The computational power of interactive recurrent neural networks

Area of Science:

• Computational neuroscience
• Theoretical computer science
• Artificial intelligence

Background:

• Classical recurrent neural networks (RNNs) have established computational benchmarks against Turing machines.
• Understanding RNNs within biologically inspired frameworks is crucial for advancing AI.
• Sequential interactivity and memory persistence are key aspects of biological computation.

Purpose of the Study:

• To investigate the computational power of interactive RNNs in a biologically relevant context.
• To compare the capabilities of interactive rational- and real-weighted RNNs with interactive Turing machines.
• To provide a mathematical characterization of these computational powers.

Main Methods:

• Developing a biologically oriented computational framework for RNNs.
• Analyzing interactive rational- and real-weighted neural networks.
• Comparing their computational power to interactive Turing machines and Turing machines with advice.
• Deriving mathematical characterizations of the identified computational powers.

Main Results:

• Interactive rational-weighted neural networks match the power of interactive Turing machines.
• Interactive real-weighted neural networks are equivalent to interactive Turing machines with advice.
• A mathematical framework was established for these computational models.
• Real-weighted networks exhibit super-Turing capabilities, performing uncountably more translations.

Conclusions:

• Interactive real-weighted neural networks possess computational power beyond traditional Turing machines.
• This research bridges theoretical computer science and neuroscience by introducing biologically plausible computational models.
• The findings suggest new avenues for developing more powerful AI systems inspired by biological principles.