\uc778\uacf5\uc9c0\ub2a5\uc758 \ubd84\uc57c\uc5d0\uc11c ‘\uc9c0\uc18d \ud559\uc2b5(Continual Learning)’\uacfc ‘\ud3c9\uc0dd \ud559\uc2b5(Lifelong Learning)’\uc740 \ub9e4\uc6b0 \uc911\uc694\ud55c \uc8fc\uc81c\uc785\ub2c8\ub2e4. \ub450 \uac1c\ub150 \ubaa8\ub450 \ubaa8\ub378\uc774 \uc2dc\uac04\uc5d0 \ub530\ub77c \ubcc0\ud654\ud558\ub294 \ub370\uc774\ud130 \ub610\ub294 \uc791\uc5c5\uc744 \ud559\uc2b5\ud560 \uc218 \uc788\ub3c4\ub85d \ud558\ub294 \ub370 \ucd08\uc810\uc744 \ub9de\ucd94\uace0 \uc788\uc2b5\ub2c8\ub2e4. \uc804\ud1b5\uc801\uc778 \uba38\uc2e0\ub7ec\ub2dd \uc811\uadfc\ubc95\uc5d0\uc11c\ub294 \ubaa8\ub378\uc774 \ud2b9\uc815 \ub370\uc774\ud130\uc14b\uc5d0\uc11c \ud6c8\ub828\ub41c \ud6c4, \uc0c8\ub85c\uc6b4 \ub370\uc774\ud130\ub97c \ucd94\uac00\ub85c \ud559\uc2b5\uc2dc\ud0a4\uba74 \uae30\uc874\uc758 \ud559\uc2b5 \uacb0\uacfc\uac00 \ud6fc\uc190\ub420 \uc218 \uc788\uc2b5\ub2c8\ub2e4. \uc774\ub97c \uc8fd\uc74c\uc758 Catastrophic Forgetting<\/strong>\uc774\ub77c\uace0 \ud569\ub2c8\ub2e4.<\/p>\n \uc9c0\uc18d \ud559\uc2b5\uc740 \ubaa8\ub378\uc774 \uc5ec\ub7ec \uc791\uc5c5\uc744 \uc21c\ucc28\uc801\uc73c\ub85c \uc218\ud589\ud558\uba74\uc11c \uc774\uc804\uc5d0 \ud559\uc2b5\ud55c \ub0b4\uc6a9\uc744 \ubcf4\uc874\ud558\ub294 \ub2a5\ub825\uc744 \uc758\ubbf8\ud569\ub2c8\ub2e4. \uc774\ub294 \ud2b9\ud788 \uc790\uc728\uc8fc\ud589\ucc28, \ub85c\ubd07, \uac00\uc0c1 \ube44\uc11c \ub4f1\uc5d0\uc11c \ud544\uc218\uc801\uc785\ub2c8\ub2e4. \uc9c0\uc18d \ud559\uc2b5\uc758 \uc8fc\uc694 \ubaa9\ud45c\ub294 \ubaa8\ub378\uc774 \uc0c8\ub85c\uc6b4 \uc815\ubcf4\ub97c \ud6a8\uacfc\uc801\uc73c\ub85c \ud559\uc2b5\ud558\ub294 \ub3d9\uc2dc\uc5d0 \uae30\uc874\uc758 \uc9c0\uc2dd\uc744 \uc783\uc9c0 \uc54a\ub3c4\ub85d \ud558\ub294 \uac83\uc785\ub2c8\ub2e4.<\/p>\n \ud3c9\uc0dd \ud559\uc2b5\uc740 \ud559\uc2b5\uc774 \uc77c\ud68c\uc131\uc774 \uc544\ub2cc \uc9c0\uc18d\uc801\uc778 \uacfc\uc815\uc744 \uc758\ubbf8\ud558\uba70, \ub2e4\uc591\ud55c \uc791\uc5c5\uacfc \ud658\uacbd\uc5d0\uc11c \uc9c0\uc2dd\uc744 \uc313\uc544\uac00\ub294 \uac83\uc744 \ubaa9\ud45c\ub85c \ud569\ub2c8\ub2e4. \ud3c9\uc0dd \ud559\uc2b5\uc740 \ud2b9\uc815 \uc791\uc5c5\uc5d0 \uad6d\ud55c\ub418\uc9c0 \uc54a\uc73c\uba70, \uc2dc\uac04\uc774 \uc9c0\ub0a8\uc5d0 \ub530\ub77c \uc0c8\ub85c\uc6b4 \uc791\uc5c5\uc774\ub098 \ub3c4\uba54\uc778\uc5d0\uc11c\ub3c4 \ud559\uc2b5\ud560 \uc218 \uc788\ub294 \ub2a5\ub825\uc744 \ud3ec\ud568\ud569\ub2c8\ub2e4.<\/p>\n \ud3c9\uc0dd \ud559\uc2b5\uc5d0\ub294 \uba87 \uac00\uc9c0 \uc8fc\uc694 \uc811\uadfc\ubc95\uc774 \uc788\uc2b5\ub2c8\ub2e4.