Chen and Colleagues Proposes Bounded Old-State Modulation for Stable Quantum Sequence Modeling

A new bounded-state modulation rule addresses a key limitation in Quantum Fast-Weight Programmers (QFWPs), a promising technique for quantum sequence modelling. Kuo-Chung Peng of National Taiwan University and colleagues found that unbounded recurrent memory within Self-Modulating QFWPs can lead to divergence when processing long sequences. The introduction of a bounded old-state modulation rule, employing a tanh gate, stabilises the recurrent memory branch without compromising the additive and new-update modulation processes. Evaluations across quantum-dynamics forecasting and Milan SMS telecommunication activity prediction reveal that this bounded gating strategy not only prevents divergence but also enhances the strong performance of Self-Modulating QFWPs, confirming accumulated-memory modulation as a key mechanism for effective sequence learning.

Stabilising recurrent memory with bounded gating enhances long sequence prediction

Milan SMS telecommunication activity prediction accuracy improved by 17% with the new bounded old-state gate, surpassing the previous limit of unbounded Self-Modulating Quantum Fast-Weight Programmers (QFWPs). These QFWPs previously diverged when processing sequences exceeding 50 steps. Applying a sign-preserving tanh function to the recurrent memory branch of the QFWP, this bounded gate stabilises the model without impeding its ability to learn from new data, effectively preventing runaway calculations in longer sequences. Accumulated-memory modulation is crucial for effective sequence learning within Self-Modulating QFWPs, and this bounded gating strategy provides a targeted solution for long-sequence stability.

Detailed analysis showed that old-state modulation consistently delivered the most significant performance increases compared to the initial Quantum Fast-Weight Programmer (QFWP) baseline. Tests on CUDA-Q quantum-dynamics forecasting revealed the bounded gate not only prevented divergence in long sequences but also enhanced overall durability across various model configurations. Controlling accumulated fast weights, the memory stored in circuit parameters, is key to the architecture’s success. A further 8 percentage points of accuracy improvement was achieved on the Milan SMS telecommunication activity prediction task, resulting in a total gain of 17% over standard models. Further investigation explored the impact of different gate parameters, revealing that the sign-preserving tanh function offered an optimal balance between stability and learning capacity, and varying these parameters could potentially yield further gains.

Stabilising quantum memory through controlled retention of past data

Self-Modulating Quantum Fast-Weight Programmers, a novel approach to quantum sequence modelling storing information in adaptable circuit parameters rather than traditional recurrent memory, have been demonstrably stabilised. Previously, unbounded memory within these programmers caused instability when processing lengthy datasets, limiting their practical application. The bounded old-state modulation rule enables more reliable processing of extended data, as demonstrated through improved Milan SMS telecommunication activity prediction and quantum-dynamics forecasting. Researchers at National Taiwan University, collaborating with several international institutions, have successfully addressed a key limitation in Self-Modulating Quantum Fast-Weight Programmers (QFWPs). These quantum computer models learn by adjusting internal settings, and this stabilisation, achieved by controlling how the system retains past information, is vital for practical applications like forecasting and telecommunications activity prediction. The team are now exploring alternative bounding methods and investigating the potential for applying this technique to other quantum machine learning architectures.

Quantum sequence modelling presents a significant challenge for both classical and quantum computation. Traditional recurrent neural networks, while effective for many sequence tasks, struggle with vanishing or exploding gradients when dealing with long sequences, hindering their ability to capture long-range dependencies. Quantum Fast-Weight Programmers (QFWPs) offer a potential solution by shifting the burden of temporal information storage from recurrent hidden states to dynamically programmed variational-circuit parameters. This approach leverages the principles of quantum computation to potentially overcome the limitations of classical recurrent networks. The initial QFWP framework, however, still faced challenges related to stability, particularly when extended to Self-Modulating versions which incorporate input-dependent updates.

The Self-Modulating QFWP introduces input-dependent gates that control both the updates to the fast weights and the accumulated fast-weight state, allowing the model to adapt more effectively to the input sequence. However, this modulation, if unbounded, can lead to exponential growth of the old-state multiplier, causing the model to diverge during training and inference with sequences longer than 50 steps. This divergence represents a critical bottleneck, preventing the application of Self-Modulating QFWPs to real-world problems involving extended temporal data. The research detailed here directly addresses this issue by proposing a bounded old-state modulation rule.

The core of the solution lies in the application of a hyperbolic tangent (tanh) function as a gate to the recurrent memory branch. The tanh function, defined as tanh(x) = (e^x, e^-x) / (e^x + e^-x), naturally constrains its output to the range of -1 to 1. By multiplying the old-state modulation by this tanh gate, the researchers effectively limit the growth of the accumulated fast weights, preventing the unbounded behaviour that previously caused divergence. Crucially, the tanh function is chosen for its sign-preserving property, ensuring that the direction of the modulation is maintained, thereby minimising disruption to the learning process. The implementation was tested using the CUDA-Q platform, a software framework for developing and running quantum algorithms on NVIDIA GPUs.

The efficacy of the bounded gating strategy was evaluated across two distinct tasks. Firstly, quantum-dynamics forecasting, a benchmark problem in quantum machine learning, was used to assess the model’s ability to predict the evolution of quantum systems. Secondly, the researchers tackled the problem of Milan SMS telecommunication activity prediction, a real-world dataset representing the number of SMS messages exchanged in the city of Milan over a period of time. The results show a substantial improvement in performance. A 17% increase in accuracy for Milan SMS prediction, achieved through the combined effect of preventing divergence and enhancing learning, highlights the practical significance of this work. The initial QFWP baseline was surpassed, and the unbounded Self-Modulating QFWP’s limitations were overcome. The 8 percentage point improvement on top of existing gains further underscores the effectiveness of the bounded gate.

Further analysis revealed that the old-state modulation consistently contributed the most significant performance gains compared to other components of the Self-Modulating QFWP. This suggests that effectively managing the accumulated memory is paramount for successful sequence learning in this architecture. While the sign-preserving tanh function proved optimal in this study, the researchers acknowledge that exploring alternative bounding functions and parameter settings could potentially yield further improvements. Future work will focus on investigating these possibilities and extending the bounded gating strategy to other quantum machine learning architectures, potentially unlocking new capabilities in areas such as natural language processing and time-series analysis.

The researchers successfully implemented a bounded old-state modulation rule within a Quantum Fast-Weight Programmer, stabilising performance on long sequences. This addresses a previous limitation where unbounded modulation could lead to divergence, improving the robustness of the model. Results from quantum-dynamics forecasting and Milan SMS telecommunication activity prediction, showing an 8 percentage point improvement on existing gains and a 17% increase in accuracy, demonstrate the effectiveness of this approach. The study identifies accumulated-memory modulation as a key component of the system, and future work will explore extending this bounded gating strategy to other quantum machine learning architectures.