本文翻譯自DENNY BRITZ的 WildML中的一篇文章AI and Deep Learning in 2017 – A Year in Review 。

2017年人工智能最成功的故事是 AlphaGo——一個擊敗了人類頂級圍棋棋手的強化學習智能體。由于圍棋游戲的搜索空間極大，人們曾經(jīng)認為機器學習技術在多年內(nèi)都無法下好圍棋。而現(xiàn)在的 AlphaGo 所向披靡，真是讓人驚喜！

第一版的 AlphaGo 利用人類棋手的棋譜數(shù)據(jù)訓練, 并通過自我對弈和蒙特卡洛樹搜索的方法得到進一步改進。不久之后，名為“AlphaGo Zero”的新版本有了進一步突破， 學會了從零開始，不用任何人類棋譜數(shù)據(jù)，僅僅靠自我對弈就學會了圍棋，并且輕而易舉地擊敗了 AlphaGo 的第一個版本。到2017年年底，AlphaGo Zero 算法拓展成為 Alpha Zero，不僅能掌握圍棋，還能使用相同技術掌握國際象棋和日本將棋。有趣的是，這些程序的棋路甚至能讓最有經(jīng)驗的人類棋手感到驚訝，這激勵著棋手們從 AlphaGo 那里學習并調(diào)整自己的下棋風格。DeepMind 還發(fā)布了一款在線工具，用 AlphaGo 技術幫助人類棋手探索圍棋新玩法。

但圍棋不是我們?nèi)〉弥卮筮M步的唯一游戲?？突仿〈髮W(CMU)的研究人員開發(fā)了一個名為 LIbratus 的人工智能系統(tǒng)，在為期20天的一對一、無限注德州撲克(Heads-up, No-Limit Texas Hold’em)人機對戰(zhàn)中，擊敗了人類頂級撲克玩家。更早之前，由布拉格大學、捷克理工大學和加拿大阿爾伯塔大學的研究人員開發(fā)的 DeepStack 系統(tǒng)，第一次擊敗人類職業(yè)撲克玩家。這兩個人工智能系統(tǒng)都使用一對一的玩法，是兩個玩家之間的游戲，這遠遠比多人參與的撲克游戲容易。后者很可能在2018年取得更多進展。

強化學習的下一個前沿應用可能是更為復雜的多人游戲，包括多人撲克等。DeepMind 正在積極研究讓 AI 學會玩星際爭霸2，為此它們發(fā)布了一個研究環(huán)境；OpenAI 在1v1的 Dota2游戲中已經(jīng)取得初步成果，它們目標是讓 AI 在不遠的將來可以玩5v5的 Dota2。

在監(jiān)督學習中，基于梯度的反向傳播算法表現(xiàn)很好，這一點短期內(nèi)不會變化。然而在強化學習中，進化策略(Evolution Strategies,ES)正卷土重來。原因是，強化學習的數(shù)據(jù)通常不是獨立且分布相同的；這些數(shù)據(jù)的信號誤差更小；而且因為需要探索其規(guī)律，不依賴梯度的算法也可以很好工作。此外，進化算法可以線性拓展至成千上萬的機器，進行極快的并行訓練。只需要成百上千的廉價 CPU 即可訓練，而無需昂貴的GPU。

2017年早些時候，OpenAI 的研究人員證明，進化策略可以達到與深度 Q-learning 等標準強化學習算法相媲美的效果。到2017年年底，Uber 的研究團隊發(fā)布了一篇博客和五篇研究論文，展示了對遺傳算法(Genetic Algorithm,GA)的最新研究。他們使用的遺傳算法極其簡單，無需任何梯度信息，就學會了玩復雜的電子游戲。在 Frostbite 這款游戲中，遺傳算法的得分是10500，而 DQN、AC3和 ES 的得分均不到1000。

2018年，我們極有可能看到更多在進化算法方面的工作進展。

2017年，Google Tacotron 2語音合成系統(tǒng)生成的音頻令人印象深刻。它基于 WaveNet 語音合成技術，這項技術也應用于 Google 助手。WaveNet 的速度在過去一年獲得巨大提升。此前，WaveNet 也曾被應用于機器翻譯，以提高循環(huán)神經(jīng)網(wǎng)絡架構的訓練速度。

