\uc6b0\ub9ac\uac00 \uc0ac\uc6a9\ud558\ub294 ChatGPT, Claude, Gemini\uc640 \uac19\uc740 \ubaa8\ub4e0 \ucd5c\uc2e0 AI \ubaa8\ub378\uc758 \uc2ec\uc7a5\ubd80\uc5d0\ub294 **’\ud2b8\ub79c\uc2a4\ud3ec\uba38(Transformer)’**\ub77c\ub294 \uc544\ud0a4\ud14d\ucc98\uac00 \uc788\uc2b5\ub2c8\ub2e4. 2017\ub144 \uad6c\uae00\uc774 \ubc1c\ud45c\ud55c \ub17c\ubb38 *”Attention Is All You Need”*\ub294 \uc778\uacf5\uc9c0\ub2a5\uc758 \uc5ed\uc0ac\ub97c \uc0c8\ub85c \uc37c\uc9c0\ub9cc, \uc218\uc2dd\uc73c\ub85c \uac00\ub4dd \ucc2c \ub17c\ubb38\uc744 \uc774\ud574\ud558\uae30\ub780 \uc27d\uc9c0 \uc54a\uc2b5\ub2c8\ub2e4.<\/p>\n
\uc774 \uae00\uc5d0\uc11c\ub294 \ubcf5\uc7a1\ud55c \uc218\ud559 \ub300\uc2e0 \uc9c1\uad00\uc801\uc778 \ube44\uc720\uc640 \uc2dc\uac01\uc801 \uad6c\uc870<\/strong>\ub97c \ud1b5\ud574 \ud2b8\ub79c\uc2a4\ud3ec\uba38\uac00 \uc5b4\ub5bb\uac8c \ubb38\ub9e5\uc744 \uc774\ud574\ud558\uace0 \uc5b8\uc5b4\ub97c \uc0dd\uc131\ud558\ub294\uc9c0 \uba85\ud655\ud558\uac8c \ud30c\ud5e4\uce69\ub2c8\ub2e4.<\/p>\n \uacfc\uac70\uc758 \uc5b8\uc5b4 \ubaa8\ub378(RNN, LSTM)\uc740 \ubb38\uc7a5\uc744 \uc77d\uc744 \ub54c \uc778\uac04\ucc98\ub7fc \uc67c\ucabd\uc5d0\uc11c \uc624\ub978\ucabd\uc73c\ub85c \ud55c \ub2e8\uc5b4\uc529 \uc21c\uc11c\ub300\ub85c \ucc98\ub9ac\ud588\uc2b5\ub2c8\ub2e4. \uc774 \ubc29\uc2dd\uc740 \ubb38\uc7a5\uc774 \uae38\uc5b4\uc9c8\uc218\ub85d \uc55e\ubd80\ubd84\uc758 \ub0b4\uc6a9\uc744 \uc78a\uc5b4\ubc84\ub9ac\ub294 \uce58\uba85\uc801\uc778 \ub2e8\uc810\uc774 \uc788\uc5c8\uc2b5\ub2c8\ub2e4.<\/p>\n \ud2b8\ub79c\uc2a4\ud3ec\uba38\ub294 **’\ubcd1\ub82c \ucc98\ub9ac(Parallel Processing)’**\uc640 **’\uc5b4\ud150\uc158(Attention)’**\uc774\ub77c\ub294 \ub450 \uac00\uc9c0 \ubb34\uae30\ub85c \uc774 \ubb38\uc81c\ub97c \ud574\uacb0\ud588\uc2b5\ub2c8\ub2e4.<\/p>\n \ud2b8\ub79c\uc2a4\ud3ec\uba38\uc758 \ud575\uc2ec\uc778 ‘Scaled Dot-Product Attention’\uc744 \ud30c\uc774\uc36c \ucf54\ub4dc\ub85c \uad6c\ud604\ud558\uba74 \uadf8 \uc6d0\ub9ac\uac00 \ub354 \uba85\ud655\ud574\uc9d1\ub2c8\ub2e4.<\/p>\n \ud2b8\ub79c\uc2a4\ud3ec\uba38\uac00 “The bank of the river”\ub77c\ub294 \ubb38\uc7a5\uc744 \ucc98\ub9ac\ud558\ub294 \uacfc\uc815\uc744 5\ub2e8\uacc4\ub85c \uc2dc\uac01\ud654\ud574 \ubcf4\uaca0\uc2b5\ub2c8\ub2e4.<\/p>\n \uc544\ub798 \ub2e4\uc774\uc5b4\uadf8\ub7a8\uc740 \ud2b8\ub79c\uc2a4\ud3ec\uba38\uc758 \uc804\uccb4\uc801\uc778 \ub370\uc774\ud130 \ud750\ub984\uc744 \ubcf4\uc5ec\uc90d\ub2c8\ub2e4.<\/p>\n1. \ud575\uc2ec \uac1c\ub150: \uc65c \ud2b8\ub79c\uc2a4\ud3ec\uba38\uc778\uac00? (RNN vs Transformer)<\/h2>\n
\n
\ud83d\udca1 \ube44\uc720: \ub3c4\uc11c\uad00 \uc0ac\uc11c<\/h3>\n
\n
2. \ucf54\ub4dc \ube14\ub85d: \uc5b4\ud150\uc158 \uba54\ucee4\ub2c8\uc998\uc758 \uc218\ud559\uc801 \uad6c\ud604 (Python)<\/h2>\n
