8.3390 Study Project: Computation at the Edge of Chaos in L2L Architectures (Part I)


Type Language Semester Credits Hours Room Time Term Year
Pr e 12 6 W 2018
MSc: Study project


Prerequisites: Recommended but not required: Linear Algebra, Analysis, Differential Equations, Neurodynamics, TensorFlow

In the first half of this study project, we will analyze the dynamics of long-short-term-memory-based (LSTM) (Andrychowicz et al., 2016) and spiking-neural-network-based (SNN) (Bellec et al., 2018) learning-to-learn (L2L) architectures. Part of the analysis will involve literature review, and part of it, implementing the networks from the papers to examine their dynamic properties. The objective is to find out whether the computation in these L2L architectures is happening at the edge of chaos (Bertschinger et al., 2004).
In the second half of the project, we will attempt to design novel learning algorithms and/or improve on existing ones from the cited papers using the insight we gained during the first half of the course. We will attempt to merge the ideas of backpropagation through time (BPTT) and computation at the edge of chaos, by supplying the BPTT algorithm with features that nudge the system towards the critical line where it can perform complex computations.