解釋得嘅 AI

解釋得嘅 AI 係指人工智能（AI）領域上一啲技術，用嚟儘量等人工智能做嘅資訊處理過程可以由人類理解。喺廿一世紀初嘅人工智能界，解釋得嘅 AI 開始受重視：噉係因為打前嘅人工智能技術好多時都有所謂嘅黑盒問題，即係佢哋能夠運算出正確結果，但人類冇辦法理解佢哋係點樣得出結果嘅－就好似一個「黑盒」噉，能夠俾到正確結果，但人對個盒嘅內部運作一無所知。因此，人工智能領域就開始思考解釋得嘅 AI 嘅問題，會著手研究啲「黑盒」型嘅 AI 係點樣得出正確結果嘅，亦會致力創造出一啲由人類角度睇嚟能夠理解嘅人工智能技術，但同時唔會完全禁止開發自動學習嘅嘗試，亦唔會要求硬係要解釋晒所有嘢^[1]。

美國嘅國防高級研究計劃局設立咗專門嘅項目嚟針對「解釋得嘅 AI」嚟做研究。

背景

解釋得嘅 AI 嘅諗頭源於廿世紀同廿一世紀人工智能技術成日有嘅黑盒問題：如果話一個人工智能程式係一個「黑盒」，意思係話佢能夠運算出正確嘅結果，但人類用家冇辦法理解個程式係點樣得出呢個結果嘅；例如係人工神經網絡就係噉^[2]：人類喺用個腦諗完嘢解決完問題之後，曉將自己嘅思考過程口頭報告返出嚟；但一個神經網絡唔會識做呢樣嘢，唔會話到俾研究者聽佢經過啲乜嘢思考過程先能夠達到佢所俾嘅輸出。所以有人將神經網絡比喻做一個黑盒－就算佢嘅輸出係準確嘅都好，個研究者都淨係知個神經網絡嘅輸入同輸出，冇辦法知道當中嘅思考過程。即係話對於「個神經網絡俾嘅答案點解係啱」呢條問題，目前唯一嘅答覆係「因為一部醒過人類嘅系統俾咗個答案出嚟」，雖然呢點唔損害神經網絡喺實用上預測現象嘅能力，但噉表示神經網絡嘅可解釋度低得好交關，令好多學者都覺得難以接受^[3]^[4]^[5]。因為噉，有好多科學家都致力研究點樣先可以拆解一個經歷咗訓練、做緊準確預測嘅神經網絡^[2]^[6]。

喺科學同工程學上，一個黑盒系統嘅輸入同輸出之間有已知嘅關係，問題係研究者唔知佢內部點樣運作，點樣由輸入產生輸出^[7]。

除咗科研上嘅好奇－「想透過瞭解 AI 做決策嗰陣嘅思路嚟改良啲模型」等－之外，解釋得嘅 AI 亦都源於社會、倫理甚至法律上嘅實際需求：到咗廿一世紀初，好多人工智能模型經已俾人廣泛噉使用，而呢啲應用好多時會牽涉重要資訊、財產甚至乎係人身安全（譬如自駕車撞起車上嚟可以搞出人命），如果將呢啲技術變成解釋得嘅 AI 嘅話，能夠有效提升人對呢啲技術嘅信任（而唔係同用家講「嗱，我哋都唔知呢個 AI 點解做到正確嘅運算，但總之信佢啦」噉嘅說話）^[8]。

玻璃箱模型

相對於傳統上對於機械學習佢黑箱式模型嘅放任，DARPA 希望啲解釋得嘅 AI 係「玻璃箱」噉嘅模型，即係唔單之啲 AI 作出嘅決策可以解釋得到，仲可以允許人類全程監管個過程、嚟去最終裁決啲 AI 嘅決策。