import numpy as np\r\nimport torch\r\nimport torch.nn.functional as F\r\n\r\ndef scaled_dot_product_attention(query, key, value, mask=None):\r\n \"\"\"\r\n Query, Key, Value \ud589\ub82c\uc744 \uc0ac\uc6a9\ud55c \uc5b4\ud150\uc158 \uc2a4\ucf54\uc5b4 \uacc4\uc0b0\r\n \"\"\"\r\n d_k = query.size(-1) # \ucc28\uc6d0 \ud06c\uae30\r\n \r\n # 1. \uc5b4\ud150\uc158 \uc2a4\ucf54\uc5b4 \uacc4\uc0b0 (Query\uc640 Key\uc758 \ub0b4\uc801)\r\n scores = torch.matmul(query, key.transpose(-2, -1)) \/ np.sqrt(d_k)\r\n \r\n # 2. \ub9c8\uc2a4\ud0b9 (\uc635\uc158: \ub514\ucf54\ub354\uc5d0\uc11c \ubbf8\ub798 \ub2e8\uc5b4 \ucc38\uc870 \ubc29\uc9c0 \ub4f1)\r\n if mask is not None:\r\n scores = scores.masked_fill(mask == 0, -1e9)\r\n \r\n # 3. Softmax \uc801\uc6a9 (\ud655\ub960 \ubd84\ud3ec\ub85c \ubcc0\ud658)\r\n attention_weights = F.softmax(scores, dim=-1)\r\n \r\n # 4. Value\uc640 \uac00\uc911\ud569 (\ucd5c\uc885 \ubb38\ub9e5 \ubca1\ud130 \uc0dd\uc131)\r\n output = torch.matmul(attention_weights, value)\r\n \r\n return output, attention_weights\r\n<\/code><\/pre>\n3. \ub2e8\uacc4\ubcc4 \uac00\uc774\ub4dc: \ud2b8\ub79c\uc2a4\ud3ec\uba38 \ub0b4\ubd80 \ub3d9\uc791 \uc6d0\ub9ac<\/h2>\n
\n
\n
\n
\n
\n
\n
\n
4. \ube44\uc8fc\uc5bc \ub370\uc774\ud130: \ud2b8\ub79c\uc2a4\ud3ec\uba38 \uad6c\uc870 \ub2e4\uc774\uc5b4\uadf8\ub7a8<\/h2>\n
graph TD\r\n Input[\"\uc785\ub825 \ubb38\uc7a5 (Inputs)\"] --> Emb[\"\uc784\ubca0\ub529 + \uc704\uce58 \uc778\ucf54\ub529\"]\r\n Emb --> EncBlock[\"\uc778\ucf54\ub354 \ube14\ub85d (Encoder Block)\"]\r\n \r\n subgraph Encoder_Layer [\"\uc778\ucf54\ub354 \ub808\uc774\uc5b4 (\ubc18\ubcf5)\"]\r\n direction TB\r\n MultiHead1[\"\uba40\ud2f0 \ud5e4\ub4dc \uc5b4\ud150\uc158 (Self-Attention)\"]\r\n AddNorm1[\"Add & Norm\"]\r\n FFN1[\"Feed Forward\"]\r\n AddNorm2[\"Add & Norm\"]\r\n \r\n MultiHead1 --> AddNorm1\r\n AddNorm1 --> FFN1\r\n FFN1 --> AddNorm2\r\n end\r\n \r\n EncBlock --> Encoder_Layer\r\n \r\n subgraph Decoder_Layer [\"\ub514\ucf54\ub354 \ub808\uc774\uc5b4 (\ubc18\ubcf5)\"]\r\n direction TB\r\n MaskedAtt[\"\ub9c8\uc2a4\ud06c\ub4dc \uba40\ud2f0 \ud5e4\ub4dc \uc5b4\ud150\uc158\"]\r\n AddNorm3[\"Add & Norm\"]\r\n EncDecAtt[\"\uba40\ud2f0 \ud5e4\ub4dc \uc5b4\ud150\uc158 (Encoder-Decoder)\"]\r\n AddNorm4[\"Add & Norm\"]\r\n FFN2[\"Feed Forward\"]\r\n AddNorm5[\"Add & Norm\"]\r\n \r\n MaskedAtt --> AddNorm3\r\n AddNorm3 --> EncDecAtt\r\n EncDecAtt --> AddNorm4\r\n AddNorm4 --> FFN2\r\n FFN2 --> AddNorm5\r\n end\r\n\r\n Encoder_Layer -- \"K, V \ud589\ub82c \uc804\ub2ec\" --> EncDecAtt\r\n OutputShifted[\"\ucd9c\ub825 \ubb38\uc7a5 (Shifted Right)\"] --> EmbDec[\"\uc784\ubca0\ub529 + \uc704\uce58 \uc778\ucf54\ub529\"]\r\n EmbDec --> Decoder_Layer\r\n Decoder_Layer --> Linear[\"Linear & Softmax\"]\r\n Linear --> FinalProb[\"\ub2e4\uc74c \ub2e8\uc5b4 \ud655\ub960\"]\r\n<\/code><\/pre>\n<\/div>\n\ud2b8\ub79c\uc2a4\ud3ec\uba38 \uc8fc\uc694 \uad6c\uc131 \uc694\uc18c \ube44\uad50<\/h3>\n