Machine Intelligence Research Frontier

Let’s start with the intersection of the Brain, Deepmind, and OpenAI research frontiers, and follow up with the workshops at ICML, NeurIPS, ICLR, EMNLP and ACL for the last 2 years. Later, I can add CVPR, AAAI.

ICML 2019, ICML 2018
NeurIPS 2019, NIPS 2018
ICLR 2019, ICLR 2018
EMNLP 2018
ACL 2018
ICML 2020
Deepmind Research Overview
Google Brain Research Overview
Open AI Research Overview

Transfer Learning / Domain Adaptation
Tools, Environment & Datasets
Reinforcement Learning
1. Model-based RL[a][b]
2. Exploration in RL[c][d]
3. Multi-Task Learning
4. Imitation Learning +1
Safety[e]
Deep Learning
1. Convolutional Neural Networks
2. Sequence 39.Modeling
  1. Recurrent Neural Networks
  2. Attention
3. Scalability and Speed
4. Parallel and Distributed Learning
5. Distillation / Compactness
Natural Language Processing & Understanding[f]
1. Word, Phrase, Paragraph, Document Representation
2. Semantics
3. Multilingual Methods
4. Information Extraction
Regularization
Multi-Modal[g]
Generative Models
1. GANs
2. VAEs
3. Normalizing Flows
Variational Inference
Linear Models
Unsupervised Learning
Clustering
Dimensionality Reduction
Autoencoders
Representation Learning
Memory
Multi-Agent Systems[h]
Metalearning
Neural Programming[i]
Hyperparameter Optimization
Loss Function Learning[j]
Evolution
Game Theory
General Machine Learning
Theory
Mean Field Theory
Infinite Width NNs
Probably Approximate Correctness
VC Dimension
Neuroscience
Interpretability[k]
Adversarial Examples
Kernel Machines
Collaborative Filtering
Graphical / Relational Learning
Optimization
Convex
Non-Convex
Bandits
Multi-Armed Bandit
Topological Data Analysis
Semi-Supervised Learning
Self-Supervised Learning
Learning to Rank & Structured Prediction
Feature Selection
Statistical Machine Learning
Bayesian Machine Learning[l]
1. Bayesian Deep Learning
2. Bayesian Nonparametrics
Statistical Processes
1. Gaussian Processes
2. Poisson Processes
MCMC
Uncertainty Estimation
Distributional Shift Robustness[m][n]
Reasoning
Causal Inference[o]
Online Learning
Active Learning[p]
Continual Learning / Life-Long Learning
Time Series
Information Theory
Intuitive Physics
Privacy, Anonymity
Security[q]
Miscellaneous
Applications
Speech Recognition
Image Categorization
Image Captioning
Natural Language Understanding
1. Machine Translation[r]
2. Language Modeling / Generation
3. Question Answering
4. Summarization
5. Search
6. Parsing
Pedestrian Detection
Grasp Detection
Go
Video
Dialogue
3D Object Reconstruction
Speaker Verification
Health Care
Theorem Proving
Music
Pose Estimation
Social Media
Speech Generation
System Design / Device Placement
Fairness
Super Resolution
Chemistry 1. Molecules and Materials
Robotics 1. Autonomous Vehicles
Physics
Games
Art
New types of hardware: neuromorphic computing or other types could SIGNIFICANTLY speed up some types of AI, with unpredictable consequences

Application:

Question for each subfield:
“If I want to do _____, what concepts do I need to understand, what facts do I need to memorize, what procedures do I need to practice?”
Ranking subfields by a standard
1. Impact on safety
2. Impact on general intelligence

[a]+ Models of the environment and its dynamics leads to interpretability [b]+1 [c]AIXI theoretical concerns on exploration: "to explore or not to explore" is a safety-related decision, and it's a fundamental uncertainty [d]+1 [e]Of course [f]+ Important for interfacing with the complexity of human goals, values [g]Making ConvNets explain their decisions using natural language and RL agents speak about decision would improve their interpretability (->safety) and how all of them work [h]Important for safety because many "human values" are derived from the need for cooperation in our evolutionary environment. Seems plausible that an AI trained in a multi-agent way would be easier to "control" or at least do positive-sum bargains with. [i]- Program Induction / Source Code Rewrites are likely to lead to unpredictable systems [j]+ Loss function learning / learned objective (for alignment with Human values) [k]+1 [l]Will help with uncertainty [m]Helpful for both inner optimizers, and using a NN to estimate a reward function [n]+1, Uncertainty Estimation is very important for behaving in the real world safely [o]+ Causal graphs of the environment, at the right level of abstraction, gives strong interpretability and controllability of the system [p]Critical for safety - plausibly amount of supervision is the bottleneck for alignment, and active learning seems the most promising way to use supervision more effectively. [q]Very important in short / medium term . Senior people I've talked to say this defending against adversaries (broader definition of security) is actually the #1 problem in ensuring recommender systems are good for users. [r]+ Translation of decisions / models of the world from neuralese into human language

Source: Original Google Doc