2017年，在機器學習領域，遠離需長時間訓練的、昂貴的循環(huán)架構似乎是大勢所趨?！?strong>Attention 就是你需要的全部”，研究人員可以完全擺脫循環(huán)或卷積的架構，使用更復雜的注意力機制，以最低的成本實現(xiàn)最好的結果。

如果簡要總結2017年，那么可以說這是“深度學習框架年”。Facebook 借助深度學習框架 PyTorch，成為攪局者。由于其動態(tài)計算圖的結構和 Chainer 框架的相似，PyTorch 深受 NLP 研究人員的喜愛，因為他們經(jīng)常要處理動態(tài)和循環(huán)的結構，而這些結構很難在 TensorFlow 之類的靜態(tài)圖框架中展示。

2017年，TensorFlow 框架運行得相當好。2017年2月，具有穩(wěn)定和后向兼容 API 的 TensorFlow 1.0發(fā)布。目前 TensorFlow 的版本是1.4.1。在主庫之外，Google 還發(fā)布了幾個 TensorFlow 的伴侶庫，包括可以生成動態(tài)計算圖的 Tensorflow Fold、方便數(shù)據(jù)預處理的 TensorFlow Transform，以及 DeepMind 的 Higher-Level Sonnet。TensorFlow 團隊還公布了一種類似于 PyTorch 動態(tài)計算圖的 eager execution 模式。當然，這一年除了 Google 和 Facebook，許多其他公司也加入了機器學習框架的搭建：

• 蘋果公司發(fā)布 CoreML 2. CoreML 移動機器學習庫

• Uber 的研究團隊發(fā)布了一種深度概率編程語言 Pyro

• Amazon 推出 Gluon , 一個可以在 MXNet 中使用的高級 API

• Uber 公布了其內(nèi)部使用的機器學習基礎平臺 Michelangelo 的詳細信息

由于現(xiàn)有深度學習框架已經(jīng)很多，F(xiàn)acebook、微軟聯(lián)合推出 ONNX 標準，用于跨平臺分享深度學習模型。比如你可以在一個框架中訓練你的模型，然后在讓該模型在另一個框架中工作。

除了通用的深度學習框架，我們還看到大量強化學習框架，其中包括：

• OpenAI，一個用于機器人模擬的開源軟件

• OpenAI Baselines，一組高質量的強化學習算法實現(xiàn)

• Tensorflow Agents，一個框架，可以使用 Tensorflow 對 RL agent 進行優(yōu)化

• Unity ML Agents，允許研究者和開發(fā)者使用 Unity 編輯器創(chuàng)建游戲和模擬使用, 并用強化學習來訓練

• Nervana Coach，允許使用最先進的強化學習算法進行實驗

• Facebook’s ELF Facebook，一個游戲 AI 研究平臺

•  DeepMind Pycolab，一個可定制的 gridworld 游戲引擎

• Geek.ai MAgent，一個游戲 AI 研究平臺，用于多主體(many-agent)游戲的強化學習

我們還看到一些為了讓深度學習更易用的 web 框架, 比如 Google 的 deeplearn.js 和 MIL WebDNN 執(zhí)行框架。但至少有一個曾經(jīng)非常受歡迎的框架已經(jīng)停止維護，那就是 Theano。開發(fā)者在 Theano 郵件列表的公告中宣布，Theano 1.0將是最后一個版本。

2017年，看到許多類似“深度學習技術解決醫(yī)學問題，擊敗人類專家”的炒作。對于沒有醫(yī)療背景的人，要了解 AI 在醫(yī)療領域真正的技術突破絕非易事。想對此有全面了解，推薦閱讀 Luke Oakden-Ranyner 的《人類醫(yī)生的終結》系列文章(https://lukeoakdenrayner./2017/04/20/the-end-of-human-doctors-introduction/)。這里我僅簡要介紹一些進展：

