Most of these items are gathered from Jack Clark's Import AI(https://jack-clark.net/about/) newsletter.

Many of these items reference articles on Cornell's arVix: https://arxiv.org/

A consortium of Chinese universities along with seventeen companies - including Alibaba, Tencent, Baidu, and ByteDance - have developed AIBench, an AI benchmarking suite meant to compete with MLPerf, an AI benchmarking suite predominantly developed by American universities and companies. http://www.benchcouncil.org/AIBench/index.html

Google researchers have trained a chatbot with uncannily good conversational skills. The bot, named Meena, is a 2.6 billion parameter language model trained on 341GB of text data, filtered from public domain social media conversations. Meena uses a seq2seq model (the same sort of technology that powers Google's “Smart Compose” feature in gmail), paired with an Evolved Transformer encoder and decoder - it's interesting to see something like this depend so much on a component developed via neural architecture search. https://arxiv.org/abs/2001.09977

MuZero from DeepMind, combines all previous programs for Atari, Go, Chess, Space Invaders. And it learns the games rules itself via reinforcement learning. AlphaZero, the previously lauded Go champion, was fed the game rules to start. MuZero gets equally good results, but was never fed the game rules; it learned the rules itself. https://arxiv.org/abs/1911.08265

JetBlue replaces passport and boarding pass with facial recognition to board an international aircraft. https://twitter.com/mackenzief/status/1118509708673998848

Real-time object detection YOLO You Only Look Once running at 30 FPS test dataset: CAM2, the Continuous Analysis of Many CAMeras project, which is built and maintained by Purdue University researchers.

Watch “Engineers Created A New Bionic Arm That Can Grow With You” on YouTube https://youtu.be/luHmXHEpF7w

[1812.02071] Vision-Based High Speed Driving with a Deep Dynamic Observer https://arxiv.org/abs/1812.02071

[1811.10201] InstaNAS: Instance-aware Neural Architecture Search https://arxiv.org/abs/1811.10201 More recently, we've seen a further push to make such so-called 'neural architecture search' (NAS) systems efficient, and approaches like ENAS (Import AI #82) and SMASH (Import AI #56) have shown how to take systems that previously required hundreds of GPUs and fit them onto one or two GPUs.

2019 Jul 11: Google. Massively Multilingual Neural Machine Translation in the Wild: Findings and Challenges https://arxiv.org/abs/1907.05019

2019 Apr 5: Google. Develop AI that can learn to prove math theorems. https://arxiv.org/abs/1904.03241

2018 Sep 14: DeepMind. Meta-learning. Choosing the algorithm. https://arxiv.org/abs/1809.05214

2018 Sep 19: San Francisco and Palo Alto. Planet runs satellites and has large datasets of earth imagery. Orbital Insight makes analtics software to analyze earth imagery, like counting cars and trucks and monitoring construction projects. They share data. https://www.prnewswire.com/news-releases/planet-and-orbital-insight-expand-satellite-imagery-partnership-300714991.html

2018 Aug 8: UC Berkeley. Courteous self-driving cars. A courtesy parameter. We're heading into a future where we deploy autonomous systems into the same environments as humans, so figuring out how to create AI systems that can adapt to human behaviors and account for the peculiarities of people will speed uptake. https://arxiv.org/abs/1808.02633

2017 Dec 22: Researchers with The State Key Laboratory for Management and Control of Complex Systems, within the Chinese Academy of Sciences in Beijing have published details on the 'ParallelEye' dataset, a 3D urban environment modeled on Beijing's Zhongguancun region. They constructed the dataset by grabbing the available Open Street Map (OSM) layout data for a 2km*3km area, then modeled that data using CityEngine, and built the whole environment in the Unity3D engine. https://arxiv.org/abs/1712.08394

2017 Dec 21: NVIDIA release Titan V. This comparison review recommends four 1080 Ti cards instead. https://medium.com/@u39kun/titan-v-vs-1080-ti-head-to-head-battle-of-the-best-desktop-gpus-on-cnns-d55a19866b7c

