在以前提到生成式人工智能(下面簡稱人工智能)時,相關的討論主要聚焦於其能力和在教育或學習上的應用方法。然而,在研究使用方面,卻缺乏足夠的探討。這可能是因為技術相對較新,大部分研究學者未必熟悉其使用方法。此外,由於每個人對研究都有自己的心得和解決問題的方法,如何與人工智能配合,並不一定能讓他們更加得心應手。

曾經讀過黃易的小說《覆雨翻雲》,其中提到一把絕世寶刀,名為鷹刀。武林高手可以藉由鷹刀傳說中的傳鷹和鷹緣活佛,窺探武學精粹,用家可吸收宇宙精神力量,提升自己的功力。故事中描述了武林中各人為了爭奪這把寶刀而瘋狂鬥爭。然而,對於真正的高手而言,即使這把寶刀放在他們面前,他們仍會猶豫不決。故事中提到,這把刀在普通兵器室內顯得不起眼,但當高手進入練武室時,大家一方面會對它產生異樣的感覺,另一方面會盡力保持內心平靜,避免受到它的影響。作者通過女主角夢瑤的解釋,揭示各大門派掌門都有類似的反應,因為他們都經過長時間的摸索,對武道有自己的見解,並對這些成果珍惜不已。如果他們憑藉這把鷹刀窺探通向武道不同之路,他們不知道是否會對自己已領悟的武道產生負面影響,進而得不償失。因此,這把刀對於高手而言, 並不一定能增加他們武學修為,反而可能會影響了他們武道的認知。基於這些判斷,這些武林中視都盡量不會顯示出他們對這把武器的感受。

鷹刀猶如人工智能,擁有它,你就能夠連接宇宙,獲得無盡的知識。然而,能否將這些力量融會貫通,運用到自己的技藝中,則取決於個人的天賦、掌握和理解能力。此外,對於教授而言,他們經過多年的學習,建立了自己的研究方式和問題解決方法。然而,人工智能是否能幫助他們的研究,並沒有一個確定的答案。學習新技能所需的時間可能比真正進行研究的時間還要長。因此,對於長期從事研究的人來說,學習如何使用人工智能技術可能並不具有很大的誘因。

回顧過去幾周有關人工智能的討論,我們不禁想象這些新技術如何有助於科學研究。然而,討論可能不會深入,並且由於技術日新月異,文章正式發表後,科技可能已經有所更新,討論的結果也可能不盡完美。儘管如此,我們可以拋磚引玉,共同探討這個議題。

其中一個對ChatGPT生成式人工智能的批評是,它的內容不一定準確。尤其是對於一些資料性的問題,可能隨機作答,給出的答案可能是似是而非的,這樣的結果不能成為我們使用人工智能幫助研究的理由。因為科研需要準確的結論,過程不能馬虎了事,每一步都需要有充足的證據去證明和立論。每個論點都必須基於事實,如果中間推論過程有任何缺失,我們研究的成果也可能因此作廢。因此,大多數人不相信生成式人工智能可以幫助我們進行研究。有些讀者可能也同意,發覺一些人工智能不僅會給出錯誤的答案,而且可能不承認自己的錯誤,這也增加了使用人工智能進行研究的風險。

我認為這其實與人類本身有些相似。有時即使在課堂上學習的知識也不一定完全正確。有時我們在課堂上提出的問題,也未必能立即給出正確的答案。有時甚至需要到下一節課才能修正答案。對於同學提出的問題,我們也未必能立即給予正確的答案。有時可能只能簡單地想一想,給出一個近似的答案,之後在課後仔細回想時才發現瑕疵,只能在下一節課重新講解。在研究時,我們在網上尋找資料,但這些資料也不一定完全正確。科學理論也需要隨著時間不斷修正和調整,以解釋更多的現象,這正是建模的過程,也是科學界多年來一直相信的理論架構。

在現今人工智能不斷進步的時代,讓人工智能幫助我們解決科研問題可能仍有一些困難。目前,我認為人工智能的能力還不足以完全解決科研問題本身。然而,這並不代表我們不能將人工智能的應用融入到我們的研究中。最簡單的應用方式當然是讓人工智能幫我們修改論文。由於ChatGPT是一個大型自然語言處理器(NLP),其發展基於對語言本身的邏輯思維,因此他最基本的使用還是在語言編輯上。他在語法、用字、文章流暢程度、邏輯等方面的能力應該都比一般人更出色,因此如果我們要求他幫我們修改文章,他的能力絕對會比大部份人強。這是人工智能最簡單的應用之一。

