22

卷積層的千絲萬縷 | Mingled Sensations in ConvNets

盧振華、吳柏瑤、李哲榮 (指導老師)、許素朱 (指導老師) Chen-Hua Lu, Po-Yao Wu, Che-Rung Lee (supervisor), Su-Chu Hsu (supervisor)

近年來,人工智慧已在不同領域掀起壯闊波瀾。在這個人工智慧世代,我們的創作理念是建立一個與認知相關的人工智慧系統,再將這個系統原始且富創造性的演算過程以藝術造型呈現出來。為此我們選用了卷積神經網路(ConvNets);其作為當代最先進人工智慧技術之一,特別適合用於視覺的識別與生成任務。首先我們設計並訓練了卷積神經網路,將輸入圖像逐層解析,並將其對空間的感知生成為中文象形字;而後將數位資訊轉換為類比呈現,讓卷積神經網路的推演過程與結構得以視覺化,特別是神經元的活化與模型權重。有別於多數現有的人工智慧藝術作品,多用以展示人工智慧的威力或反映人工智慧對於人類社會的影響,期許這件作品能幫助大家對現代人工智慧系統的內部結構有更好的認識,以多樣的角度欣賞其複雜性的美感。 Recently, Artificial Intelligence (AI) has flourished in diverse areas. In the era of AI, our concept is to create an AI system that showcases its perception, then we reveal the its primality and ingenuity in the computing process by installation art. We built the convolutional neural network (ConvNet, or CNN), which is a state-of-art AI model, and specializes in visual recognition and image generation. First, we designed and trained ConvNets that can analyze images and transform their local perception layer-by-layer to the corresponding Chinese pictograms. Then, a set of installation art pieces that transforms the digital structure of CNN to an analog representation is constructed to visualize the algorithmic and structural details of ConvNets, which emphasizes neural activations and model weights. Unlike most existing AI artworks created to demonstrate the power of AI, or to show the reflection on the influence of AI to human society, we expect this work would help people have a better understanding of the inside architecture of the modern AI system, and appreciate the beauty of its complexity.