Computable AI — A Machine Intelligence Blog

A machine intelligence blog — notes on deep reinforcement learning, probability, and the mathematics of learning machines.

Cox’s Theorem: Establishing Probability Theory

arXiv highlights

Cox’s theorem is the strongest argument for the use of standard probability theory. Here we examine the axioms to establish a firm foundation for the interpretation of probability theory as the unique extension of true-false logic to degrees of belief.

Nov 3, 2019

10 min

Comments on Eight Abstracts

arXiv highlights

An unfocused sweep of eight abstracts from a very busy week in AI research: Emergent tool use, why hierarchical learning can work so well, brain-inspired hardware for artificial neural networks, pretraining and transfer learning for RL, chromatic network compression, semi-supervised reward shaping, WGAN model imitation for model-based RL, and navigation in turbulent flows!

Oct 6, 2019

18 min

Active Perception in Adversarial Scenarios

arXiv highlights

Accumulating evidence about peers to discriminate potential threats.

Sep 22, 2019

5 min

Discovery of Useful Questions as Auxiliary Tasks

arXiv highlights

Learning more like a human, and more like a scientist, by actively seeking useful auxiliary questions during learning.

Sep 15, 2019

7 min

Deep Reinforcement Learning without Catastrophic Forgetting

arXiv highlights

Long-term learning of multiple tasks without forgetting old skills, using a new technique called Pseudo-Rehearsal.

Sep 9, 2019

4 min

Reward tampering

arXiv highlights

Improving safety and control by preventing all manner of reward tampering by the agent itself.

Aug 25, 2019

5 min

DRL Not Superhuman on Atari

arXiv highlights

DRL may not be superhuman on Atari after all, and how to avoid making mistakes like that in the future.

Aug 18, 2019

3 min

Three Method Comparison for Traffic Signal Control

arXiv highlights

Comparing supervised learning, random search, and deep reinforcement learning on traffic signal control.

Aug 11, 2019

7 min

Learning Compound and Composable Policies

arXiv highlights

Straightforward hierarchical RL for concurrent discovery of sub-policies and their controller.

Aug 4, 2019

5 min

Efficient exploration with self-imitation learning

arXiv highlights

I wonder if that happens every time…

Jul 28, 2019

4 min

Keeping to the Narrow Path

arXiv highlights

Better imitation learning with self-correcting policies by negative sampling.

Jul 21, 2019

4 min

Look at This: Where We See Shapes, AI Sees Textures

Look at This

CNNs trained in “the usual way” tend to learn something different than you might expect. They learn to recognize textures (local structure) rather than shapes (global structure).

Jul 16, 2019

2 min

Way Off-Policy Batch DRL

arXiv highlights

Pre-training using a generative model of pre-recorded trajectories and bias correction.

Jul 14, 2019

6 min

A New Series arXiv Sampler

arXiv highlights

Beginning a new series highlighting a few interesting RL papers on the arXiv each week. This week: Simple curriculum learning, learning to interact with humans, and warm starting RL with propositional logic.

Jul 7, 2019

8 min

Boltzmann Machines: Differentiation Work

Math

My differentiation work while reading Ilya Sutskever on the biological plausibility of Boltzmann machines.

Mar 10, 2019

2 min

Inaugural Post

Miscellany

The purpose statement and introduction to Computable AI.

Feb 16, 2019

1 min

Categories