On Differentiable Interpreters

Neural networks have transformed the fields of Machine Learning and Artificial Intelligence with the ability to model complex features and behaviours from raw data. They quickly became instrumental models, achieving numerous state-of-the-art performances across many tasks and domains. Yet the successes of these models often rely on large amounts of data. When data is scarce, resourceful ways of using background knowledge often help. However, though different types of background knowledge can be used to bias the model, it is not clear how one can use algorithmic knowledge to that extent. In this thesis, we present differentiable interpreters as an effective framework for utilising algorithmic background knowledge as architectural inductive biases of neural networks. By continuously approximating discrete elements of traditional program interpreters, we create differentiable interpreters that, due to the continuous nature of their execution, are amenable to optimisation with gradient descent methods. This enables us to write code mixed with parametric functions, where the code strongly biases the behaviour of the model while enabling the training of parameters and/or input representations from data. We investigate two such differentiable interpreters and their use cases in this thesis. First, we present a detailed construction of ∂4, a differentiable interpreter for the programming language FORTH. We demonstrate the ability of ∂4 to strongly bias neural models with incomplete programs of variable complexity while learning missing pieces of the program with parametrised neural networks. Such models can learn to solve tasks and strongly generalise to out-of-distribution data from small datasets. Second, we present greedy Neural Theorem Provers (gNTPs), a significant improvement of a differentiable Datalog interpreter NTP. gNTPs ameliorate the large computational cost of recursive differentiable interpretation, achieving drastic time and memory speedups while introducing soft reasoning over logic knowledge and natural language.