今年醫(yī)療領域最熱門的新聞是斯坦福大學的一個研究團隊發(fā)布的深度學習算法，這個算法可以幫皮膚科醫(yī)生識別皮膚癌。斯坦福大學的另一個研究小組開發(fā)了一種可以診斷心律不齊的模型，可以比心臟病專家更好地從單導聯(lián)心電圖信號中診斷出心律失常。

但是今年并非沒有遺憾。 Deepmind 與英國 NHS 的數(shù)據(jù)共享協(xié)議充滿了“不可原諒的”錯誤。美國國家衛(wèi)生研究院向科學界公布了一個胸部 x 射線數(shù)據(jù)集, 但更細致的研究發(fā)現(xiàn)，它并不真正適合訓練用于疾病診斷的 AI 模型。

今年開始越來越受關注的應用還有圖片、音樂、繪畫和視頻的生成模型。 2017年的 NIPS 會議首次舉辦了機器學習創(chuàng)造力與設計研討會(Machine Learning for Creativity and Design workshop)。

這類應用程序中最流行的是 Google 的 QuickDraw，它使用一個神經(jīng)網(wǎng)絡來識別你的涂鴉是什么。使用公開數(shù)據(jù)集，你甚至可以讓機器為你畫畫。

生成式對抗網(wǎng)絡(GANs)今年取得了重大進展。例如，CycleGAN、DiscoGAN 和 StarGAN 等在人臉模型方面取得了令人矚目的成果。傳統(tǒng)上，GANs 很難生成真實的高分辨率圖像，但 pix2pixHD 令人印象深刻的結果表明，我們正在解決這些問題。所以 GANs 會成為新的畫筆嗎？

自動駕駛汽車的大玩家是共享汽車公司 Uber 和 Lyft，以及 Alphabet 的 Waymo 和 Tesla。Uber 在今年年初遭遇一些挫折，他們的自動駕駛汽車在舊金山闖了幾次紅燈，原因是軟件錯誤，而非之前報道所說的人為失誤。事后 Uber 分享了其內(nèi)部使用的汽車可視化平臺的詳細信息。到12月，Uber 的自動駕駛汽車項目已經(jīng)達到了200萬英里的里程。

與此同時，Waymo 的自動駕駛汽車在四月開始了人類駕駛員的第一次試行，后來在亞利桑那州的鳳凰城試行時，則完全沒有了人類駕駛員。Waymo 還公布了他們的測試和模擬技術的細節(jié)。

Lyft 宣布正在開發(fā)自動駕駛的硬件和軟件，它在波士頓的第一個試驗點已經(jīng)啟用。特斯拉的 Autpilot 沒有看到太多的更新，但是這個領域又加入了一個新人：Apple。蒂姆·庫克(Tim Cook)證實，Apple 公司正致力于開發(fā)自動駕駛汽車的軟件，Apple 的研究人員已經(jīng)在 arXiv 發(fā)表了一篇關于與地圖相關的論文。

在過去一整年，一些研究人員對學術論文結果的可重復性提出了擔憂。深度學習模型通常依賴大量的超參數(shù)，必須優(yōu)化這些參數(shù)，才能獲得足夠好的結果并發(fā)表。這種參數(shù)優(yōu)化的成本可能越來越高，以至于只有 Google 和 Facebook 這樣的公司才能負擔得起。研究人員并不總會公開他們的代碼，有時他們會忘記把重要的細節(jié)寫進論文，或者使用略微不同的評估步驟，或者通過在同一個分裂點上反復優(yōu)化超參數(shù)來適應數(shù)據(jù)集，從而出現(xiàn)過擬合，這些使得可重復性成為一個大問題。在強化學習研究中，研究人員發(fā)現(xiàn)同樣的算法從不同的代碼庫中獲得的結果大不相同，而且差異很大:

在《GANs 是否造成了不公？一項大型研究》(Are GANs Created Equal? A Large-Scale Study)一文中，研究人員發(fā)現(xiàn)調(diào)整良好的 GAN 如果使用昂貴的超參數(shù)搜索，就可以擊敗那些自稱最優(yōu)的更復雜的方法。同樣, 在《神經(jīng)語言模型評價的現(xiàn)狀》(On the State of the Art of Evaluation in Neural Language Models)一文中，研究人員表示，簡單的 LSTM 架構，如果經(jīng)過適當?shù)囊?guī)范和調(diào)整, 也可以勝過最新的模型。

