Neural Finite State Machines

• Neural FSMs are computational architectures that embed classical finite state machine dynamics within neural networks using learnable state embeddings and dynamic transition tables.
• They are implemented through various architectures such as feedforward, recurrent, and attractor networks to ensure temporal regularity and interpretability in sequential tasks.
• Neural FSMs offer practical insights for video analysis and reinforcement learning by providing transparent state extraction and enhancing phase recognition and policy diagnostics.

A Neural Finite State Machine (Neural FSM or NFSM) is any computational architecture—specifically, a neural or neuro-inspired system—that simulates, implements, or extracts the state-transition dynamics of a classical finite-state machine (FSM) within its structural or operational framework. Across the literature, variants include modules that impose explicit state-transition constraints in modern deep learning pipelines, constructive realizations of deterministic and nondeterministic automata in feedforward networks, symbolic attractor-based FSMs in recurrent networks, and extraction protocols for representing the behavior of a recurrent neural policy as an FSM abstraction. Neural FSMs thus bridge discrete automata theory with the representational and learning power of neural architectures.

1. Neural FSM Principles and Mathematical Formulation

Classical FSMs are defined by a finite set of states, an input alphabet, and a transition function, $\delta: S \times \Sigma \to S$. Neural FSMs generalize this construct by embedding the discrete state and transition structure in a neural parameterization, typically making both states and transitions trainable or learnable.

A prototypical formulation is the Neural FSM (NFSM) module for video phase recognition (Ding et al., 2024), which introduces:

• Learnable global state embeddings: Each discrete FSM state is assigned a distinct trainable vector $e_g \in \mathbb{R}^{s \times d}$, serving as a continuous identifier.
• Dynamic transition tables: For each frame, dynamic state embeddings $e_{dt}^i \in \mathbb{R}^{s \times d}$ are computed, defining transition matrices $T_t^i \in \mathbb{R}^{s \times s}$ via scaled-dot-product attention:

$$T_t^i = \mathrm{softmax}\left(\frac{1}{\sqrt{d}}\,e_{dt}^i e_g^\top\right).$$

• Transition-aware prediction: The current frame’s prediction $\hat p_t \in \Delta^s$ is combined with transition matrices to produce temporally regularized outputs $\tilde p_t^i = \hat p_t T_t^i$.
• Supervised losses: The design enforces both direct prediction accuracy and state-coherent transitions via composite loss functions:

$$\mathcal{L} = \mathcal{L}_c + \alpha \mathcal{L}_{\mathrm{trans}},$$

where $\mathcal{L}_c$ is the cross-entropy for direct predictions, and $\mathcal{L}_{\mathrm{trans}}$ enforces transition correctness over a history and pseudo-future window.

Other formulations include exact NFA and DFA constructions in fixed depth, width-bounded feedforward networks (Dhayalkar, 30 May 2025, Dhayalkar, 16 May 2025), and distributed symbolic attractors in recurrent networks (Cotteret et al., 2022).

2. Constructive Neural Implementations of FSMs

Neural FSMs can be realized in numerous architectural paradigms:

• Feedforward ReLU Networks for DFA/NFA Simulation: Both deterministic and nondeterministic automata can be exactly simulated by fixed-depth ReLU networks. Each automaton state maps to a binary or one-hot vector; transitions correspond to sparse weight matrices. For NFAs, a three-layer network of width $e_g \in \mathbb{R}^{s \times d}$0 is sufficient and can match the acceptance behavior of any $e_g \in \mathbb{R}^{s \times d}$1-state machine (Dhayalkar, 30 May 2025):

$e_g \in \mathbb{R}^{s \times d}$2

$e_g \in \mathbb{R}^{s \times d}$3-closure is handled by repeated application of $e_g \in \mathbb{R}^{s \times d}$4 and ReLU.

• Explicit DFA-Unrolled Neural Networks: For DFAs, each state and input is encoded, and the entire transition sequence is “unrolled” in network depth. Transition functions are implemented as MLP submodules, with exponentially compressed representations possible via binary threshold units (Dhayalkar, 16 May 2025). The Myhill-Nerode equivalence classes are preserved via embedding in a continuous latent space.
• Hopfield Networks and VSAs: Arbitrary FSMs are realized by encoding states and input symbols as high-dimensional random vectors. The transition relation $e_g \in \mathbb{R}^{s \times d}$5 is implemented by a weight matrix $e_g \in \mathbb{R}^{s \times d}$6, such that network dynamics $e_g \in \mathbb{R}^{s \times d}$7 recover the correct successor state upon symbol binding (Cotteret et al., 2022). Capacity scales linearly or quadratically with dimension, depending on sparsity.
• Coupled sWTA/Attractor Circuits: Robust FSMs are built from two coupled recurrent soft winner-take-all networks, representing memory states as paired activity patterns stabilized via cross-map excitation. “Transition neurons” implement state changes conditionally on symbol input, yielding a biophysically plausible neural DFA (0809.4296).

