My colleagues and I are fascinated by the science of moving dots. So what are these dots? Well, it's all of us. And we're moving in our homes, in our offices, as we shop and travel throughout our cities and around the world. And wouldn't it be great if we could understand all this movement? If we could find patterns and meaning and insight in it. And luckily for us, we live in a time where we're incredibly good at capturing information about ourselves. So whether it's through sensors or videos, or apps, we can track our movement with incredibly fine detail.
我同事和我對移動的點 背後的科學非常著迷。 這些點是什麼? 是我們每一個人。 我們在家、在辦公室裡走動, 在整個城市裡、 甚至世界各地旅遊和購物。 我們若能夠瞭解這些動作, 不是很棒嗎? 如果能從中找出模式、 意義和背後意涵的話。 而我們有幸 活在一個擷取自身資訊 非常容易的時代, 無論透過感應器、影片或應用程式, 我們都能非常精細地追蹤自己的動作。
So it turns out one of the places where we have the best data about movement is sports. So whether it's basketball or baseball, or football or the other football, we're instrumenting our stadiums and our players to track their movements every fraction of a second. So what we're doing is turning our athletes into -- you probably guessed it -- moving dots.
然後我們發現,最適合 獲取動作資訊的一個地方, 就是運動場。 無論是籃球、棒球、橄欖球或足球, 我們都能在場館裡、甚至球員身上 安裝儀器,追蹤他們的動作 -- 每個瞬間的動作。 所以我們要做的 就是把運動員變成── 你們大概已經猜到了── 移動的點。
So we've got mountains of moving dots and like most raw data, it's hard to deal with and not that interesting. But there are things that, for example, basketball coaches want to know. And the problem is they can't know them because they'd have to watch every second of every game, remember it and process it. And a person can't do that, but a machine can. The problem is a machine can't see the game with the eye of a coach. At least they couldn't until now. So what have we taught the machine to see?
我們收集了海量的移動的點, 但就像大部分的原始數據一樣, 它們不易處理、也不甚有趣。 但藏在數據裡的是,舉例來說, 籃球教練想知道的事。 而問題是他們無法得知這些事, 因為他們得看著每一秒鐘的比賽, 記住內容並處理它。 沒有任何人能做到這件事, 但機器可以。 問題是,機器無法從 教練的角度觀看比賽, 至少在此之前,它們都做不到。 所以我們教會了機器看些什麼呢?
So, we started simply. We taught it things like passes, shots and rebounds. Things that most casual fans would know. And then we moved on to things slightly more complicated. Events like post-ups, and pick-and-rolls, and isolations. And if you don't know them, that's okay. Most casual players probably do. Now, we've gotten to a point where today, the machine understands complex events like down screens and wide pins. Basically things only professionals know. So we have taught a machine to see with the eyes of a coach.
從簡單的開始。 我們教它判斷傳球、 投籃、搶籃板等動作, 一些大部分普通球迷都知道的事。 然後我們進入稍微複雜一點的動作, 像是低位單打、擋切和清空單打。 如果你不知道這些動作,沒關係。 打球的人大概都清楚。 接著,我們到達今天的地步, 機器已經可以讀出複雜的動作, 例如:向下掩護和無球掩護(wide pin), 一些基本上是專業人士才懂的動作。 我們教會了機器 用教練的角度來看比賽。
So how have we been able to do this? If I asked a coach to describe something like a pick-and-roll, they would give me a description, and if I encoded that as an algorithm, it would be terrible. The pick-and-roll happens to be this dance in basketball between four players, two on offense and two on defense. And here's kind of how it goes. So there's the guy on offense without the ball the ball and he goes next to the guy guarding the guy with the ball, and he kind of stays there and they both move and stuff happens, and ta-da, it's a pick-and-roll.