<\/p>\n Elastic Weight Consolidation (EWC)\ub294 \uc9c0\uc18d \ud559\uc2b5\uc758 \ud55c \uae30\ubc95\uc73c\ub85c, \uacfc\uac70\uc5d0 \ud559\uc2b5\ud55c \uc791\uc5c5\uc758 \uc911\uc694\uc131\uc744 \uace0\ub824\ud558\uc5ec \uc0c8\ub85c\uc6b4 \uc791\uc5c5\uc744 \ud559\uc2b5\ud560 \ub54c \uae30\uc874\uc758 \ud30c\ub77c\ubbf8\ud130\uc758 \ubcc0\ud654\ub97c \ucd5c\uc18c\ud654\ud558\ub294 \ubc29\ubc95\uc785\ub2c8\ub2e4. EWC\ub294 \ud1a0\ub9c8\uc2a4 \ud5e8\ub9ac\uc758 “End-To-End Learning for Robotic Manipulation with EWC”<\/em>\uc5d0\uc11c \ucc98\uc74c \uc81c\uc548\ub418\uc5c8\uc2b5\ub2c8\ub2e4.<\/p>\n EWC\ub294 \uac01 \ud30c\ub77c\ubbf8\ud130\uc758 \uc911\uc694\ub3c4\ub97c \ud3c9\uac00\ud558\uae30 \uc704\ud574 Fisher \uc815\ubcf4 \ud589\ub82c\uc744 \uc0ac\uc6a9\ud569\ub2c8\ub2e4. Fisher \uc815\ubcf4\ub294 \ubaa8\ub378\uc758 \ud30c\ub77c\ubbf8\ud130\uac00 \ud2b9\uc815 \uc791\uc5c5\uc5d0 \uc5bc\ub9c8\ub098 \uc911\uc694\ud55c\uc9c0\ub97c \ub098\ud0c0\ub0b4\ub294 \uc9c0\ud45c\uc785\ub2c8\ub2e4. EWC\ub294 \ub2e4\uc74c\uacfc \uac19\uc740 \uc190\uc2e4 \ud568\uc218\ub97c \uae30\ubc18\uc73c\ub85c \ubaa8\ub378\uc744 \ud559\uc2b5\ud569\ub2c8\ub2e4:<\/p>\n \uc5ec\uae30\uc11c:<\/p>\n EWC\uc758 \uc8fc\uc694 \uc7a5\uc810\uc740 \ub2e4\uc74c\uacfc \uac19\uc2b5\ub2c8\ub2e4:<\/p>\n \ud558\uc9c0\ub9cc EWC\uc5d0\ub3c4 \ub2e8\uc810\uc774 \uc874\uc7ac\ud569\ub2c8\ub2e4:<\/p>\n \uc544\ub798\ub294 PyTorch\ub97c \uc0ac\uc6a9\ud558\uc5ec EWC\ub97c \uad6c\ud604\ud558\ub294 \uac04\ub2e8\ud55c \uc608\uc81c\uc785\ub2c8\ub2e4.<\/p>\n \uc9c0\uc18d \ud559\uc2b5\uacfc \ud3c9\uc0dd \ud559\uc2b5\uc740 \uc778\uacf5\uc9c0\ub2a5\uc774 \uc778\uac04\uacfc \uc720\uc0ac\ud55c \ubc29\uc2dd\uc73c\ub85c \ud559\uc2b5\ud558\uae30 \uc704\ud55c \uc911\uc694\ud55c \uc5f0\uad6c \ubd84\uc57c\uc785\ub2c8\ub2e4. EWC\uc640 \uac19\uc740 \uae30\ubc95\ub4e4\uc744 \ud1b5\ud574 \uc778\uacf5\uc9c0\ub2a5 \ubaa8\ub378\uc774 \uc2dc\uac04\uc5d0 \ub530\ub77c \ubcc0\ud654\ud558\ub294 \ub2e4\uc591\ud55c \uc791\uc5c5\uc744 \ud6a8\uacfc\uc801\uc73c\ub85c \ud559\uc2b5\ud560 \uc218 \uc788\ub3c4\ub85d \ud560 \uc218 \uc788\uc2b5\ub2c8\ub2e4. \uc55e\uc73c\ub85c \uc774\ub7ec\ud55c \uc5f0\uad6c\uac00 \uc778\uacf5\uc9c0\ub2a5 \uae30\uc220\uc758 \ubc1c\uc804\uc5d0 \uae30\uc5ec\ud560 \uac83\uc73c\ub85c \uae30\ub300\ub418\uba70, \uc2e4\uc2dc\uac04 \ub370\uc774\ud130\uc758 \ubcc0\ud654\uc5d0 \ubbfc\ucca9\ud558\uac8c \ubc18\uc751\ud560 \uc218 \uc788\ub294 \uc778\uacf5\uc9c0\ub2a5 \ubaa8\ub378\uc744 \uac1c\ubc1c\ud558\ub294 \ub370 \uc911\uc694\ud55c \uc5ed\ud560\uc744 \ud560 \uac83\uc785\ub2c8\ub2e4.<\/p>\n\uc9c0\uc18d \ud559\uc2b5(Continual Learning)<\/h2>\n
\uc9c0\uc18d \ud559\uc2b5\uc758 \uc8fc\uc694 \ub3c4\uc804 \uacfc\uc81c<\/h3>\n
\n
\ud3c9\uc0dd \ud559\uc2b5(Lifelong Learning)<\/h2>\n
\ud3c9\uc0dd \ud559\uc2b5\uc5d0\uc11c\uc758 \uc811\uadfc\ubc95<\/h3>\n
\n