3. Extraction and Interpretation of Neural FSMs from RNNs

FSM extraction from trained recurrent neural networks is operationalized via state-space clustering and quantized observation encoding (Danesh et al., 2020). The extraction pipeline:

1. Data collection: Sample environment interactions and record RNN hidden activations and discretized observations.
2. Clustering: Aggregate hidden vectors into $e_g \in \mathbb{R}^{s \times d}$8 clusters (states) using K-means or similar; map observations to a discrete alphabet.
3. Transition mapping: Reconstruct empirical transition tables $e_g \in \mathbb{R}^{s \times d}$9 by tracking hidden and input cluster sequences.
4. Reduction and Pruning: Post-process to merge functionally equivalent states, collapse trivial loops, and focus on decision points (states with nontrivial outgoing transitions).
5. Attention-based analysis: Integrated gradients or similar methods identify which observations drive key transitions.

The extracted neural FSM can be further simplified while preserving decision semantics, offering transparent, actionable abstractions of complex policies in reinforcement learning.

4. Expressivity, Learnability, and Computational Boundaries

• Regular Language Characterization: Feedforward networks of bounded depth and width match precisely the class of regular languages (DFA/NFA). All such networks admit exact simulation of any regular language; no fixed-size network can recognize non-regular languages such as $e_{dt}^i \in \mathbb{R}^{s \times d}$0 (Dhayalkar, 30 May 2025, Dhayalkar, 16 May 2025).
• Gradient Descent and Structure Preservation: Gradient descent over structure-preserving networks (masks enforcing symbolic structure) permits learning while maintaining exact automata semantics in the trained model (Dhayalkar, 30 May 2025).
• Undecidability of Equivalence: For general weighted RNNs (e.g., with ReLU or piecewise-linear activations), the equivalence problem with PDFA/WFA/DPFA is undecidable, and no non-trivial distance can be recursively computed (Marzouk et al., 2020, Marzouk, 2020). Even truncated finite-support equivalence is EXP-hard, and the Chebyshev distance is NP-hard to approximate.
• Practical Extraction Algorithms: For small and smooth RNNs (low norm, entropy, or spectral radius), DFA extraction via clustering or active learning can be empirically reliable, but theoretical hardness persists for arbitrary models (Marzouk, 2020).
• Stability and Robustness: Attractor-based neural FSMs are robust to synaptic noise, quantization, and severe weight sparsity; feedforward implementations offer exponential compression via binary coding of hidden states (Cotteret et al., 2022, Dhayalkar, 16 May 2025).

5. Applications: Sequential Video Analysis, Policy Interpretabiliy, and Biological Plausibility

Neural FSMs play a critical role in applications where process or temporal stage structure is intrinsic:

• Surgical Phase Recognition: As an addon module, NFSM provides explicit state-transition supervision, dynamic transition forecasting, and interpretable state assignments, improving temporal coherence, phase-level metrics, and overall accuracy in both surgical and general long-range sequential video analysis. As demonstrated on Cholec80, BernBypass70, and the Breakfast dataset, state embeddings and transition-aware supervision are the primary drivers of gains, with dynamic transition tables and forward-prediction offering additional robustness (Ding et al., 2024).
• Policy Analysis in RL: Extraction of neural FSMs from RNN policies yields decision diagrams that illuminate the structure and sensitivities of learned controllers, enabling transparent diagnosis of behavior as observed in Atari and classical control benchmarks (Danesh et al., 2020).
• Distributed Memory in Attractor Networks: Construction of FSMs via Hopfield or coupled sWTA networks offers models for symbolic memory and conditional sequence computation with high capacity and error tolerance, supporting hypotheses of biological neural implementation of state-dependent computation (Cotteret et al., 2022, 0809.4296).

6. Future Directions and Methodological Implications

Current research continues to expand neural FSM methodology along multiple axes:

• Integration with Modern Architectures: Drop-in FSM modules for deep sequence models, exploiting differentiable transition tables and learnable embeddings, serve as practical tools for enforcing explicit temporal structure in large-scale video and event modeling (Ding et al., 2024).
• Theoretical Characterization: Fully constructive proofs demonstrate the limitations and compressibility of neural FSMs, with transparent conversion between symbolic automata and their neural realizations (Dhayalkar, 30 May 2025, Dhayalkar, 16 May 2025).
• Extraction Complexity and Approximate Semantics: Given undecidability and hardness results for general neural-to-FSM extraction, ongoing work targets tractable approximate equivalence, margin-aware training, spectral regularization, and adaptive abstraction mechanisms (Marzouk, 2020, Marzouk et al., 2020).
• Biological and Hardware Realizability: Attractor and sWTA-based FSMs provide pathways for neuromorphic or nano-device implementations, robust to substantial noise and quantization, aligning with cortical processing constraints (Cotteret et al., 2022, 0809.4296).

Neural FSMs thus form an active research interface connecting deep learning, automata theory, interpretable AI, and computational neuroscience, with an overview of discrete transition structure and neural computation as their foundational paradigm.