Scientists Are Trying to Teach AI How to Smell科學家嘗試讓人工智能獲得嗅覺

Describing odors can be surprisingly complicated， even for a complex computer.

用文字來描述氣味，即便是借助于復雜的計算機系統，也可能會出人意料地困難。

It’s hard to overstate the power of the nose—research says humans can distinguish more than a trillion odors. This is especially impressive when you remember that each individual odor is a chemical with a unique structure. Experts have been trying to discern patterns or logic in how chemical structure dictates smell， which would make it much easier to synthetically replicate scents or discover new ones. But that’s incredibly challenging—two very similarly structured chemicals could smell wildly different. When identifying smells is such a complicated task， scientists are asking： Can we get a computer to do it？

我們很難充分形容鼻子的功能有多么強大。研究顯示，人類可以分辨超過一萬億種氣味，這是一個驚人的數字，尤其是考慮到每種氣味都是具有獨特結構的化學物質。專家們一直在嘗試找出化學結構決定氣味的規律或邏輯，如此人工合成氣味或發現新的氣味便容易得多。但這極為困難，因為兩種結構非常近似的化學物質，氣味也可能截然不同。既然識別氣味如此艱難，科學家便提出了這樣的疑問：能否讓電腦來完成這項任務？

Smell remains more mysterious to scientists than our senses of sight or hearing. While we can “map” what we see as a spectrum of light wavelengths， and what we hear as a range of sound waves with frequencies and amplitudes， we have no such understanding for smell. In new research， published in September 2023 in the journal Science， scientists trained a neural network with 5，000 compounds from two perfumery databases of odorants—molecules that have a smell—and corresponding smell labels like “fruity” or “cheesy.” The AI was then able to produce a “principal odor map” that visually showed the relationships between different smells. And when the researchers introduced their artificial intelligence to a new molecule， the program was able to descriptively predict what it would smell like.

相較于視覺或聽覺，嗅覺對科學家來說更加神秘難解。我們可以把視覺“映射”為光譜，把聽覺“映射”為具有頻率和振幅的一系列聲波，然而，氣味卻不能如此解讀。在2023年9月的《科學》雜志上發表的最新研究中，科學家從兩個香水呈香物質（散發氣味的分子）數據庫中提取了5000種化合物及相應的氣味標簽，如“果香”或“奶酪香”等，以此訓練出一個神經網絡。人工智能借此生成了一個“主氣味圖”，直觀地展示出不同氣味間的關系。當研究人員將新的分子輸入人工智能系統時，該程序能夠以文字描述的形式預測分子的氣味。

The research team then asked a panel of 15 adults with different racial backgrounds living near Philadelphia to smell and describe that same odor. They found that “the neural network’s descriptions are better than the average panelist， most of the time，” says Alex Wiltschko， one of the authors of the new paper. Wiltschko is the CEO and co-founder of Osmo， a company whose mission is “to give computers a sense of smell” and that collaborated with researchers from Google and various US universities for this work.

研究團隊隨后請居住在費城周邊、不同種族背景的15名成年人嗅聞同樣的氣味并進行描述。這篇新論文的作者之一亞歷克斯·維爾奇科說，研究人員發現，“大多數情況下，神經網絡給出的描述比普通小組成員的描述更加準確”。維爾奇科是歐斯莫公司的首席執行官及聯合創始人，這家公司以“賦予計算機嗅覺”為使命，并為此與谷歌公司及多家美國大學的研究者開展合作。

“Smell is deeply personal，” says Sandeep Robert Datta， a neurobiology professor at Harvard University. （Datta has previously acted as a nominal advisor to Osmo， but was not involved in the new study.） And so， any research related to how we describe and label smells has to come with the caveat that our perception of smells， and how smells might relate to each other， is deeply entwined with our memories and culture. This makes it difficult to say what the “best” description of a smell even is， he explains. Despite all this， “there are common aspects of smell perception that are almost certainly driven by chemistry， and that’s what this map is capturing.”

哈佛大學神經生物學教授桑迪普·羅伯特·達塔說：“對氣味的感受是因人而異的。”（達塔曾是歐斯莫公司的名義顧問，但未參與這項新研究。）所以，每一項與氣味描述和氣味標簽相關的研究都必須明確：我們對氣味的感知及氣味間可能存在的相互關聯，是與我們的記憶和文化息息相關的。達塔解釋道，正因如此，很難說一種氣味的“最佳”描述是什么。不過，“氣味感知有一些共同點，基本確定是受化學成分的影響，這正是這張圖要記錄的內容”。

It’s important to note that this team is not the first or only to use computer models to investigate the relationship between chemistry and smell perception， Datta adds. There are other neural networks， and many other statistical models， that have been trained to match chemical structures with smells. But the fact that this new AI produced an odor map and was able to predict the smells of new molecules is significant， he says.