法令

2018 年修定嘅歐盟一般資料保護規範（英文：General Data Protection Regulation）（GDPR）引入唨「得解釋權（英文：Right to explanation）」（Right to explanation），希望噉樣解決得啲日漸受重視嘅演算法可能帶埋嘅問題，之但係同時 GDPR 只係要求局部嘅解釋。

睇埋

引咗

↑ Holzinger, A., Biemann, C., Pattichis, C. S., & Kell, D. B. (2017). What do we need to build explainable AI systems for the medical domain?. arXiv preprint arXiv:1712.09923.
↑ ^2.0 ^2.1 THE BLACK BOX PROBLEM CLOSES IN ON NEURAL NETWORKS 互聯網檔案館嘅歸檔，歸檔日期2018年6月22號，.. The Next Platform.
↑ Trust, and don’t verify: the AI black box problem 互聯網檔案館嘅歸檔，歸檔日期2018年7月30號，.. Medium.
↑ Peres, D. J.; Iuppa, C.; Cavallaro, L.; Cancelliere, A.; Foti, E. (1 October 2015). "Significant wave height record extension by neural networks and reanalysis wind data". Ocean Modelling. 94: 128–140.
↑ Dwarakish, G. S.; Rakshith, Shetty; Natesan, Usha (2013). "Review on Applications of Neural Network in Coastal Engineering". Artificial Intelligent Systems and Machine Learning. 5 (7): 324–331.
↑ Dewdney, A. K. (1 April 1997). Yes, we have no neutrons: an eye-opening tour through the twists and turns of bad science. Wiley. p. 82.
↑ Holzinger, Andreas; Plass, Markus; Holzinger, Katharina; Crisan, Gloria Cerasela; Pintea, Camelia-M.; Palade, Vasile (3 August 2017). "A glass-box interactive machine learning approach for solving NP-hard problems with the human-in-the-loop". arXiv:1708.01104
↑ Edwards, Lilian; Veale, Michael (2017). "Slave to the Algorithm? Why a 'Right to an Explanation' Is Probably Not the Remedy You Are Looking For". Duke Law and Technology Review. 16: 18.

拎

（英文）解釋得嘅 AI：令機械對人類嚟講可以理解

[1] Holzinger, A., Biemann, C., Pattichis, C. S., & Kell, D. B. (2017). What do we need to build explainable AI systems for the medical domain?. arXiv preprint arXiv:1712.09923.

[blackbox-2] 2.0 ^2.1 THE BLACK BOX PROBLEM CLOSES IN ON NEURAL NETWORKS 互聯網檔案館嘅歸檔，歸檔日期2018年6月22號，.. The Next Platform.

[3] Trust, and don’t verify: the AI black box problem 互聯網檔案館嘅歸檔，歸檔日期2018年7月30號，.. Medium.

[4] Peres, D. J.; Iuppa, C.; Cavallaro, L.; Cancelliere, A.; Foti, E. (1 October 2015). "Significant wave height record extension by neural networks and reanalysis wind data". Ocean Modelling. 94: 128–140.

[5] Dwarakish, G. S.; Rakshith, Shetty; Natesan, Usha (2013). "Review on Applications of Neural Network in Coastal Engineering". Artificial Intelligent Systems and Machine Learning. 5 (7): 324–331.

[6] Dewdney, A. K. (1 April 1997). Yes, we have no neutrons: an eye-opening tour through the twists and turns of bad science. Wiley. p. 82.

[7] Holzinger, Andreas; Plass, Markus; Holzinger, Katharina; Crisan, Gloria Cerasela; Pintea, Camelia-M.; Palade, Vasile (3 August 2017). "A glass-box interactive machine learning approach for solving NP-hard problems with the human-in-the-loop". arXiv:1708.01104

[8] Edwards, Lilian; Veale, Michael (2017). "Slave to the Algorithm? Why a 'Right to an Explanation' Is Probably Not the Remedy You Are Looking For". Duke Law and Technology Review. 16: 18.

[1]

[2]

[3]

[4]

[5]

[6]

[7]

[8]