在一次 NIPS 演講中，Ali Rahimi 把最近深度學習方法和中世紀的煉金術做了比較，呼吁更嚴格的實驗設計，引起研究人員的廣泛共鳴。Yann LeCun 第二天立即作出回應，認為這是對研究者的冒犯。

隨著美國移民政策的收緊，對國際人才需求很大的科技公司正越來越多地在海外開設辦公室，加拿大是首選目的地。Google 在多倫多開了一間新的辦公室，DeepMind  在加拿大的埃德蒙頓也開了一間新辦公室，F(xiàn)acebook AI 研究院也在向加拿大蒙特利爾擴張。

中國是另一個受到廣泛關注的目的地。由于中國擁有大量資金、龐大的人才庫以及隨時可以獲得的政府數(shù)據(jù)，在人工智能開發(fā)和生產(chǎn)部署方面，它已經(jīng)與美國展開了正面競爭。Google 還宣布不久將在北京開設一個新的實驗室。

眾所周知，現(xiàn)代深度學習技術需要昂貴的 GPU 來訓練最先進的模型。截至目前，NVIDIA 是最大贏家。今年，它公布了泰坦5號旗艦 GPU。順便說一句, 它是金色的。

但競爭正在加劇。Google 的 TPUs 現(xiàn)在可以在其云平臺上使用，英特爾下屬 Nervana 公司推出了一套新的芯片, 甚至 Tesla 也承認它正在開發(fā)自己的人工智能硬件。競爭也可能來自中國，那里專門從事比特幣采礦的硬件制造商，也希望進入 GPU 領域。

過度炒作要擔責。主流媒體的報道幾乎從不符合實驗室或工業(yè)界的實際情況。 IBM Watson 就是一個過度營銷的典型代表。2017年，每個人都討厭 IBM Watson，這并不奇怪，因為它被宣傳得很厲害，但在醫(yī)療方面卻屢次失敗。

最引人注目的可能是 Facebook 的“研究人員關閉了發(fā)明自己語言的人工智能”這件事，我并不了解媒體這樣報道的真實意圖是什么。但它已經(jīng)造成很大的負面影響，你甚至可以在 Google 上搜索到相關報道。當然了，這個標題與事實相去甚遠。實際情況只是研究者停止了一項結果不理想的實驗。

但是，對炒作負有責任的不僅僅是媒體。確實也有一部分研究者突破了底線，在論文中使用不能反映實際實驗結果的標題和摘要，比如這篇自然語言研究論文(https:///@yoav.goldberg/an-adversarial-review-of-adversarial-generation-of-natural-language-409ac3378bd7)，和這篇機器學習研究論文(http:///2017/08/06/fitting-to-noise-or-nothing-at-all-machine-learning-in-markets/)。

Coursera 的聯(lián)合創(chuàng)始人 Andrew Ng（吳恩達）今年多次出現(xiàn)在新聞報道中，他因為機器學習 MOOC 課程視頻而出名。今年3月，Andrew Ng 離開了由自己擔任 AI 團隊領導者的百度公司，籌集1.5億美元資金，成立了名為 landing.AI 的創(chuàng)業(yè)公司，專注于制造業(yè)。其他這類新聞還有：Gary Marcus 辭去了 Uber 人工智能實驗室的主管職務，F(xiàn)acebook 雇傭了 Siri 的自然語言理解主管，幾位杰出的研究人員離開了 OpenAI，開辦了一家新的機器人公司。

學術界流失科學家的趨勢也在繼續(xù)。大學實驗室紛紛抱怨：他們能為 AI 人才提供的薪水，無法與工業(yè)巨頭競爭。

以下為英文原文，點擊閱讀原文也可查看。

The year is coming to an end. I did not write nearly as much as I had planned to. But I’m hoping to change that next year, with more tutorials around Reinforcement Learning, Evolution, and Bayesian Methods coming to WildML! And what better way to start than with a summary of all the amazing things that happened in 2017? Looking back through my Twitter history and the WildML newsletter, the following topics repeatedly came up. I’ll inevitably miss some important milestones, so please let me know about it in the comments!