除了修改文章外,ChatGPT 技術還可以用來簡化文章,以便更容易了解不同研究方法。透過從其他人不同文章中提取段落,我們可以要求人工智能用簡單的語言編寫文章,讓我們能夠更快速地了解論文的主旨和問題,從而讓研究人員更快地掌握新知識,以便應用在自己的研究中。此外,我們還可以利用人工智能,整合不同文章的內容,以幫助自己撰寫論文的引言。但需要注意的是,這些方法並不能代替學習,研究人員不能僅依靠剪貼拼湊的過程來學習新知識。

在科學研究領域,英語仍然是主流的論文編寫語言。如果我們的英語能力不夠強,沒有辦法閱讀大量英文文章,那麼人工智能可以幫助我們將整篇文章翻譯成我們喜歡的語言。此外,如果我們希望了解其他語言的研究文章,也可以要求軟件翻譯,這樣我們就可以了解來自世界各地的研究成果。隨著語言處理技術的不斷發展,可能將來博士課程中不再需要博士生掌握多種語言。如果我們想閱讀用不同語言撰寫的論文,只需要將論文提交給人工智能,他就可以使用最快捷的方法翻譯給我們,讓我們更加方便地閱讀和理解。

然而, 除了上面關於語文方面的幫助,人工智能的應用可能更加廣泛。例如,我們可以向ChatGPT提出關於研究方向的問題,例如詢問某特定方向的研究題目,相關的研究項目和比較等等。

下面是一個有關球體表面數據插值問題的例子。

Tim:「given a set of ordered data points on the unit sphere, summarize existing methods to interpolate these points to get a curve. tell me a new research direction and explain why the proposed approach is better」

GPT-3.5-Turbo:「

Existing methods for interpolating a set of ordered data points on the unit sphere to get a curve can be broadly categorized into two types: parametric and non-parametric methods.

Parametric methods involve defining a parametric curve that passes through the given data points. One popular approach is to use spherical splines, which are piecewise polynomial functions that interpolate the data points and have continuous first and second derivatives. Another approach is to use circular arcs or great circles to connect the data points.

Non-parametric methods involve defining a surface that passes through the data points and then projecting this surface onto the sphere to obtain a curve. One approach is to use radial basis functions, which involve fitting a weighted sum of radial basis functions to the data points. Another approach is to use spherical harmonic functions, which are a sum of orthogonal functions that can be used to represent the surface.

A new research direction for interpolating data points on the unit sphere could be to use deep learning techniques. Specifically, we could use a neural network to learn a mapping from the input data points to the output curve. This approach would involve training the neural network on a large dataset of input-output pairs and using it to predict the output curve for new input data points.

The proposed approach is better than existing methods in several ways. First, it is more flexible and can handle a wider range of input data. Second, it can learn complex patterns and relationships in the data, which may not be captured by traditional interpolation methods. Third, it can be easily scaled to handle large datasets. Finally, it can be adapted to different types of input data, such as images or point clouds, by using appropriate neural network architectures. Overall, the proposed approach has the potential to significantly improve the accuracy and efficiency of interpolation on the unit sphere.

GPT-4:「

Here is a summary of existing methods for interpolating ordered data points on the unit sphere:

  1. Great Circle Interpolation: This is the simplest method, which involves finding the shortest path (geodesic) between each pair of consecutive points. The curve is composed of several great circle segments, which are the intersections of the sphere and the plane containing the two points and the sphere’s center.
  2. Spherical Splines: Spherical splines are natural extensions of cubic splines to the unit sphere. They are smooth curves that pass through the data points and have continuous first and second derivatives. Some popular spherical splines include the Spherical Linear Interpolation (SLERP) and the Spherical Cubic Spline Interpolation (SCSI).
  3. Quaternion Interpolation: Quaternions are a more general way to represent rotations in 3D space. Converting the data points to quaternions and using quaternion-based interpolation methods such as SLERP or SQUAD (Spherical and Quadrangle Averaging) can create smooth curves that pass through the data points.
  4. Manifold Learning Techniques: These methods project the data points to a lower-dimensional space and perform interpolation in that space. Some examples include multidimensional scaling (MDS), locally linear embedding (LLE), and Isomap.

New Research Direction: Geodesic Neural Networks

A potential new research direction could be the development of Geodesic Neural Networks (GNNs) for interpolating ordered data points on the unit sphere. GNNs would be specifically designed to learn the underlying structure and curvature of the data points in the spherical space, taking into account the geodesic distances between points.