2017 Dec 20: Automated RNN Architecture Generation, the BC3 cell https://arxiv.org/abs/1712.07316

2017 Dec 16: TorontoCity benchmark dataset. Raquel Urtasun, University of Toronto/Vector Institute/Uber. self-driving cars https://arxiv.org/abs/1612.00423

2017 Dec 5: AlphaZero now plays Chess, Shogi, as well as Go. Learning by self-play. https://arxiv.org/abs/1712.01815

2017 Dec 4: DeepMind Arcade Learning Environment (ALE), which plays Atari games like Breakout, available on GitHub here https://github.com/mgbellemare/Arcade-Learning-Environment

2017 Dec 4: Mozilla offers new voice dataset. https://voice.mozilla.org/data

2017 Nov 30: Amazon AWS gives money to CalTech, to promote MXNet. https://aws.amazon.com/blogs/ai/aws-and-caltech-partner-to-accelerate-ai-and-machine-learning-through-a-new-research-collaboration/

2017 Nov 29: Household Multimodal Environment (HoME) integrates over 45,000 diverse 3D house layouts based on the SUNCG dataset. https://arxiv.org/abs/1711.11017v1

2017 Nov 28: Multi-Level Residual Network (ResNet) with Stacked Simple Recurrent Unit (SRU)s for Action Recognition: https://arxiv.org/abs/1711.08238

2017 Nov 27: from DeepMind Population Based Training (PBT) - optimize a population of models and their hyperparameters hyperparameter selection, such as model architecture, loss function, and optimisation algorithm. By combining multiple steps of gradient descent followed by weight copying by exploit, and perturbation of hyperparameters by explore… https://deepmind.com/blog/population-based-training-neural-networks/

2017 Nov 20: Self-Supervised Intrinsic Image Decomposition Rendered Intrinsics Network (RIN) https://arxiv.org/abs/1711.03678

2017 Nov 16: Jumping and back-flipping robot from Boston Dynamics. https://www.youtube.com/watch?v=fRj34o4hN4I

2017 Nov 13: There is a growing sense that neural networks need to be interpretable to humans. The field of neural network interpretability has formed in response to these concerns. As it matures, two major threads of research have begun to coalesce: feature visualization and attribution. https://distill.pub/2017/feature-visualization/

2017 Nov 12: Extremely Large Minibatch SGD: Training ResNet-50 on ImageNet in 15 Minutes, using 1000 Tesla P100's. https://arxiv.org/abs/1711.04325

2017 10 Nov: Poverty Prediction with Public Landsat 7 Satellite Imagery and Machine Learning: https://arxiv.org/abs/1711.03654

2017 Nov 6: A paper from DeepMind on neural architecture search (NAS) techniques using fractal math. https://arxiv.org/abs/1711.00436

2017 Nov 6: An article from Google Brain on automated design of neural nets. meta learning
https://research.googleblog.com/2017/11/automl-for-large-scale-image.html?utm//source=feedburner&utm_medium=feed&utm//campaign=Feed:+blogspot/gJZg+(Official+Google+Research+Blog)

2017 Nov 6: A paper from DeepMind on biases in semantic learning. “The agent learns words more quickly if the range of words to which it is exposed is limited at first and expanded gradually as its vocabulary develops.” https://arxiv.org/abs/1710.09867

2017 Nov 6: Uber AI releases a programming language Pyro to open source. https://eng.uber.com/pyro/

2017 Oct 30: Meta Learning Shared Hierarchies, an algorithm is able to efficiently learn to navigate mazes by switching between various sub-components, while traditional methods would typically struggle due to the lengthy timesteps required to learn to solve the environment. meta learning https://arxiv.org/abs/1710.09767

2017 Oct 30: The House3D dataset will be announced at ICLR 2018. It consists of 45,622 human-designed scenes of houses, with an average of 8.9 rooms and 1.3 floors per scene. Can be used to train household robots, or room-navigating agents. https://openreview.net/forum?id=rkaT3zWCZ&noteId=rkaT3zWCZ

