BEG Neural Networks: Ternary Memory

• BEG neural networks are fully connected associative memory models with ternary neurons that leverage explicit pattern dilution for enhanced storage and retrieval.
• They employ a two-sector Hamiltonian with tailored Hebbian learning and Guerra interpolation to analyze both serial and parallel recall regimes.
• The model’s phase diagram, capacity scaling, and graded-response generalizations provide actionable insights for designing high-capacity, multitasking neural systems.

The Blume-Emery-Griffiths (BEG) neural network is a fully connected associative memory model generalizing the Hopfield paradigm, enabling richer neuron state spaces, incorporating explicit pattern sparsity, and supporting serial and parallel recall regimes. Neurons are ternary ($\sigma_i \in \{-1,0,+1\}$), and patterns employ dilution—some entries are zero (“inactive”). The BEG Hamiltonian’s two-sector structure, coupled with tailored Hebbian learning rules and threshold terms, provides enhanced storage and multitasking capabilities relative to classical binary models. Rigorous analysis via Guerra interpolation and replica-symmetric calculations elucidates the model’s phase diagram, storage scaling, and retrieval properties in both mild and extreme dilution regimes, including generalizations to graded-response neuron states and relations to inverse-freezing phenomena.

1. Formalism: Network Architecture and Hamiltonian

The BEG associative memory network consists of $N$ neurons, each taking values $\sigma_i \in \{-1,0,+1\}$, storing $K$ random ternary patterns $\xi_i^\mu \in \{-1,0,+1\}$ with dilution parameter $a$: $P(\xi_i^\mu = \pm 1) = a/2$, $P(\xi_i^\mu = 0) = 1-a$ (Albanese et al., 12 Jan 2026). The energy function comprises two Hebbian terms:

$$H(\sigma \mid \xi) = -\frac{1}{2Na} \sum_{i, j, \mu} \xi_i^\mu \xi_j^\mu \sigma_i \sigma_j - \frac{1}{2N a(1-a)} \sum_{i, j, \mu} \eta_i^\mu \eta_j^\mu \sigma_i^2 \sigma_j^2,$$

where $\eta_i^\mu = (\xi_i^\mu)^2 - a$ centers the second-order pattern statistics. Recasting in terms of Mattis overlaps:

• $N$0 (retrieval quality),
• $N$1,

the Hamiltonian simplifies to

$N$2

For $N$3, the system reduces to the standard Hopfield model. The second quadratic term uniquely enables isolation of statistical contributions from inactive patterns.

2. Order Parameters and Self-Consistency

Internal states and retrieval dynamics are characterized by the Mattis overlaps $N$4 and $N$5, quantifying the signal and quadratic correlation with stored patterns, respectively (Albanese et al., 12 Jan 2026). For high storage loads ($N$6), replica theory introduces overlaps

• $N$7,
• $N$8

for distinct replicas $N$9, $\sigma_i \in \{-1,0,+1\}$0, capturing fluctuation phenomena. Auxiliary order parameters $\sigma_i \in \{-1,0,+1\}$1, $\sigma_i \in \{-1,0,+1\}$2 emerge in the interpolating framework. Closed self-consistency equations for all parameters generalize the classical Hopfield mean-field equations.

3. Guerra Interpolation and Replica Symmetry Free Energy

Rigorous computation of the BEG thermodynamic limit employs the Guerra interpolation method, constructing a partition function $\sigma_i \in \{-1,0,+1\}$3 interpolating between the fully coupled BEG model ($\sigma_i \in \{-1,0,+1\}$4) and single-site decoupled problems ($\sigma_i \in \{-1,0,+1\}$5) (Albanese et al., 12 Jan 2026). The pressure $\sigma_i \in \{-1,0,+1\}$6 evolves under

$\sigma_i \in \{-1,0,+1\}$7

with $\sigma_i \in \{-1,0,+1\}$8 comprising terms in the order parameters and auxiliary fields under replica-symmetric (RS) assumptions. The RS free energy is expressed as

$\sigma_i \in \{-1,0,+1\}$9

All macroscopic observables are determined from stationary points of this pressure subject to coupled RS equations.

4. Pattern Dilution: Serial and Parallel Recall Regimes

Pattern dilution ($K$0) introduces a fraction $K$1 of truly inactive neurons (‘blank sites’) per pattern, fundamentally altering recall dynamics (Albanese et al., 12 Jan 2026). In the low-load regime ($K$2), pure-state serial recall with $K$3 is feasible; however, blank sites enable lower-energy configurations by aligning with other patterns.

A key transition from serial to parallel recall occurs: when the total overlap with subleading patterns balances that of the leading pattern, energetic optimization favors simultaneous recall (“parallel recall”). For $K$4 at zero temperature,

$K$5

where $K$6 marks the critical dilution threshold. For larger $K$7, $K$8 is defined by $K$9. Energetic analyses demonstrate that when patterns activate disjoint neuron subsets, fully parallel recall yields a lower energy than strictly serial recall.