我們是怎麼做到的? 如果我請教練形容 某個動作,例如擋切, 他們會給我一段敘述, 如果我把這個敘述寫成 一個演算法,大概會慘不忍睹。 擋切在籃球中,恰如 四個球員之間的舞蹈, 兩個進攻方、兩個防守方。 大概是這樣的過程: 有一個未持球的進攻球員 他跑到持球球員的旁邊幫他掩護, 在原地稍作停留, 然後他們都移動,事情就發生了 ——嗒啦——這就是擋拆。
(Laughter)
(笑聲)
So that is also an example of a terrible algorithm. So, if the player who's the interferer -- he's called the screener -- goes close by, but he doesn't stop, it's probably not a pick-and-roll. Or if he does stop, but he doesn't stop close enough, it's probably not a pick-and-roll. Or, if he does go close by and he does stop but they do it under the basket, it's probably not a pick-and-roll. Or I could be wrong, they could all be pick-and-rolls. It really depends on the exact timing, the distances, the locations, and that's what makes it hard. So, luckily, with machine learning, we can go beyond our own ability to describe the things we know.
剛剛我示範了一個差勁的演算法。 如果擋人那名球員──我們稱掩護者── 慢慢靠近,但他並未停留, 這可能就不是擋切。 或是他有停留,但距離不夠近, 這可能也不是擋切。 或是他慢慢靠近、也確實停留, 但這發生在籃下,可能也不是擋切。 也可能我判斷錯誤,這些全都是擋切。 這一切取決於精確的 時機、距離和位置, 這也正是困難的地方。 很幸運地,透過機器學習技術, 我們得以用超越自己的能力, 來描述我們知道的事情。
So how does this work? Well, it's by example. So we go to the machine and say, "Good morning, machine. Here are some pick-and-rolls, and here are some things that are not. Please find a way to tell the difference." And the key to all of this is to find features that enable it to separate. So if I was going to teach it the difference between an apple and orange, I might say, "Why don't you use color or shape?" And the problem that we're solving is, what are those things? What are the key features that let a computer navigate the world of moving dots? So figuring out all these relationships with relative and absolute location, distance, timing, velocities -- that's really the key to the science of moving dots, or as we like to call it, spatiotemporal pattern recognition, in academic vernacular. Because the first thing is, you have to make it sound hard -- because it is.
如何做到這個技術?舉個例。 我們跑到機器面前說:「早安,機器。 這裡有些擋切的例子, 也有些不是的例子。 請找出區分它們的方式。」 而一切的關鍵在於找到 可以進行區別的特徵。 如果我要教它區分 蘋果和橘子的差異, 我可能會說: 「不妨用顏色或形狀來區分?」 而我們要解決的問題是, 類似這樣的特徵是什麼? 哪些是能讓電腦判讀這些移動的點時, 能暢行無阻的重要特徵? 所以搞清楚每件事情的關聯── 包含相對和絕對位置、 距離、時機、速率── 絕對是研究這些移動的點的重要關鍵, 或者用我們喜歡的稱呼方式: 「時空模式識別」這樣的學術用語。 因為最重要的是, 必須讓它聽起來很難, 因為它真的很難。
The key thing is, for NBA coaches, it's not that they want to know whether a pick-and-roll happened or not. It's that they want to know how it happened. And why is it so important to them? So here's a little insight. It turns out in modern basketball, this pick-and-roll is perhaps the most important play. And knowing how to run it, and knowing how to defend it, is basically a key to winning and losing most games. So it turns out that this dance has a great many variations and identifying the variations is really the thing that matters, and that's why we need this to be really, really good.