This approach would be better for several reasons:

  1. Adaptability: GNNs could be trained to adapt to different interpolation tasks, making them suitable for a wide range of applications. The network could learn the optimal weighting and smoothness for the specific dataset at hand.
  2. Robustness: By learning the curvature and structure of the data points, GNNs could be more robust to noise and outliers in the data.
  3. Scalability: Neural networks can handle large datasets more efficiently than some of the traditional interpolation methods. As the number of data points increases, GNNs’ performance would scale better than other interpolation methods.
  4. End-to-end learning: GNNs can be trained end-to-end, allowing for a seamless integration with other deep learning techniques and creating opportunities for joint optimization with other tasks (e.g., data augmentation, denoising, or classification).
  5. Generalization: Once trained, GNNs can potentially generalize to new datasets with similar characteristics, making them a versatile tool for interpolating ordered data points on the unit sphere.

In summary, Geodesic Neural Networks could offer a promising new direction for interpolating ordered data points on the unit sphere by leveraging the power of deep learning to learn the underlying structure and curvature of the data points in spherical space.

這是我最近感興趣的研究方向,問題聽起來相對簡單,但困難在於如何做得好。

上面的問題,要求人工智慧回答關於這個研究方向的現有方法,並比較它們的效果,然後提供一個新的研究方向,並解釋其優點。我使用了Poe裡的GPT-3.5-Turbo和GPT-4,看看他們如何回答這個問題。GPT-3.5的回答提到了兩個大方向。第一個大方向類似於我想要探討的方法,嘗試使用一些更直接的技巧在球體表面上線性插值這些點。我之前的研究都是基於這個方向。然而,他並未詳細介紹這些方法,科研人員仍需進一步搜尋現有的方法以了解這個大方向的詳情。第二個方向是非參數方法,其中提到了RBF和Spherical Harmonic。第一次看到這些答案時,我需要花時間思考,不確定他是否在亂拼湊。最大的問題是他未提供出處,我們難以了解他為什麼要提及這些數學方法。我認為所謂的非參數方法在ordered data 上應該比較困難。我不知道有誰使用這些方法去解決這個問題。如果我們只需要找出一條在球體表面的曲線穿越這些點,這些方法其實是可行的。在另一篇論文中,我使用水平值Level Set Method 解決了這個問題。至於GPT-4提到的問題前半部分,他給出的答案比較詳細和準確,還提到了SLERP、SQUAD和QUATERNIONS等方法,這些都是我研究中提到的方法。他還提到了使用機器學習的方法,我不確定適用這些針對高維空間的方法於三維空間是否真的有好處。

在問題的下半部分,要求人工智慧提供一個新的研究方向來解決這個數值問題。兩個不同的人工智慧都提到了使用神經網絡。GPT-4提到了一種名為Geodesic Neural Network(GNN)的方法,但在Google上並沒有相關的討論,最接近的是一種名為Geodesic Convolutional Neural Network (GCNN)。簡單的了解一下那篇文章,這種方法可能與我們的插值問題沒有太大關聯。GPT-3.5並未給出太多討論,我不確定如何使用神經網絡解決這個數學問題。此外,他得出的結論是這些方法可以大幅提高答案的準確性,但我對此完全不相信。

這個例子顯示出,人工智能對於已知知識的處理能力似乎非常出色。它可以提供一些基本的知識,讓科研人員能夠進一步搜尋、了解研究範疇,並對該研究方向有更深入的認識。然而,如果我們完全依賴人工智能的建議,在其建議的方向進行研究,結果並不一定能夠成功,因為它提供的建議可能只是胡亂拼湊,結論不一定正確。

儘管人工智能並不能直接教導我們如何進行研究或提供研究方向,但這並不代表它不能應用於研究方向的設定中。有時學者需要跳脫傳統思維,嘗試進行腦力激盪,將不同的知識範疇進行連結,從而產生新的知識和方法。這種方法可以產生創新和新穎的想法。我認為,如果我們不太重視產生結果的合理性,人工智能在這個創新領域可能比大多數人做得更好。由於人工智能的回答具有隨機性,它可以提供一些以往未曾想過的方法和研究道路。即使這些道路不一定成功,我們也可以參考它們,看看是否有一些新的想法和創意。

隨著人工智能的發展日新月異,它已經能夠做出一些以前難以想像的事情。現在,由人工智能製造的東西越來越難以區分是由人類還是電腦所製造。儘管如此,在科學研究方面,人工智能暫時還無法像人類一樣,能夠發現和解決問題,並遵循嚴謹的邏輯進行研究。當然,現在的局面並不代表未來不會有所改變。任何行業都有可能被取代,因此每個人都應該不斷學習,提升自己的競爭力,以應對這個快速變化的世界。