Reinforcement Learning beats humans at their own games

The biggest success story of the year was probably AlphaGo (Nature paper), a Reinforcement Learning agent that beat the world’s best Go players. Due to its extremely large search space, Go was thought to be out of reach of Machine Learning techniques for a couple more years. What a nice surprise!

The first version of AlphaGo was bootstrapped using training data from human experts and further improved through self-play and an adaptation of Monte-Carlo Tree Search. Soon after, AlphaGo Zero (Nature Paper) took it a step further and learned to play Go from scratch, without human training data whatsoever, using a technique simultaneously published in the Thinking Fast and Slow with Deep Learning and Tree Search paper. It also handily beat the first version of AlphaGo. Towards the end of the year, we saw yet another generalization of the AlphaGo Zero algorithm, called AlphaZero, which not only mastered Go, but also Chess and Shogi, using the exact same techniques. Interestingly, these programs made moves that surprised even the most experienced Go players, motivating players to learn from AlphaGo and adjusting their own play style accordingly. To make this easier, DeepMind also released an AlphaGo Teach tool.

But Go wasn’t the only game where we made significant progress. Libratus (Science paper), a system developed by researchers from CMU, managed to beat top Poker players in a 20-day Heads-up, No-Limit Texas Hold’em tournament. A little earlier, DeepStack, a system developed by researchers from Charles University, The Czech Technical University, and the University of Alberta, became the first to beat professional poker players. Note that both of these systems played Heads-up poker, which is played between two players and a significantly easier problem than playing at a table of multiple players. The latter will most likely see additional progress in 2018.

The next frontiers for Reinforcement Learning seem to be more complex multi-player games, including multi-player Poker. DeepMind is actively working on Starcraft 2, releasing a research environment, and OpenAI demonstrated initial success in 1v1 Dota 2, with the goal of competing in the the full 5v5 game in the near future.

Evolution Algorithms make a Comeback

For supervised learning, gradient-based approaches using the back-propagation algorithm have been working extremely well. And that isn’t likely to change anytime soon. However, in Reinforcement Learning, Evolution Strategies (ES) seem to be making a comeback. Because the data typically is not iid (independent and identically distributed), error signals are sparser, and because there is a need for exploration, algorithms that do not rely on gradients can work quite well. In addition, evolutionary algorithms can scale linearly to thousands of machines enabling extremely fast parallel training. They do not require expensive GPUs, but can be trained on a large number (typically hundreds to thousands) of cheap CPUs.

Earlier in the year, researchers from OpenAI demonstrated that Evolution Strategies can achieve performance comparable to standard Reinforcement Learning algorithms such as Deep Q-Learning. Towards the end of the year, a team from Uber released a blog post and a set of five research papers, further demonstrating the potential of Genetic Algorithms and novelty search. Using an extremely simple Genetic Algorithm, and no gradient information whatsoever, their algorithm learns to play difficult Atari Games. Here’s a video of the GA policy scoreing 10,500 on Frostbite. DQN, AC3, and ES score less than 1,000 on this game.

Most likely, we’ll see more work in this direction in 2018.

WaveNets, CNNs, and Attention Mechanisms

Google’s Tacotron 2 text-to-speech system produces extremely impressive audio samples and is based on WaveNet, an autoregressive model which is also deployed in the Google Assistant and has seen massive speed improvements in the past year. WaveNet had previously been applied to Machine Translation as well, resulting in faster training times that recurrent architectures.

The move away from expensive recurrent architectures that take long to train seems to be larger trend in Machine Learning subfields. In Attention is All you Need, researchers get rid of recurrence and convolutions entirely, and use a more sophisticated attention mechanism to achieve state of the art results at a fraction of the training costs.

The Year of Deep Learning frameworks

If I had to summarize 2017 in one sentence, it would be the year of frameworks. Facebook made a big splash with PyTorch. Due to its dynamic graph construction similar to what Chainer offers, PyTorch received much love from researchers in Natural Language Processing, who regularly have to deal with dynamic and recurrent structures that hard to declare in a static graph frameworks such as Tensorflow.