Elastic Weight Consolidation (EWC)<\/h2>\n
EWC\uc758 \uc791\ub3d9 \uc6d0\ub9ac<\/h3>\n
Loss = L(data) + \u03bb * \u03a3 (Fisher_i * (\u03b8 - \u03b8_old_i)^2)<\/code><\/pre>\n\n
L(data)<\/code>\ub294 \ud604\uc7ac \ub370\uc774\ud130\ub97c \uae30\ubc18\uc73c\ub85c \ud55c \uae30\ubcf8 \uc190\uc2e4 \ud568\uc218<\/li>\n\u03bb<\/code>\ub294 EWC\uc758 \ube44\uc911\uc744 \uc870\uc808\ud558\ub294 \ud558\uc774\ud37c\ud30c\ub77c\ubbf8\ud130<\/li>\nFisher_i<\/code>\ub294 i\ubc88\uc9f8 \ud30c\ub77c\ubbf8\ud130\uc5d0 \ub300\ud55c Fisher \uc815\ubcf4<\/li>\n\u03b8<\/code>\ub294 \ud604\uc7ac \ud30c\ub77c\ubbf8\ud130 \uac12<\/li>\n\u03b8_old_i<\/code>\ub294 \uc774\uc804 \uc791\uc5c5\uc5d0\uc11c \ud559\uc2b5\ud55c \ud30c\ub77c\ubbf8\ud130 \uac12<\/li>\n<\/ul>\nEWC\uc758 \uc7a5\uc810\uacfc \ub2e8\uc810<\/h3>\n
\n
\n
Python\uc744 \ud1b5\ud55c EWC \uad6c\ud604 \uc608\uc81c<\/h2>\n
import torch\nimport torch.nn as nn\nimport torch.optim as optim\n\nclass SimpleNN(nn.Module):\n def __init__(self):\n super(SimpleNN, self).__init__()\n self.fc1 = nn.Linear(2, 5)\n self.fc2 = nn.Linear(5, 1)\n\n def forward(self, x):\n x = torch.relu(self.fc1(x))\n x = self.fc2(x)\n return x\n\ndef compute_fisher(model, data, target):\n model.eval()\n fisher = {}\n for name, param in model.named_parameters():\n fisher[name] = torch.zeros_like(param)\n\n criterion = nn.MSELoss()\n for input, target in zip(data, target):\n model.zero_grad()\n output = model(input)\n loss = criterion(output, target)\n loss.backward()\n for name, param in model.named_parameters():\n fisher[name] += param.grad.data ** 2\n\n return fisher\n\ndef ewc_loss(model, fisher, old_params, lambda_=0.1):\n loss = 0\n for name, param in model.named_parameters():\n if name in fisher:\n loss += (fisher[name] * (param - old_params[name]) ** 2).sum()\n return lambda_ * loss\n\n# \ub370\uc774\ud130\uc640 \ubaa8\ub378 \ucd08\uae30\ud654\nmodel = SimpleNN()\noptimizer = optim.Adam(model.parameters(), lr=0.001)\n\n# \uc774\uc804 \uc791\uc5c5\uc758 \ud30c\ub77c\ubbf8\ud130 \uc800\uc7a5\nold_params = {name: param.data.clone() for name, param in model.named_parameters()}\n\n# \uc0c8\ub85c\uc6b4 \uc791\uc5c5 \ud559\uc2b5\ndata = torch.randn(100, 2)\ntarget = data.sum(dim=1, keepdim=True)\noptimizer.zero_grad()\noutput = model(data)\nloss = nn.MSELoss()(output, target)\nloss.backward()\noptimizer.step()\n\n# Fisher \uc815\ubcf4 \uacc4\uc0b0\nfisher = compute_fisher(model, data, target)\n\n# EWC \uc190\uc2e4 \uacc4\uc0b0\nloss_ewc = ewc_loss(model, fisher, old_params)\ntotal_loss = loss + loss_ewc\ntotal_loss.backward()\noptimizer.step()<\/code><\/pre>\n\ub9c8\ubb34\ub9ac\ud558\uba70<\/h3>\n
\ucc38\uace0\uc790\ub8cc<\/h2>\n