088/0305-4470/21/4/013. https://doi.org/10.1088/0305-4470/21/4/013
10. H. Sompolinsky, A. Crisanti, H.-J. Sommers, 'Chaos in random neural networks', Phys. Rev. Lett. 61 (1988) 259-262. DOI: 10.1103/PhysRevLett.61.259. https://doi.org/10.1103/PhysRevLett.61.259
11. M. Bauer, W. Martienssen, 'Quasi-periodicity route to chaos in neural networks', Europhys. Lett. 10 (1989) 427-431. DOI: 10.1209/0295-5075/10/5/007. https://doi.org/10.1209/0295-5075/10/5/007
12. M. Bauer, W. Martienssen, 'Lyapunov exponents and dimensions of chaotic neural networks', J. Phys. A: Math. Gen. 24 (1991) 4557-4566. DOI: 10.1088/0305-4470/24/19/019. https://doi.org/10.1088/0305-4470/24/19/019
13. P. Spitzner, W. Kinzel, 'Hopfield network with directed bonds', Z. Phys. B 74 (1989) 539-545. DOI: 10.1007/BF01311400. https://doi.org/10.1007/BF01311400
14. H. Rieger, M. Schreckenberg, J. Zittartz, 'Glauber dynamics of the asymmetric SK-model', Z. Phys. B 74 (1989) 527-538. DOI: 10.1007/BF01311399. https://doi.org/10.1007/BF01311399
15. M. Schreckenberg, 'Attractors in the fully asymmetric SK-model', Z. Phys. B 86 (1992) 453-459. DOI: 10.1007/BF01323739. https://doi.org/10.1007/BF01323739
16. Y. Mori, P. Davis, S. Nara, 'Pattern retrieval in an asymmetric neural network with embedded limit cycles', J. Phys. A: Math. Gen. 22 (1989) L525-L532. DOI: 10.1088/0305-4470/22/11/013. https://doi.org/10.1088/0305-4470/22/11/013
17. E. Gardner, H. Gutfreund, I. Yekutieli, 'The phase space of interactions in neural networks with definite symmetry', J. Phys. A: Math. Gen. 22 (1989) 1995-2008. DOI: 10.1088/0305-4470/22/12/005. https://doi.org/10.1088/0305-4470/22/12/005
18. T. Fukai, 'Metastable states of neural networks incorporating the physiological Dale hypothesis', J. Phys. A: Math. Gen. 23 (1990) 249-258. DOI: 10.1088/0305-4470/23/2/020. https://doi.org/10.1088/0305-4470/23/2/020
19. T. Tanaka, M. Yamada, 'A new associative memory which inhibits a meaningless output', Proc. IEEE International Conference on Neural Networks (1993) 1057-1062. DOI: 10.1109/ICNN.1993.298704. https://doi.org/10.1109/ICNN.1993.298704
20. S. Nara, P. Davis, 'Chaotic wandering and search in a cycle-memory neural network', Prog. Theor. Phys. 88 (1992) 845-855. DOI: 10.1143/ptp/88.5.845. https://doi.org/10.1143/ptp/88.5.845
21. I. Tsuda, 'Dynamic link of memory - chaotic memory map in nonequilibrium neural networks', Neural Networks 5 (1992) 313-326. DOI: 10.1016/S0893-6080(05)80029-2. https://doi.org/10.1016/S0893-6080(05)80029-2
22. U. Bastolla, G. Parisi, 'Attractors in fully asymmetric neural networks', J. Phys. A: Math. Gen. 30 (1997) 5613-5631. DOI: 10.1088/0305-4470/30/16/007. https://doi.org/10.1088/0305-4470/30/16/007
23. U. Bastolla, G. Parisi, 'Relaxation, closing probabilities and transition from oscillatory to chaotic attractors in asymmetric neural networks', J. Phys. A: Math. Gen. 31 (1998) 4583-4602. DOI: 10.1088/0305-4470/31/20/003. https://doi.org/10.1088/0305-4470/31/20/003
24. J. A. Horas, E. A. Bea, 'Distinguishing spurious and nominal attractors applying unlearning to an asymmetric neural network', Int. J. Neural Syst. 12 (2002) 109-116. DOI: 10.1142/S0129065702001047. https://doi.org/10.1142/S0129065702001047
25. H. Zhao, 'Designing asymmetric neural networks with associative memory', Phys. Rev. E 70 (2004) 066137. DOI: 10.1103/PhysRevE.70.066137. https://doi.org/10.1103/PhysRevE.70.066137
26. M. Fruchart, R. Hanai, P. B. Littlewood, V. Vitelli, 'Non-reciprocal phase transitions', Nature 592 (2021) 363-369. DOI: 10.1038/s41586-021-03375-9. https://doi.org/10.1038/s41586-021-03375-9
27. F. Aguirre-López, M. Pastore, S. Franz, 'Satisfiability transition in asymmetric neural networks', J. Phys. A: Math. Theor. 55 (2022) 305001. DOI: 10.1088/1751-8121/ac79e5. https://doi.org/10.1088/1751-8121/ac79e5
28. U. Pereira-Obilinovic, J. Aljadeff, N. Brunel, 'Forgetting leads to chaos in attractor networks', Phys. Rev. X 13 (2023) 011009. DOI: 10.1103/PhysRevX.13.011009. arXiv:2112.00119. https://doi.org/10.1103/PhysRevX.13.011009
29. S. J. Fournier, P. Urbani, 'High-dimensional dynamical systems: co-existence of attractors, phase transitions, maximal Lyapunov exponent and response to periodic drive', J. Stat. Mech. (2026) 043302. DOI: 10.1088/1742-5468/ae4f7e. https://doi.org/10.1088/1742-5468/ae4f7e
30. C. Albers, K. Pawelzik, 'Self-organization of asymmetric associative networks', BMC Neuroscience 10 (Suppl. 1) (2009) P340. DOI: 10.1186/1471-2202-10-S1-P340. https://doi.org/10.1186/1471-2202-10-S1-P340