Tensorflow had quite a run in 2017. Tensorflow 1.0 with a stable and backwards-compatible API was released in February. Currently, Tensorflow is at version 1.4.1. In addition to the main framework, several Tensorflow companion libraries were released, including Tensorflow Fold for dynamic computation graphs, Tensorflow Transform for data input pipelines, and DeepMind’s higher-level Sonnet library. The Tensorflow team also announced a new eager execution mode which works similar to PyTorch’s dynamic computation graphs.

In addition to Google and Facebook, many other companies jumped on the Machine Learning framework bandwagon:

• Apple announced its CoreML mobile machine learning library.

• A team at Uber released Pyro, a Deep Probabilistic Programming Language.

• Amazon announced Gluon, a higher-level API available in MXNet.

• Uber released details about its internal Michelangelo Machine Learning infrastructure platform.

And because the number of framework is getting out of hand, Facebook and Microsoft announced the ONNX open format to share deep learning models across frameworks. For example, you may train your model in one framework, but then serve it in production in another one.

In addition to general-purpose Deep Learning frameworks, we saw a large number of Reinforcement Learning frameworks being released, including:

• OpenAI Roboschool is an open-source software for robot simulation.

• OpenAI Baselines is a set of high-quality implementations of reinforcement learning algorithms.

• Tensorflow Agents contains optimized infrastructure for training RL agents using Tensorflow.

• Unity ML Agents allows researchers and developers to create games and simulations using the Unity Editor and train them using Reinforcement Learning.

• Nervana Coach allows experimentation with state of the art Reinforcement Learning algorithms.

• Facebook’s ELF platform for game research.

• DeepMind Pycolab is a customizable gridworld game engine.

• Geek.ai MAgent is a research platform for many-agent reinforcement learning.

With the goal of making Deep Learning more accessible, we also got a few frameworks for the web, such as Google’s deeplearn.js and the MIL WebDNN execution framework. But at least one very popular framework died. That was Theano. In an announcement on the Theano mailing list, the developers decided that 1.0 would be its last release.

Applications: AI & Medicine

2017 saw many bold claims about Deep Learning techniques solving medical problems and beating human experts. There was a lot of hype, and understanding true breakthroughs is anything but easy for someone not coming from a medical background. For an comprehensive review, I recommend Luke Oakden-Rayner’s The End of Human Doctors blog post series. I will briefly highlight some developments here.

Among the top news this year was a Stanford team releasing details about a Deep learning algorithm that does as well as dermatologists in identifying skin cancer. You can read the Nature article here. Another team at Stanford developed a model which can diagnose irregular heart rhythms, also known as arrhythmias, from single-lead ECG signals better than a cardiologist.

But this year was not without blunders. DeepMind’s deal with the NHS was full of “inexcusable” mistakes. The NIH released a chest x-ray dataset to the scientific community, but upon closer inspection it was found that it is not really suitable for training diagnostic AI models.

Applications: Art & GANs

Another application that started to gain more traction this year is generative modeling for images, music, sketches, and videos. The NIPS 2017 conference featured a Machine Learning for Creativity and Design workshop the first time this year.

Among the most popular applications was Google’s QuickDraw, which uses a neural network to recognize your doodles. Using the released dataset you may even teach machines to finish your drawings for you.

Generative Adversarial Networks (GANs), made significant progress this year. New models such as CycleGAN, DiscoGAN and StarGAN achieved impressive results in generating faces, for example. GANs traditionally have had difficulty generating realistic high-resolution images, but impressive results from pix2pixHD demonstrate that we’re on track to solving these. Will GANs become the new paintbrush?

Applications: Self-driving cars

The big players in the self-driving car space are ride-sharing apps Uber and Lyft, Alphabet’s Waymo, and Tesla. Uber started out the year with a few setbacks as their self-driving cars missed several red lights in San Francisco due to software error, not human error as had been reported previously. Later on, Uber shared details about its car visualization platform used internally. In December, Uber’s self driving car program hit 2 million miles.

In the meantime, Waymo’s self-driving cars got their first real riders in April, and later completely took out the human operators in Phoenix, Arizona. Waymo also published details about their testing and simulation technology.