2017 Oct 30: “Deep learning-based memory systems are currently hard to train and of dubious utility. One candidate memory system is the Neural Turing Machine (NTM), which was introduced by DeepMind in a research paper in 2014. NTMs can let networks - in certain, quite limited tasks - perform tasks with less training and higher accuracy than other systems. Successors, like the Neural GPU, extended the capabilities of the NTM to work on harder tasks, like being able to multiply numbers. Then it was further extended with the Evolvable Neural Turing Machine. Now, researchers in Denmark have proposed the HyperENTM, which uses evolution techniques to figure out how to wire together the memory interface, and connect to an external memory component.” https://arxiv.org/abs/1710.04748

2017 Oct 30: Pretrained.ml is a website that pulls together a bunch of pretrained models, including OpenAI's unsupervised sentiment neuron classifier, along with reference information. http://pretrained.ml/

2017 Oct 30: Geoffrey Hinton released a paper revealing new ideas in his “Capsules” theory, which he has been working on with collaborators at Google Brain in Toronto.
Video from 2014: https://www.youtube.com/watch?v=rTawFwUvnLE
Recently released paper: https://arxiv.org/abs/1710.09829

2017 Oct 23: Google releases Atomic Video Actions AVA video dataset for action recognition, ie, tagging videos. Contains 57,000 distinct video clips, tagged with 80 labels including “hitting”, “martial arts”, “shovel”, “digging”, “lift a person”, “play with kids”, “hug”, and “kiss”. Compares favorably to other benchmarking datasets, e.g., UCF101, ActivityNet, DeepMind’s Kinetics, [Joint-annotated Human Motion Database (JHMDB)](http://jhmdb.is.tue.mpg.de/). https://research.googleblog.com/2017/10/announcing-ava-finely-labeled-video.html

2017 Oct 23: New version of AlphaGo, named AlphaGo Zero, way better than previous versions, learns via self-play only. https://deepmind.com/blog/alphago-zero-learning-scratch/

2017 Oct 16: DeepMind invents Rainbow, a composite system developed by DeepMind that
“cobbles together several recent RL techniques”, like A3C, Prioritized Experience Replay, Dueling Networks, Distributional RL, and so on. …[good results] on the benchmark suite of 57 Atari 2600 games from the Arcade Learning Environment (Bellemare et al. 2013), both in terms of data efficiency and of final performance… Atari 2600 benchmark, Meta Learning https://arxiv.org/abs/1710.02298

2017 Oct 16: “One of the more popular tropes within AI research is that of the paperclip maximizer - the worry that if we build a super-intelligent AI and give it overly simple objectives (eg, make paper clips), it will seek to achieve those objectives to the detriment of everything else.” Here is webgame to do just that. http://www.decisionproblem.com/paperclips/

2017 Oct 16: “Kudos to Ilya Kostrikov at NYU for being so inspired by OpenAI Baselines to re-write the PPO, A3C, and ACKTR algorithms into PyTorch.” https://github.com/ikostrikov/pytorch-a2c-ppo-acktr

2017 Oct 16: Quasi Recurrent Neural Network (QRNN), from Salesforce on PyTorch https://github.com/salesforce/pytorch-qrnn

2017 Oct 16: “AI is basically made of matrix multiplication.” Speed up using 16-bit floating point numbers instead of or along with 32-bit. https://arxiv.org/abs/1710.03740

2017 Oct 16: “While most people are focusing on different evolutionary optimization algorithms when using deep learning (eg, REINFORCE, HYPERNEAT, NEAT, and so on), Ant Colony Optimization (ACO) is an interesting side-channel: you get a bunch of agents - ants_ - to go and explore the problem space and, much like their real world insect counterparts, lay down synthetic pheromones for their other ant chums to follow when they find something that approximates to _food”. https://arxiv.org/abs/1710.03753

2017 Oct 16: “a new multi-agent competitive environment, RoboSumo”, simulated environment https://arxiv.org/abs/1710.03641

2017 Apr 25: Biomimicry, Snake-bot, Butterfly, Woman, Sophia from Hanson Robotics, Jimmy Fallon Tonight Show: http://www.hansonrobotics.com/robot/sophia/