Transparency of Classical and QM AI

• What are the implications of Quantum information theory for AI?
  • Quantum-enchanced machine learning
  • We can also use the help of ML to work on QM problems

• Transparency paper by Adrian Weller (arXiv: 1708.01870)

• Transparency:
  • Understanding of how the system works,
  • Can localise and diagnose errors,
  • Fix them.
• In terms of debugging:
  • It would be good to identify that a particular step is in error locally, without waiting for the end result.
  • Identify the causes of deviation.
  • But with quantum systems the log is more complicated because events are quantum events.
    • Interesting note: many worlds is the first theory to consider?
• Quantum computer as an agent:
  • Measure a qubit at a time (system changes state for the next measurement)
  • This is similar to an agent that gets percepts and outputs actions, however the agent programm is replaced by a quantum system.
• Quantum agents vs. environment:
  • CC CQ QC QQ (depending on which part(s) you quantize)
  • The last case is the most crazy because it's quantum system learning/interacting with a quantum history.
• Projective simulation + episodic memory:
  • https://arxiv.org/pdf/1104.3787.pdf
  • Creative Machines in Scientific Reports 2, 522 (2012)
  • History of the agent in a physical world is a line in (s, a, t)-space where s=percepts, a=actions, t=time.
  • Applications:
    • Textbook problems: gridworld, mountain car, etc.
    • Robotics: learning to manipulate objects.
    • Design of novel quantum experiments.
      • Create highly-dimensional entangled states:
        • Optical table, we can add and remove elements.
        • People tried brute-force search, but we could use reinforcement learning with PS instead.
        • We ended up with some setups of the optical table that produce interesting experiments (high-dim entangled states). These are then useful for humans who work with quantum optics.
• PS can meta-learn (tune learning parameters of itself).

• PS gives a clear route to quantizing.
  • Clips can become quantum states in a Hilbert space and we replace a classical search with quantum search.

• Modelling of free agency under determinism (a paper wth T. Müller)
  • Cat thinking about jumping on the table: deliberations and actions connected into a Markov-ish network. The log of deliberation can be seen as an explanation,
    • If we see a sequence of deliberations that we don't like, we can give agents learning scenarios that would lead it to learn and update in the direction we like. Or we can directly change the network of clips.
    • This system is transparent, however, if we use quantum search, we can't produce the log without losing peformance. How do we balance transparency with power?

• Q/A
  • Is deliberation process better described as a random walk or as a rational process driven by some normative constraints.
  • There's no planning here, at least not an explicit intentional plan.