重要的是,對NBA教練來說, 他們想知道的 不是擋切是否發生, 而是擋切是怎麼發生的。 為什麼這對他們如此重要? 這裡有一些發現。 原來在現代籃球中, 擋切可能就是最重要的戰術。 知道如何執行、如何防守擋切, 基本上是大部分比賽輸贏的關鍵。 所以結果是,擋切這種舞步變化多端, 真正重要的是要辨別這些變化, 所以我們需要有非常完善的演算法。
So, here's an example. There are two offensive and two defensive players, getting ready to do the pick-and-roll dance. So the guy with ball can either take, or he can reject. His teammate can either roll or pop. The guy guarding the ball can either go over or under. His teammate can either show or play up to touch, or play soft and together they can either switch or blitz and I didn't know most of these things when I started and it would be lovely if everybody moved according to those arrows. It would make our lives a lot easier, but it turns out movement is very messy. People wiggle a lot and getting these variations identified with very high accuracy, both in precision and recall, is tough because that's what it takes to get a professional coach to believe in you. And despite all the difficulties with the right spatiotemporal features we have been able to do that.
這裡有個例子。 兩個進攻球員和兩個防守球員 準備進行擋切的攻防, 持球者可以選擇利用或拒絕擋切, 他的隊友則可以選擇切入或後撤。 防守持球者的球員可選擇 從前繞開、或從後繞開。 他的隊友可選擇上前補防、 隨球盯人或向後消極防守, 他們也可以選擇換防或夾擊。 一開始我並不知道大部分的動作, 覺得如果大家能照那些箭頭 的方向移動就太棒了, 這會讓我們的生活更加容易; 但我們的動作往往非常雜亂。 人的動作有大量扭動,要在精準度和完整度上 精確辨識這些變化,是相當困難的, 因為唯有如此, 才能取得專業教練的信任。 而儘管找到這些正確的 時空特徵困難重重, 我們還是做到了。
Coaches trust our ability of our machine to identify these variations. We're at the point where almost every single contender for an NBA championship this year is using our software, which is built on a machine that understands the moving dots of basketball. So not only that, we have given advice that has changed strategies that have helped teams win very important games, and it's very exciting because you have coaches who've been in the league for 30 years that are willing to take advice from a machine. And it's very exciting, it's much more than the pick-and-roll. Our computer started out with simple things and learned more and more complex things and now it knows so many things. Frankly, I don't understand much of what it does, and while it's not that special to be smarter than me, we were wondering, can a machine know more than a coach? Can it know more than person could know? And it turns out the answer is yes.
教練們信任我們機器 辨識這些變化的能力。 我們已經達到,今年幾乎每一個 爭奪NBA冠軍的隊伍 都在使用我們的軟體, 安裝在可以讀懂籃球場上 移動的點的機器裡。 不只如此,這些隊伍也根據 我們的建議改變一些戰術, 幫助他們贏了一些很重要的比賽, 這令人感到非常興奮, 因為我們讓這些在聯盟裡 打滾了30年的教練, 願意聽一台機器的建議。 不只擋切戰術,更讓我們興奮的是, 我們的電腦從簡單的開始, 學會越來越複雜的動作, 現在它已經有豐富的知識。 老實說,它懂得已經比我多了, 但由於比我聰明也沒什麼特別的, 我們更想知道, 機器有可能懂得比教練還多嗎? 可能懂得比人類還多嗎? 結果答案是:可以。
The coaches want players to take good shots. So if I'm standing near the basket and there's nobody near me, it's a good shot. If I'm standing far away surrounded by defenders, that's generally a bad shot. But we never knew how good "good" was, or how bad "bad" was quantitatively. Until now.
教練都希望球員掌握好的投籃時機, 如果離籃框很近, 旁邊沒人防守,就是好的投籃時機。 如果我離籃框很遠,又被防守者包圍, 通常就是很差的投籃時機。 但我們從來無法從量化數據得知, 「好」有多好、「差」有多差。 但現在不同了。
So what we can do, again, using spatiotemporal features, we looked at every shot. We can see: Where is the shot? What's the angle to the basket? Where are the defenders standing? What are their distances? What are their angles? For multiple defenders, we can look at how the player's moving and predict the shot type. We can look at all their velocities and we can build a model that predicts what is the likelihood that this shot would go in under these circumstances? So why is this important? We can take something that was shooting, which was one thing before, and turn it into two things: the quality of the shot and the quality of the shooter. So here's a bubble chart, because what's TED without a bubble chart?