5. Dilution Phases: Hierarchical and Equal-Strength Recall

Two major dilution regimes govern the BEG network's multitasking properties (Albanese et al., 12 Jan 2026):

• Mild dilution ($\xi_i^\mu \in \{-1,0,+1\}$0 fixed in $\xi_i^\mu \in \{-1,0,+1\}$1, small $\xi_i^\mu \in \{-1,0,+1\}$2): hierarchical recall, with overlaps decaying as $\xi_i^\mu \in \{-1,0,+1\}$3 for $\xi_i^\mu \in \{-1,0,+1\}$4, and up to $\xi_i^\mu \in \{-1,0,+1\}$5 patterns recalled. Amplitudes are distributed hierarchically, and resource exhaustion rapidly limits total multitasking.
• Extreme dilution ($\xi_i^\mu \in \{-1,0,+1\}$6, $\xi_i^\mu \in \{-1,0,+1\}$7, $\xi_i^\mu \in \{-1,0,+1\}$8): $\xi_i^\mu \in \{-1,0,+1\}$9 balances central limit noise effects, enabling simultaneous recall of $a$0 patterns, all with equal overlap strength $a$1, yielding the “flat multitasking” phase. The corresponding phase diagram in $a$2 space shows single-recall, hierarchical-serial, and fully parallel domains as $a$3 and $a$4 are varied.

6. Graded-Response and Ghatak-Sherrington Generalizations

The BEG model admits graded-response generalizations by extending neuron states to $a$5 for $a$6 (Albanese et al., 12 Jan 2026). For $a$7, the standard BEG model is recovered; for $a$8, the binary Hopfield model is embedded. Patterns take similarly graded values with identical dilution $a$9. The Hamiltonian’s variance terms are rescaled to $P(\xi_i^\mu = \pm 1) = a/2$0 and $P(\xi_i^\mu = \pm 1) = a/2$1; order parameters are correspondingly renormalized.

The replica-symmetric free energy and self-consistency equations generalize directly in the Guerra framework, now summing over multiple level indices. This construction imports phenomena such as Ghatak-Sherrington inverse freezing into the associative memory context, linking BEG-type architectural features to broader classes of multi-state spin-glass models.

7. Sparse BEG Networks: Storage Capacity and Comparisons

In the extreme sparse regime with activity $P(\xi_i^\mu = \pm 1) = a/2$2, BEG networks can store up to

$P(\xi_i^\mu = \pm 1) = a/2$3

patterns as fixed points under the zero-temperature retrieval dynamics (Heusel et al., 2017). The network update is governed by a thresholded, hybrid asynchronous rule:

$P(\xi_i^\mu = \pm 1) = a/2$4

where $P(\xi_i^\mu = \pm 1) = a/2$5 denotes the bilinear Hebb sum, $P(\xi_i^\mu = \pm 1) = a/2$6 the quadratic threshold, and $P(\xi_i^\mu = \pm 1) = a/2$7 is optimized at $P(\xi_i^\mu = \pm 1) = a/2$8. The ternary state space and explicit chemical potential favor the zero state, gating crosstalk noise. Compared to other sparse associative memories (Willshaw, Amari, Gripon-Berrou, sparse Hopfield), BEG achieves substantially higher $P(\xi_i^\mu = \pm 1) = a/2$9 parameter and capacity:

Model	Capacity $P(\xi_i^\mu = 0) = 1-a$0
BEG	$P(\xi_i^\mu = 0) = 1-a$1
Willshaw	$P(\xi_i^\mu = 0) = 1-a$2–$P(\xi_i^\mu = 0) = 1-a$3
Amari	$P(\xi_i^\mu = 0) = 1-a$4–$P(\xi_i^\mu = 0) = 1-a$5
Gripon–Berrou	$P(\xi_i^\mu = 0) = 1-a$6–$P(\xi_i^\mu = 0) = 1-a$7
Sparse Hopfield	$P(\xi_i^\mu = 0) = 1-a$8–$P(\xi_i^\mu = 0) = 1-a$9

The threshold $$H(\sigma \mid \xi) = -\frac{1}{2Na} \sum_{i, j, \mu} \xi_i^\mu \xi_j^\mu \sigma_i \sigma_j - \frac{1}{2N a(1-a)} \sum_{i, j, \mu} \eta_i^\mu \eta_j^\mu \sigma_i^2 \sigma_j^2,$$0 is critical for optimizing sparse recall performance. This scaling holds with high probability over i.i.d. random patterns, though real-world nonuniformities may affect the constant.

