英文部分擷自於BBC NEWS
The mushroom that AI thinks is a pretzel AI把香菇誤認成蝴蝶餅
Butterflies labelled as washing machines, alligators as hummingbirds and dragonflies that become bananas.
蝴蝶誤認成洗衣機，鱷魚誤認成蜂鳥，就連蜻蜓在AI的眼裡也只是個香蕉！
These are just some of the examples of tags which artificial intelligence system have given images.
這只是部分的例子，AI在許多圖片上都辨認錯誤。
Now researchers have released a database of 7,500 images that AI systems are struggling to identify correctly.
研究人員已將數據庫中7,500張機器人難以判讀的圖片公開出來，
One expert said it was crucial to solve the issue if these systems were going to be used in the real world.
專家認為若未來要將AI技術運用於生活中，就必須先解決此問題。
“No-one quite knows why they are failing to recognise these images which don’t look that hard,” said Calum Chace, an expert in the field.
AI領域的專家Calum Chace表示：「沒有人知道為什麼機器人對於看似容易判斷的圖片會辨認不出來，
“And while no-one knows what the solution is, my hunch is that it won’t hold up AI research for long because there is an enormous amount of money and talent that can be thrown at the problem to solve it.”
雖然還沒有什麼解決辦法出現，但有這麼多的資金及人才投入，我認為這樣的問題不會拖很久。」
The researchers from UC Berkeley, and the Universities of Washington and Chicago, said the images they have compiled — in a dataset called ImageNet-A — have the potential to seriously affect the overall performance of image classifiers, which could have knock-on effects on how such systems operate in applications such as facial recognition or self-driving cars.
加州大學柏克萊分校、華盛頓大學和芝加哥大學的研究人員表示，他們在ImageNet數據庫中所匯集到的圖片，可能會對圖片分類器的整體表現造成極大的影響，同時這類研究也將連帶影響人臉辨識APP和自駕車的系統運作。
“The problem has to be solved before systems like self-driving cars become standard,” said Mr Chace.
Calum Chace也認同在AI自駕技術發展完善之前，必須解決誤判問題。
The images were all collected online and none had been digitally altered.
這些數據庫中的圖片都是從網路上收集而來的，並且沒有修圖過。
Researchers hope the database will help experts improve the accuracy of how AI systems classify images.
研究人員期望透過這個數據庫，能夠幫助專家們提升AI辨認圖片的準確度，
Previous images tested on AI may have been too simple and not fully represent the ones the systems will encounter “in the real word”, the researchers said.
然而至今所測試過的圖片都算是比較容易辨識的，在真實世界中人物呈現的形式都將更加複雜。
AI often misidentifies objects because it is over-generalising, so for instance a shadow in a picture of a sundial will lead algorithms to label shadows as sundials. Or it may think that all cars are limousines.
AI會判斷錯誤可能是因為細節上的忽略，舉例來說，一張圖片裡如果出現了日晷和日晷的影子，那麼AI把影子也認為是日晷的一部分；一張圖片裡出現了許多車輛接連在一起，AI會誤判為是一輛高級豪華車。
The original ImageNet was used to train neural networks — systems that could teach themselves — and was part of a rebirth of AI, as computer power and huge databases combined to make far more capable systems.
原本ImageNet數據庫是用來訓練神經網路，能夠讓系統自我學習，若與AI技術結合，計算機與龐大的數據庫能夠讓系統更加強大。
