The 2015 FLI Open Letter on Robust and Beneficial AI
A 2015 Future of Life Institute open letter, signed by thousands of researchers, called for research to keep AI robust and beneficial.
Keras gives deep learning a human-friendly API
Francois Chollet released Keras in March 2015, a high-level deep learning API whose ease of use brought neural networks to a much wider audience.
Google Photos searches your pictures by what is in them
Google launches Photos with free unlimited storage and machine-learning search that finds images by people, places, and objects without tags.
DARPA Robotics Challenge Finals
DARPA's disaster-response robot contest reached its finals; Team KAIST's DRC-Hubo won, showing how hard real-world humanoid robotics still was.
Google's DeepDream turns neural networks into image generators
Google researchers showed that image-classifying networks could run in reverse to generate dreamlike imagery, the technique they called Inceptionism.
Spotify launches Discover Weekly, a personalized playlist for everyone
Spotify rolls out Discover Weekly, a 30-song playlist refreshed every Monday for each listener using recommendation algorithms.
AI researchers' open letter against autonomous weapons
In 2015 thousands of AI researchers signed an open letter calling for a ban on offensive autonomous weapons beyond meaningful human control.
Neural style transfer repaints photos in the style of famous artists
Gatys, Ecker, and Bethge showed a neural network could separate the content of one image from the style of another and recombine them.
Gmail suggests your replies with a neural network
Google launches Smart Reply, a sequence-to-sequence neural network that proposes short email responses, putting generative AI into everyday email.
Google open-sources TensorFlow
Google released TensorFlow, its internal machine-learning system, as free open-source software in 2015 - an industrial-grade deep learning framework for all.
ResNet trains ultra-deep networks
Residual connections let neural networks grow to 152 layers and win ImageNet 2015, unlocking much deeper architectures.
OpenAI is founded
OpenAI launched in December 2015 as a nonprofit AI lab with a 1 billion dollar commitment and a mission to ensure AGI benefits all of humanity.
Jill Watson, the AI teaching assistant students could not spot
In 2016 Georgia Tech ran an IBM Watson-based teaching assistant named Jill Watson that fielded forum questions without students noticing.
GLAD forest alerts bring near-real-time deforestation detection
University of Maryland's GLAD alerts use Landsat imagery and decision-tree models to flag tropical deforestation at 30-meter resolution.
AlphaGo defeats Lee Sedol
DeepMind's AlphaGo beat Go world champion Lee Sedol 4-1, a feat experts had thought was a decade away.
Facebook ships automatic alt text for blind users
In April 2016 Facebook used object recognition to auto-describe photos for blind users, one of the first AI accessibility features at scale.
NVIDIA builds the DGX-1 and hand-delivers the first one to OpenAI
NVIDIA announced the DGX-1 deep learning supercomputer in 2016 and Jensen Huang hand-delivered the first unit to the young OpenAI lab.
The Next Rembrandt prints a new 'Rembrandt' from data
A bank-and-Microsoft project analyzed all 346 Rembrandt paintings to 3D-print a new portrait in his style, an early data-driven art stunt.
DARPA christens Sea Hunter, an autonomous warship
In April 2016 DARPA christened Sea Hunter, a crewless trimaran built to track submarines for months at sea on its own.
GDPR Article 22 Limits Automated Decision-Making (2016)
The EU's data protection regulation gives people a right not to be subject to solely automated decisions with legal or similarly significant effects.
OpenAI releases Gym, a toolkit for reinforcement learning
OpenAI Gym gave reinforcement learning a common interface and a shared set of benchmark environments, standardizing how RL is tested.
AEGIS lets a Mars rover choose its own laser targets
From May 2016 NASA's AEGIS software let the Curiosity rover autonomously pick rock targets for its ChemCam laser without waiting for Earth.
AEGIS lets a Mars rover pick its own laser targets
From May 2016 NASA's Curiosity rover used AEGIS to choose its own ChemCam laser targets on Mars without waiting for Earth.