所以我們可以,同樣地, 利用時空特徵條件, 檢視每一次投籃。 我們可以得知:投籃位置在哪? 和籃框的角度是幾度? 防守者站在哪裡?距離多少? 角度多大? 防守者不只一個時, 我們可以觀察球員如何移動 來預測投籃類型。 我們可以觀察速率,建立一個可以預測 在此情況下的命中率模型。 為什麼這很重要? 我們可以將投籃 這種過去的單一行為,轉化為兩件事: 投籃動作本身的品質,和投籃者的品質。 這裡有一張泡泡圖, 沒有泡泡圖還像TED嗎?
(Laughter)
(笑聲)
Those are NBA players. The size is the size of the player and the color is the position. On the x-axis, we have the shot probability. People on the left take difficult shots, on the right, they take easy shots. On the [y-axis] is their shooting ability. People who are good are at the top, bad at the bottom. So for example, if there was a player who generally made 47 percent of their shots, that's all you knew before. But today, I can tell you that player takes shots that an average NBA player would make 49 percent of the time, and they are two percent worse. And the reason that's important is that there are lots of 47s out there. And so it's really important to know if the 47 that you're considering giving 100 million dollars to is a good shooter who takes bad shots or a bad shooter who takes good shots. Machine understanding doesn't just change how we look at players, it changes how we look at the game.
這些都是NBA球員。 泡泡大小代表球員體型大小, 顏色代表他打的位置。 X軸是進球的機率, 左邊的球員做了許多勉強的投籃動作, 右邊的球員在有空檔時才會出手。 Y軸是球員的投籃能力, 擅長投籃的球員在上方, 不擅長的在下方。 舉例來說,如果有個球員, 平均命中率大約47%, 這是過去所有你知道的資訊。 但今天,我能告訴你以 這個球員出手投籃的狀況, NBA球員的平均命中率是49%, 他比平均低了2%。 這之所以重要,是因為 有這麼多47%命中率的球員。 重點就是要搞清楚, 如果你要用100美金 簽下一個47%的球員, 應該要找能投得準但出手時機不佳的, 還是不那麼準但出手時機很好的球員。 機器的理解力不只改變了 我們對球員的看法, 也改變了我們對比賽的看法。
So there was this very exciting game a couple of years ago, in the NBA finals. Miami was down by three, there was 20 seconds left. They were about to lose the championship. A gentleman named LeBron James came up and he took a three to tie. He missed. His teammate Chris Bosh got a rebound, passed it to another teammate named Ray Allen. He sank a three. It went into overtime. They won the game. They won the championship. It was one of the most exciting games in basketball. And our ability to know the shot probability for every player at every second, and the likelihood of them getting a rebound at every second can illuminate this moment in a way that we never could before. Now unfortunately, I can't show you that video. But for you, we recreated that moment at our weekly basketball game about 3 weeks ago.
兩年前在NBA總冠軍系列戰, 有一場非常刺激的比賽。 邁阿密熱火隊落後3分, 時間還剩20秒。 他們即將把冠軍拱手讓人。 一位叫勒布朗·詹姆士的先生 出手一顆三分球企圖追平比賽, 球沒進。 他的隊友克里斯·波許搶到籃板, 傳給另一位隊友雷·艾倫。 他命中了一顆三分球, 將比賽帶入延長賽, 最後他們贏了比賽,拿到冠軍。 這是籃球場上最刺激的球賽之一。 而因為我們可以得知每一個球員 在每一秒鐘投進的機率, 以及每一秒鐘他們搶到籃板的機率, 讓我們得以用前所未有的方式 窺見這個時刻的全貌。 很可惜,現在我無法播放那段影片, 但為了各位,我們重建了那個時刻, 就在大約三週前 我們每週例行的籃球比賽裡。
(Laughter)
(笑聲)
And we recreated the tracking that led to the insights. So, here is us. This is Chinatown in Los Angeles, a park we play at every week, and that's us recreating the Ray Allen moment and all the tracking that's associated with it. So, here's the shot. I'm going to show you that moment and all the insights of that moment. The only difference is, instead of the professional players, it's us, and instead of a professional announcer, it's me. So, bear with me.