8. Retrieval Accuracy, Multitasking, and Performance Trade-offs

BEG/GS associative networks display distinct scaling regimes in retrieval accuracy and multitasking (Albanese et al., 12 Jan 2026):

• Storage capacity: For serial (high-load) recall, $$H(\sigma \mid \xi) = -\frac{1}{2Na} \sum_{i, j, \mu} \xi_i^\mu \xi_j^\mu \sigma_i \sigma_j - \frac{1}{2N a(1-a)} \sum_{i, j, \mu} \eta_i^\mu \eta_j^\mu \sigma_i^2 \sigma_j^2,$$1 at $$H(\sigma \mid \xi) = -\frac{1}{2Na} \sum_{i, j, \mu} \xi_i^\mu \xi_j^\mu \sigma_i \sigma_j - \frac{1}{2N a(1-a)} \sum_{i, j, \mu} \eta_i^\mu \eta_j^\mu \sigma_i^2 \sigma_j^2,$$2 (Hopfield-like, with slight modifications from the quadratic sector). Sparse networks push this to $$H(\sigma \mid \xi) = -\frac{1}{2Na} \sum_{i, j, \mu} \xi_i^\mu \xi_j^\mu \sigma_i \sigma_j - \frac{1}{2N a(1-a)} \sum_{i, j, \mu} \eta_i^\mu \eta_j^\mu \sigma_i^2 \sigma_j^2,$$3 for extreme sparsity.
• Retrieval accuracy: $$H(\sigma \mid \xi) = -\frac{1}{2Na} \sum_{i, j, \mu} \xi_i^\mu \xi_j^\mu \sigma_i \sigma_j - \frac{1}{2N a(1-a)} \sum_{i, j, \mu} \eta_i^\mu \eta_j^\mu \sigma_i^2 \sigma_j^2,$$4 in low-load, declines with $$H(\sigma \mid \xi) = -\frac{1}{2Na} \sum_{i, j, \mu} \xi_i^\mu \xi_j^\mu \sigma_i \sigma_j - \frac{1}{2N a(1-a)} \sum_{i, j, \mu} \eta_i^\mu \eta_j^\mu \sigma_i^2 \sigma_j^2,$$5 in mild dilution, but maintains finite $$H(\sigma \mid \xi) = -\frac{1}{2Na} \sum_{i, j, \mu} \xi_i^\mu \xi_j^\mu \sigma_i \sigma_j - \frac{1}{2N a(1-a)} \sum_{i, j, \mu} \eta_i^\mu \eta_j^\mu \sigma_i^2 \sigma_j^2,$$6 per pattern in extreme dilution multimode recall.
• Multitasking: In mild dilution, up to $$H(\sigma \mid \xi) = -\frac{1}{2Na} \sum_{i, j, \mu} \xi_i^\mu \xi_j^\mu \sigma_i \sigma_j - \frac{1}{2N a(1-a)} \sum_{i, j, \mu} \eta_i^\mu \eta_j^\mu \sigma_i^2 \sigma_j^2,$$7 patterns can be hierarchically recalled. Extreme dilution enables $$H(\sigma \mid \xi) = -\frac{1}{2Na} \sum_{i, j, \mu} \xi_i^\mu \xi_j^\mu \sigma_i \sigma_j - \frac{1}{2N a(1-a)} \sum_{i, j, \mu} \eta_i^\mu \eta_j^\mu \sigma_i^2 \sigma_j^2,$$8 patterns with equal amplitude.
• Trade-off: Increasing the dilution $$H(\sigma \mid \xi) = -\frac{1}{2Na} \sum_{i, j, \mu} \xi_i^\mu \xi_j^\mu \sigma_i \sigma_j - \frac{1}{2N a(1-a)} \sum_{i, j, \mu} \eta_i^\mu \eta_j^\mu \sigma_i^2 \sigma_j^2,$$9 enhances pure-state capacity but suppresses multitasking, while decreasing $\eta_i^\mu = (\xi_i^\mu)^2 - a$0 promotes parallel recall but reduces per-pattern overlap strength $\eta_i^\mu = (\xi_i^\mu)^2 - a$1.

A plausible implication is that moderate pattern dilution transforms classic serial associative memory into a genuine multitasking architecture, providing design guidelines for multi-level neural coding in both biological and synthetic memory systems.

9. Limitations, Assumptions, and Future Directions

The rigorous results for BEG networks rely on assumptions including zero-temperature retrieval, i.i.d. pattern distributions, and optimal threshold tuning. Thermal noise or correlated patterns may require alternate threshold choices or induce performance degradation. Scaling laws for capacity and retrieval persist under mild variations, but precise constants depend on idealized noise models. Finite-size networks may require practical adjustment of threshold parameters for optimal fixed-point retrieval. The broad phenomenology—dilution-driven serial–parallel transitions, rich phase diagrams, and graded-response flexibility—suggests applicability in high-capacity sparse memory architectures, and motivates further study of multi-level coding and multitasking in neural substrates.

References: (Albanese et al., 12 Jan 2026, Heusel et al., 2017)