ΑΙhub.org
deep dive
Unsupervised skill discovery with contrastive intrinsic control
BAIR blog
01 Apr 2022
Unsupervised reinforcement learning (RL), where RL agents pre-train with self-supervised rewards, is an emerging paradigm for developing RL agents that are capable of generalization.
Assessing generalization of SGD via disagreement
ML@CMU
21 Mar 2022
We demonstrate that a simple procedure can accurately estimate the generalization error with only unlabeled data.
imodels: leveraging the unreasonable effectiveness of rules
BAIR blog
14 Mar 2022
imodels provides a simple unified interface and implementation for many state-of-the-art interpretable modeling techniques, particularly rule-based methods.
Why spectral normalization stabilizes GANs: analysis and improvements
ML@CMU
07 Mar 2022
We investigate the training stability of generative adversarial networks (GANs).
The unsupervised reinforcement learning benchmark
BAIR blog
14 Feb 2022
We consider the unsupervised RL problem - how do we learn useful behaviors without supervision and then adapt them to solve downstream tasks quickly?
Improving RL with lookahead: learning off-policy with online planning
ML@CMU
11 Feb 2022
We suggest using a policy that looks ahead using a learned model to find the best action sequence.
Sequence modeling solutions for reinforcement learning problems
BAIR blog
03 Feb 2022
We tackle large-scale reinforcement learning problems with the toolbox of sequence modeling.
An energy-based perspective on learning observation models
ML@CMU
21 Jan 2022
We propose a conceptually novel approach to mapping sensor readings into states.
Which mutual information representation learning objectives are sufficient for control?
BAIR blog
10 Jan 2022
How can we best design representation learning objectives?
Understanding user interfaces with screen parsing
ML@CMU
03 Jan 2022
We introduce the problem of screen parsing, which we use to predict structured user interface models from visual information.
Bridge data: boosting generalization of robotic skills with cross-domain datasets
BAIR blog
30 Dec 2021
With our proposed dataset and multi-task, multi-domain learning approach, we have shown one potential avenue for making diverse datasets reusable in robotics.
Why generalization in RL is difficult: epistemic POMDPs and implicit partial observability
BAIR blog
21 Dec 2021
In this blog post, we will aim to explain why generalization in RL is fundamentally hard, even in theory.
Designs from data: offline black-box optimization via conservative training
BAIR blog
10 Dec 2021
In this post, we discuss offline model-based optimization and some recent advances in this area.
A first-principles theory of neural network generalization
BAIR blog
22 Nov 2021
Find out more about research trying to shed light on the workings of deep neural networks.
Compression, transduction, and creation: a unified framework for evaluating natural language generation
ML@CMU
18 Nov 2021
Our framework classifies language generation tasks into compression, transduction, and creation.
Making RL tractable by learning more informative reward functions: example-based control, meta-learning, and normalized maximum likelihood
BAIR blog
15 Nov 2021
This article presents MURAL, a method for learning uncertainty-aware rewards for RL.
PICO: Pragmatic compression for human-in-the-loop decision-making
BAIR blog
05 Nov 2021
In this post, we outline a pragmatic compression algorithm called PICO.
Unsolved ML safety problems
BAIR blog
26 Oct 2021
We provide a new roadmap for ML Safety and aim to refine the technical problems that the field needs to address.
Distilling neural networks into wavelet models using interpretations
BAIR blog
18 Oct 2021
We propose a method which distills information from a trained DNN into a wavelet transform.
What can I do here? Learning new skills by imagining visual affordances
BAIR blog
30 Sep 2021
Can we use an analogous strategy to a child learning in a robotic learning system?
The limitations of limited context for constituency parsing
ML@CMU
24 Sep 2021
We consider the representational power of two important frameworks for constituency parsing.
Strategic instrumental variable regression: recovering causal relationships from strategic responses
ML@CMU
08 Sep 2021
How can we infer causal relationships in data while employing a decision-making model?
Universal weakly supervised segmentation by pixel-to-segment contrastive learning
BAIR blog
17 Aug 2021
Can a machine learn from a few labeled pixels to predict every pixel in a new image?
A unifying, game-theoretic framework for imitation learning
ML@CMU
13 Aug 2021
In this post we explore imitation learning and consider online, offline, and interactive methods.
The surprising effectiveness of PPO in cooperative multi-agent games
BAIR blog
06 Aug 2021
A study of cooperative multi-agent tasks, in which a group of agents is trying to optimize a shared reward function.
