Tags / AI

• Neuroevolution Paper relating gene expression to development priors I met Dániel Barabási at the Neuroscience Imbizo in 2019, and chatted about his toy “XOY” model of how gene expression in different neuron types could establish broad-scale connectivity patterns in the developing brain. Later, while attending the Cosyne neuroscience conference, I realized there was a simple way to recast the evolutionary problem of designing particular neural connectivity to make an organism capable of rapid learning as a metalearning problem in which backpropogation …

• DLI: "Array Algebra" NumPy practical for the Deep Learning Indaba in Tunisia For the 2022 Deep Learning Indaba (held in Tunis), I contributed a practical entitled “Array Algebra”. The goal was to equip students with a better understanding of multidimensional arrays of various order, since array manipulation is the bread and butter of programming deep learning architectures. […] I used 2D and 3D illustrations of the fundamental operations of slicing, aggregation, and transposition, with a particular focus on manipulating image arrays as these illustrate …

• MCTS in Swift AlphaZero and MCTS implementation for OpenSpiel This project is a work-in-progress fork of the OpenSpiel library. It adds an MCTS implementation that will form the basis of a simple AlphaZero trainer. The design leverages some of the powerful type system abilities of Swift to achieve a maximally composable design, which allows the underlying tree-building implementation to be shared between several algorithms (various variants of MCTS, as well as Alpha-Beta search), which additionally makes them very clear. The fundamental Swift constructs …

• DLI: "Build Your Own TensorFlow" TensorFlow practicals for the Deep Learning Indaba in Kenya For the 2019 Deep Learning Indaba (held in Nairobi, Kenya), I was privileged to co-organize the tutorial sessions with Jamie Allingham, with the extensive help and support of Avishkar Boopchand, Stephan Gouws, and Ulrich Paquet of DeepMind. We had more than 50 tutors this year, covering more 500 students across 2 parallel sessions each day. Our goal for this year was to expand the range of material that the tutorials covered, to address pain points we saw in previous years, and to further …
• Spieeltjie experiment with multi-agent RL on zero-sum differentiable games Spieeltjie is a single-file package for doing simple experiments with multi-agent reinforcement learning on symmetric zero-sum games. For more information see “Open-ended learning in Symmetric Zero-Sum Games” and “A Unified Game-Theoretic Approach to Multiagent Reinforcement Learning”. The name “spieeltjie” comes from the Afrikaans word for “tournament”. […] This first set of images shows trajectories when starting from a set of random …
• Wasserstein GAN DFL contributed to Depth First Learning course on Wasserstein GANs While tutoring at the 2019 Deep Learning Indaba, I got to know the multi-talented Cinjon Resnik, who is currently doing his PhD with Kyunghyun Cho at NYU. After the Indaba, Cinjon invited me to join an experiment he is running in distributed teaching and learning called Depth First Learning. One innovation that particularly resonated with me was the effort DFL makes to plot a path through “paper space”, as a way to explain a core idea or story. The story we chose as the backbone for …
• Biologically Plausible Backprop feedback alignment & activity perturbation in PyTorch SOAP (Second Order Activity Perturbation) is a package for experimenting with a computational neuroscience phenomenon called feedback alignment in PyTorch. It formed my project for the 3-week IBRO-Simons Computational Neuroscience Summer School in 2019. It relates to an important challenge in reconciling how learning works in artificial neural networks with what we know about how real neurons behave, a topic called biologically plausible back-propogation. […] When a neuron fires, its axon …

• Deep Learning for Mathematica high-level, flexible, friendly deep learning From 2016 to 2018, I led development of the neural network framework, which encompassed 50+ functions and objects. Since version 10.3 it has been part of the core Mathematica language. It covers a wide range of state-of-the-art deep learning functionality, in an interface that is extremely easy for scientists and engineers to use, rivaling and exceeding the ease-of-use of libraries like Keras. One notable use of the framework was to replace the filter-bank based gravitional wave detectors used …

• Floatworld evolutionary sim of RNN-controlled agents on a gridworld floatworld is an C++ library and associated GUI that serves as a laboratory for a multi-agent RNN-based simulation in which agents compete for space and energy on a two-dimensional grid. I wrote floatworld, along with a short paper about it, during my undergraduate degree. Each agent is controlled by a recurrent neural network and harvests energy in order to reproduce. Time and space are discretized. Agents obtain a variety of data from their environments that is fed to their neural network …