達塔補充道，需要指出的是，這個團隊并不是第一個或唯一一個使用計算機模型研究化學成分與氣味感知間關系的團隊。還有其他神經網絡及許多其他統計模型用于訓練以實現化學結構與氣味的匹配。不過，達塔表示，這款新型人工智能生成了氣味圖，并能預測新分子的氣味，這一點具有重要意義。

This neural network strictly looks at chemical structure and smell， but that doesn’t really capture the complexity of the interactions between chemicals and our olfactory receptors， Anandasankar Ray， who studies olfaction at the University of California， Riverside， and was not involved in the research， writes in an email. In his work， Ray has predicted how compounds smell based on which of the approximately 400 human odorant receptors are activated. We know that odorant receptors react when chemicals attach to them， but scientists don’t know exactly what information these receptors transmit to the brain， or how the brain interprets these signals. It’s important to make predictive models while keeping biology in mind， he wrote.

加利福尼亞大學河濱分校的嗅覺研究專家阿南達桑卡爾·雷在郵件中寫道，該神經網絡把關注點全部放在化學結構和氣味上，卻未真正體現出化學物質與人類嗅覺受體間相互作用的復雜性。雷沒有參與這項研究。在他的研究工作中，雷根據大約400個人體氣味受體中哪個被激活，來預測化合物的氣味。我們知道，當化學物質附著到氣味受體上時，受體會做出反應，但科學家并不知道這些受體向大腦傳遞的具體信息，也不知道大腦是如何解讀這些信號的。他寫道，在創建預測模型時，必須考慮生物學因素。

Additionally， to really see how general the model could go， Ray points out that the team should have tested their neural network on more datasets separate from the training data. But until they do that， we can’t say how widely useful this model is， he adds.

此外，雷指出，為了真正評估該模型的普適性，開發團隊除了測試訓練數據之外，還應在更多的數據集上測試他們的神經網絡。他補充道，在此之前，我們無法判斷該模型的廣泛適用性。

What’s more， the neural network doesn’t take into account how our perceptions of a smell can change with varying concentrations of odorants. “A really great example of this is a component of cat urine called MMB; it’s what makes cat pee stink，” says Datta. But at very low concentrations， it smells quite appealing and even delicious—it’s found in some coffees and wines. It’ll be interesting to see if future models can take this into account， Datta adds.

此外，該神經網絡沒有考慮我們對氣味的感知是如何隨著呈香物質的濃度變化而改變的。“一個很好的例子是貓尿中一種叫MMB的成分，它是貓尿臭味的來源。” 達塔說。但在濃度極低時，它的氣味卻相當誘人，甚至是美味。某些咖啡和紅酒中就含有此物質。達塔補充道，未來的模型能否考慮到這一點，值得關注。

Overall， it’s important to note that this principal odor map “doesn’t explain the magic of how our nose sifts through a universe of chemicals and our brain alights on1 a descriptor，” says Datta. “That remains a profound mystery.” But it could facilitate experiments that help us interrogate how the brain perceives smells.

總之，需要指出的是，這份主氣味圖“并未解開我們的鼻子如何區分海量化學物質，以及我們的大腦如何識別特征信息的奧秘”，達塔說，“這些疑問依然奇妙難解。”不過，這份氣味圖可能為實驗提供便利，幫助我們探究大腦是如何感知氣味的。

Witschko and his collaborators are aware of other limitations of their map. “With this neural network， we’re making predictions on one molecule at a time. But you never smell one molecule at a time—you always smell blends of molecules，” says Witschko. From a flower to a cup of morning coffee， most “smells” are actually a mixture of many different odorants. The next step for the authors will be to see if neural networks can predict how combinations of chemicals might smell.

維爾奇科及其團隊明白他們這份氣味圖在其他方面的局限性。“使用這個神經網絡，我們一次只對一種分子進行預測。但是你從來不會一次只聞到一種分子的氣味，你聞到的永遠是多種分子的混合氣味。” 維爾奇科表示。無論是一朵花，還是一杯晨間的咖啡，大多數“氣味”實際上是多種不同呈香物質的混合。接下來，開發團隊將嘗試探索神經網絡能否預測混合化學物質的氣味。

Eventually， Wiltschko envisions a world where smell， like sound and vision， is fully digitizable. In the future he hopes machines will be able to detect smells and describe them， like speech to text capabilities on smartphones. Or they would be able to exude specific smells on demand.

最終，在維爾奇科設想的世界里，氣味會像聲音和畫面一樣實現完全數字化。未來他希望機器能夠檢測氣味并進行描述，就像智能手機上的語音轉文字功能一樣。或是設備能根據要求釋放特定氣味。

（譯者為“《英語世界》杯”翻譯大賽獲獎者）

1 alight on偶然碰見；無意間發覺。