我們也重建了讓我們窺見 比賽全貌的追蹤數據。 所以,這就是我們, 在洛杉磯的中國城, 我們每週都在這個公園比賽。 我們正在重建雷·艾倫的經典時刻, 以及所有相關的追蹤數據。 這就是那經典的一球。 我將和各位展示那個時刻, 以及那個時刻背後的一切。 唯一的不同是,表演者是我們, 不是那些職業球員。 還有是我在播報,不是職業播報員, 所以請大家見諒。
Miami. Down three. Twenty seconds left. Jeff brings up the ball. Josh catches, puts up a three!
邁阿密熱火。 3分落後。 剩下20秒。 傑夫帶球過來。 喬許接到了球,三分出手!
[Calculating shot probability]
[計算命中率]
[Shot quality]
[投籃品質]
[Rebound probability]
[籃板機率]
Won't go!
球沒進!
[Rebound probability]
[籃板機率]
Rebound, Noel. Back to Daria.
諾爾搶到籃板。 向後傳給朵莉雅。
[Shot quality]
[投籃品質]
Her three-pointer -- bang! Tie game with five seconds left. The crowd goes wild.
她三分出手... 球進! 追平比賽,剩下5秒! 觀眾陷入瘋狂。
(Laughter)
(笑聲)
That's roughly how it happened.
事情大概就是這樣發生的。
(Applause)
(掌聲)
Roughly.
大概啦。
(Applause) That moment had about a nine percent chance of happening in the NBA and we know that and a great many other things. I'm not going to tell you how many times it took us to make that happen.
(掌聲) 這個時刻發生在NBA 賽場上的機率大約是9%, 我們能算出這件事 和其他許多事的機率。 我絕對不會告訴各位 我們試了幾次才成功。
(Laughter)
(笑聲)
Okay, I will! It was four.
好啦,告訴你們,四次。
(Laughter)
(笑聲)
Way to go, Daria.
幹得不錯,朵莉雅。
But the important thing about that video and the insights we have for every second of every NBA game -- it's not that. It's the fact you don't have to be a professional team to track movement. You do not have to be a professional player to get insights about movement.
但重要的並不是這則影片, 也不是它對NBA每場比賽 每一秒鐘的意涵。 重要的是,不是只有職業球隊 才能追蹤人的動作, 也不是只有職業球員 才需要這些動作背後的意涵。
In fact, it doesn't even have to be about sports because we're moving everywhere. We're moving in our homes, in our offices, as we shop and we travel throughout our cities and around our world. What will we know? What will we learn? Perhaps, instead of identifying pick-and-rolls, a machine can identify the moment and let me know when my daughter takes her first steps. Which could literally be happening any second now.
事實上,它也不是只能應用在運動場上, 因為我們隨時都不斷地在各地移動。 我們在家裡移動、 在辦公室裡移動、 購物和旅遊、 穿梭在我們的城市裡 以及全世界。 我們可以知道什麼?可以學到什麼? 或許,除了辨識擋切戰術之外, 機器還能辨識出感動時刻,讓我知道 我女兒何時踏出她的第一步。 這真的是每一秒鐘都可能發生的事。
Perhaps we can learn to better use our buildings, better plan our cities. I believe that with the development of the science of moving dots, we will move better, we will move smarter, we will move forward.
或許我們能夠學會更善加 利用建築物、做更好的都市規劃。 我相信藉由移動的點的科學進步, 我們將能更好地移動、更聰明地移動, 並真正向未來移動。
Thank you very much.
非常感謝各位。
(Applause)
(掌聲)