- 1 Introduction
- 2 Background
- 3 A computational framework for making sense of sensory sequences
- 4 Computer implementation
- 6 Noisy apperception
- AI 2021
- https://www.sciencedirect.com/science/article/pii/S0004370220301855
- an ILP system for sequences. predict, retrodict, impute
- unsupervised program synthesis
- PI, object invention
- noised but low-dim inputs with only a few types of labels
1 Introduction
- symbolic theory
- explain, and unity
- causal language, \(Datalog^\ni\), generates a \(Datalog^\ni\) program
- relatively human-readable
- data-efficient
- elementary cellular automata, music, Seek Whence (sequences), multi-modal binding, occlusion
1.1 Related work
- model-based RL or MCTS
- accurate model of the game dynamics
- learning models
- three dimensions: latent? symbolic? prior?
- HMM
- only transition
- transition, perception, render
- ours: latent states, latent objects
- vectors: hard. symbols: relatively easy to understand
- some: state symbols, transition tensors
- prior: conv? event calculus? rules?
2 Background
- Datalog clause, interpreter in ASP
- subset-minimal Herbrand model
- ASP solvers, weak constraints
3 A computational framework for making sense of sensory sequences
3.1 - 3.4
- unambiguous symbolic sensory sequence
- theory, type, initial conditions, rules, constraints
- static rule, causal rule
- unary, binary, uniqueness constraint
- disallowing constants
- constraint, incompossible
- covered by
- example: three cycled states
- unity
- cost
3.5 -
- different interpretations
- trivial interpretation, upper bound
4 Computer implementation
- template, type signature, constants (static, causal, body atoms)
- increasing complexity
- lowest cost
- two non-trivial parts
- enumerate templates
- diagonalization, \((T,n)\) pairs
- infinite list of finite lists of: objects, predicates, variables
- find the best theory
- deduction, abduction, induction, combine (facts, rules, outputs)
- \(datalog^\ni\) interpreter in ASP
- ASP encoding, meta-interpreter
- complexity
- optimization
6 Noisy apperception
- length: increasing performance
- percentage of mislabelled data: decreasing performance