Memory modules for deep learning Theory of Neural Systems

Description

Memory modules such as differentiable neural computer [1] and end-to-end memory networks [2] are often used in deep learning to store medium and long term memories. These contribute to the success of architectures that can solve extremely difficult tasks such as natural language question answering and multi-player capture the flag [3].

Recent alternatives have emerged that are inspired by the brain while being simpler and having the potential to be more efficient — for example using Hebbian plasticity [4].

This Masters thesis will explore further alternatives to memory modules that use forms of memory and plasticity that have further advantages in terms of memory capacity and efficiency. These could be based on, for example, [5], [6]. Exploring connections to biological forms of plasticity will be encouraged. The modules will be implemented and evaluated for supervised learning and/or reinforcement learning tasks. Such modules are especially of interest in reinforcement learning to emulate episodic memory.

Required skills:

Knowledge of Python and deep learning frameworks such as Tensorflow/Pytorch/JAX.

References:

[1] Graves, Alex, et al. ‘Hybrid Computing Using a Neural Network with Dynamic External Memory’. Nature, vol. advance online publication, Oct. 2016. doi:10.1038/nature20101.

[2] Sukhbaatar, Sainbayar, et al. ‘End-to-End Memory Networks’. Proceedings of the 28th International Conference on Neural Information Processing Systems - Volume 2, MIT Press, 2015, pp. 2440–48.

[3] Jaderberg, Max, et al. ‘Human-Level Performance in 3D Multiplayer Games with Population-Based Reinforcement Learning’. Science, vol. 364, no. 6443, May 2019, pp. 859–65. doi:10.1126/science.aau6249.

[4] Limbacher, Thomas, and Robert Legenstein. ‘H-Mem: Harnessing Synaptic Plasticity with Hebbian Memory Networks’. Advances in Neural Information Processing Systems, vol. 33, 2020, pp. 21627–37.

[5] Jaeckel, Louis A. A Class of Designs for a Sparse Distributed Memory. 1 July 1989. NASA NTRS.

[6] Willshaw, D. J., et al. ‘Non-Holographic Associative Memory’. Nature, vol. 222, no. 5197, June 1969, pp. 960–62. doi:10.1038/222960a0.

The Institut für Neuroinformatik (INI) is a central research unit of the Ruhr-Universität Bochum. We aim to understand the fundamental principles through which organisms generate behavior and cognition while linked to their environments through sensory systems and while acting in those environments through effector systems. Inspired by our insights into such natural cognitive systems, we seek new solutions to problems of information processing in artificial cognitive systems. We draw from a variety of disciplines that include experimental approaches from psychology and neurophysiology as well as theoretical approaches from physics, mathematics, electrical engineering and applied computer science, in particular machine learning, artificial intelligence, and computer vision.

Universitätsstr. 150, Building NB, Room 3/32
D-44801 Bochum, Germany

Tel: (+49) 234 32-28967
Fax: (+49) 234 32-14210