A Waymo simulation showing improved vehicle navigation

Lyft announced that it is building its own autonomous driving hard- and software. Its first pilot in Boston is now underway. Tesla Autpilot hasn’t seen much of an update, but there’s a newcomer to the space: Apple. Tim Cook confirmed that Apple is working on software for self-driving cars, and researchers from Apple published a mapping-related paper on arXiv.

Deep Learning, Reproducibility, and Alchemy

Throughout the year, several researchers raised concerns about the reproducibility of academic paper results. Deep Learning models often rely on a huge number of hyperparameters which must to be optimized in order to achieve results that are good enough to publish. This optimization can become so expensive that only companies such as Google and Facebook can afford it. Researchers do not always release their code, forget to put important details into the finished paper, use slightly different evaluation procedures, or overfit to the dataset by repeatedly optimizing hyperparameters on the same splits. This makes reproducibility a big issue. In Reinforcement Learning That Matters, researchers showed that the same algorithms taken from different code bases achieve vastly different results with high variance:

In Are GANs Created Equal? A Large-Scale Study, researchers showed that a well-tuned GAN using expensive hyperparameter search can beat more sophisticated approaches that claim to be superior. Similarly, in On the State of the Art of Evaluation in Neural Language Models, researchers showed that simple LSTM architectures, when properly regularized and tuned, can outperform more recent models.

In a NIPS talk that resonated with many researchers, Ali Rahimi compared recent Deep Learning approaches to Alchemy and called for more rigorous experimental design. Yann LeCun took it as an insult and promptly respondedthe next day.

Artificial Intelligence made in Canada and China

With United States immigration policies tightening, it seems that companies are increasingly opening offices overseas, with Canada being a prime destination. Google opened a new office in Toronto, DeepMind opened a new office in Edmonton, Canada, and Facebook AI Research is expanding to Montreal as well.

China is another destination that is receiving a lot of attention. With a lot of capital, a large talent pool, and government data readily available, it is competing head to head with the United States in terms of AI developments and production deployments. Google also announced that it will soon open a new lab in Beijing.

Hardware Wars: Nvidia, Intel, Google, Tesla

Modern Deep Learning techniques famously require expensive GPUs to train state-of-the-art models. So far, NVIDIA has been the big winner. This year, it announced its new Titan V flagship GPU. It comes in gold color, by the way.

But competition is increasing. Google’s TPUs are now available on its cloud platform, Intel’s Nervana unveiled a new set of chips, and even Tesla admitted that it is working on its own AI hardware. Competition may also come from China, where hardware makers specializing in Bitcoin mining want to enter the Artificial Intelligence focused GPU space.

Hype and Failures

With great hype comes great responsibility. What the mainstream media reports almost never corresponds to what actually happened in a research lab or production system. IBM Watson is the poster-child over overhyped marketing and failed to deliver corresponding results. This year, everyone was hating on IBM Watson, which is not surprising after its repeated failures in healthcare.

The story capturing the most hype was probably Facebook’s “Researchers shut down AI that invented its own language”, which I won’t link to on purpose. It has already done enough damage and you can google it. Of course, the title couldn’t have been further from the truth. What happened was researchers stopping a standard experiment that did not seem to give good results.

But it’s not only the press that is guilty of hype. Researchers also overstepped boundaries with titles and abstracts that do not reflect the actual experiment results, such as in this natural language generation paper, or this Machine Learning for markets paper.

High-Profile Hires and Departures

Andrew Ng, the Coursera co-founder who is probably most famous for his Machine Learning MOOC, was in the news several times this year. Andrew left Baidu where he was leading the AI group in March, raised a new $150M fund, and announced a new startup, landing.ai, focused on the manufacturing industry. In other news, Gary Marcus stepped down as the director of Uber’s artificial intelligence lab, Facebook hired away Siri’s Natural Language Understanding Chief, and several prominent researchers left OpenAI to start a new robotics company.

The trend of Academia losing scientists to the industry also continued, with university labs complaining that they cannot compete with the salaries offered by the industry giants.

And finally, Happy New Year! Thanks for sticking with this post for so long :)