Google reveals the Tensor Processing Unit
Google disclosed the Tensor Processing Unit, a custom neural-network chip already running in its data centers, starting the custom-silicon era for AI.
Google launches Magenta to make art and music with machine learning
Google Brain started Magenta, an open-source research effort asking whether machine learning could create compelling art and music.
Sunspring: a short film written by a neural network
Sunspring, a 2016 sci-fi short, had its screenplay written by an LSTM neural network named Benjamin, an early AI-authored film.
State v. Loomis: a court upholds sentencing with a secret risk algorithm
Wisconsin's top court allowed COMPAS risk scores at sentencing despite the algorithm being a proprietary secret, with cautions.
DeepMind cuts Google data center cooling energy by 40 percent
In 2016 DeepMind applied neural networks to Google data center operations and cut the energy used for cooling by up to 40 percent.
DARPA Cyber Grand Challenge Final
The first all-machine hacking tournament, where autonomous systems found and patched software flaws in real time without humans.
The Washington Post deploys Heliograf for Olympics and election coverage
The Washington Post launched its in-house Heliograf system to auto-generate short news updates during the 2016 Rio Olympics and US election.
Hello Games launches No Man's Sky
Hello Games launches No Man's Sky, a deterministic procedural universe of 18 quintillion planets from one seed.
YouTube describes its deep-learning recommendation system
A YouTube paper details the deep neural network behind its recommendations, a two-stage candidate-generation and ranking design copied across the industry.
Google Translate switches to neural machine translation
Google deploys GNMT, a deep neural translation system that cuts translation errors by 55 to 85 percent on major language pairs.
Google launches the Assistant across phones and the home
Google ships the Google Assistant in the Pixel phone and Google Home speaker, a conversational AI built on its search and machine learning.
Team Delft wins Amazon Picking Challenge
The Amazon Picking Challenge pushed warehouse grasping robots; Team Delft won the 2016 picking and stowing events using deep-learning perception.
Google's zero-shot translation and the hint of a neural interlingua
In 2016 Google showed one neural model could translate between language pairs it never trained on, hinting at an emergent interlingua.
Mastercard rolls out Decision Intelligence, AI fraud scoring across its network
Mastercard launched Decision Intelligence, which it called the first use of AI at global scale on its network, to cut false declines.
Memo Akten's Learning to See turns a webcam into machine dreams
Artist Memo Akten's neural-network installation reinterprets a live camera feed in real time, dramatizing how a machine 'sees.'
PyTorch released by Facebook AI Research
Facebook AI Research released PyTorch, a Python-first, define-by-run deep learning framework that became the dominant tool for AI research.
Stanford CNN matches dermatologists at classifying skin cancer
A 2017 Nature paper trained a CNN on 129,450 images to classify skin cancer on par with 21 board-certified dermatologists across key diagnostic tasks.
Libratus beats poker professionals
Carnegie Mellon's Libratus defeated four top professionals at heads-up no-limit poker, the first AI to master a major game of imperfect information.
DeepStack reaches expert poker with deep learning
DeepStack beat professional players at heads-up no-limit poker using deep learning and local search instead of a full precomputed strategy.
Samsung introduces Bixby
Samsung unveils Bixby, a voice and touch interface for the Galaxy S8 built on completeness, context awareness, and cognitive tolerance.
Pan-Canadian AI Strategy
Canada launched the world's first national AI strategy in 2017, funding the Amii, Mila, and Vector institutes through CIFAR.
NVIDIA Volta and the V100: Tensor Cores Arrive
NVIDIA's Volta architecture and Tesla V100 introduced Tensor Cores, dedicated units for the matrix math behind deep learning.
Woebot puts a CBT chatbot to a randomized clinical trial
A 2017 randomized trial found a CBT chatbot, Woebot, cut depression symptoms in college students over two weeks versus an information control.
The Transformer is introduced
Google researchers publish 'Attention Is All You Need', introducing the Transformer architecture that underpins modern AI.
Baidu opens the Apollo self-driving platform (2017)
Baidu open-sourced Apollo in 2017, an autonomous-driving software stack it pitched as a shared platform for the whole industry.