{"id":2025,"date":"2026-07-04T10:03:06","date_gmt":"2026-07-04T03:03:06","guid":{"rendered":"https:\/\/blog.datacore.vn\/?p=2025"},"modified":"2026-07-04T12:43:16","modified_gmt":"2026-07-04T05:43:16","slug":"thien-kien-ket-cau-cnn-imagenet","status":"publish","type":"post","link":"https:\/\/blog.datacore.vn\/vi\/thien-kien-ket-cau-cnn-imagenet\/","title":{"rendered":"Thi\u00ean Ki\u1ebfn K\u1ebft C\u1ea5u Trong M\u1ea1ng N\u01a1-ron: B\u00e0i H\u1ecdc T\u1eeb CNN Hu\u1ea5n Luy\u1ec7n Tr\u00ean ImageNet"},"content":{"rendered":"

T\u00f3m t\u1eaft nhanh:<\/strong> M\u1ed9t nghi\u00ean c\u1ee9u quan tr\u1ecdng c\u1ee7a Geirhos v\u00e0 c\u1ed9ng s\u1ef1 (2018) cho th\u1ea5y c\u00e1c m\u00f4 h\u00ecnh \u1ea3nh ti\u00eau chu\u1ea9n nh\u1eadn di\u1ec7n v\u1eadt th\u1ec3 ch\u1ee7 y\u1ebfu d\u1ef1a v\u00e0o k\u1ebft c\u1ea5u b\u1ec1 m\u1eb7t, trong khi con ng\u01b0\u1eddi d\u1ef1a v\u00e0o h\u00ecnh d\u1ea1ng. Thi\u00ean ki\u1ebfn k\u1ebft c\u1ea5u n\u00e0y gi\u1ea3i th\u00edch nhi\u1ec1u th\u1ea5t b\u1ea1i c\u1ee7a th\u1ecb gi\u00e1c m\u00e1y t\u00ednh, v\u00e0 b\u00e0i b\u00e1o \u0111\u01b0a ra m\u1ed9t c\u00e1ch kh\u1eafc ph\u1ee5c th\u1ef1c t\u1ebf: hu\u1ea5n luy\u1ec7n l\u1ea1i m\u1ea1ng tr\u00ean m\u1ed9t t\u1eadp d\u1eef li\u1ec7u \u0111\u00e3 b\u1ecb lo\u1ea1i b\u1ecf t\u00edn hi\u1ec7u k\u1ebft c\u1ea5u, gi\u00fap m\u00f4 h\u00ecnh chuy\u1ec3n sang thi\u00ean ki\u1ebfn h\u00ecnh d\u1ea1ng gi\u1ed1ng con ng\u01b0\u1eddi, c\u1ea3i thi\u1ec7n c\u1ea3 \u0111\u1ed9 ch\u00ednh x\u00e1c l\u1eabn \u0111\u1ed9 b\u1ec1n v\u1eefng.<\/p>\n\n\n\n

\"Thi\u00ean<\/figure>\n\n\n\n

Thi\u00ean ki\u1ebfn k\u1ebft c\u1ea5u trong m\u1ea1ng n\u01a1-ron t\u00edch ch\u1eadp l\u00e0 g\u00ec?<\/h2>\n\n\n\n

M\u1ea1ng n\u01a1-ron t\u00edch ch\u1eadp (CNN) l\u00e0 m\u00f4 h\u00ecnh h\u1ecdc s\u00e2u h\u1ecdc c\u00e1c \u0111\u1eb7c tr\u01b0ng th\u1ecb gi\u00e1c t\u1eeb \u1ea3nh qua nhi\u1ec1u l\u1edbp b\u1ed9 l\u1ecdc x\u1ebfp ch\u1ed3ng. Gi\u1ea3 \u0111\u1ecbnh ph\u1ed5 bi\u1ebfn l\u00e0 CNN x\u00e2y d\u1ef1ng hi\u1ec3u bi\u1ebft theo t\u1ea7ng, t\u1eeb c\u1ea1nh v\u00e0 m\u00e0u s\u1eafc l\u00ean t\u1edbi h\u00ecnh d\u1ea1ng ph\u1ee9c t\u1ea1p nh\u01b0 \u00f4 t\u00f4 hay con ch\u00f3. Thi\u00ean ki\u1ebfn k\u1ebft c\u1ea5u l\u00e0 ph\u00e1t hi\u1ec7n \u0111i ng\u01b0\u1ee3c l\u1ea1i b\u1ee9c tranh \u0111\u00f3.<\/p>\n\n\n\n

N\u00f3 m\u00f4 t\u1ea3 xu h\u01b0\u1edbng c\u1ee7a m\u1ed9t CNN hu\u1ea5n luy\u1ec7n tr\u00ean ImageNet ra quy\u1ebft \u0111\u1ecbnh d\u1ef1a v\u00e0o k\u1ebft c\u1ea5u c\u1ee5c b\u1ed9, t\u1ee9c hoa v\u0103n b\u1ec1 m\u1eb7t c\u1ee7a v\u1eadt th\u1ec3, thay v\u00ec h\u00ecnh d\u1ea1ng t\u1ed5ng th\u1ec3. \u0110i\u1ec1u n\u00e0y quan tr\u1ecdng v\u00ec k\u1ebft c\u1ea5u v\u00e0 h\u00ecnh d\u1ea1ng th\u01b0\u1eddng tr\u00f9ng kh\u1edbp, n\u00ean m\u00f4 h\u00ecnh c\u00f3 th\u1ec3 \u0111\u00fang v\u00ec l\u00fd do sai. Thi\u00ean ki\u1ebfn k\u1ebft c\u1ea5u ch\u1ec9 l\u1ed9 ra khi hai t\u00edn hi\u1ec7u n\u00e0y b\u1ecb \u0111\u1eb7t v\u00e0o th\u1ebf xung \u0111\u1ed9t.<\/p>\n\n\n\n

C\u00e1c nh\u00e0 nghi\u00ean c\u1ee9u \u0111o thi\u00ean ki\u1ebfn k\u1ebft c\u1ea5u b\u1eb1ng c\u00e1ch n\u00e0o?<\/h2>\n\n\n\n

Nh\u00f3m nghi\u00ean c\u1ee9u d\u00f9ng c\u00e1c \u1ea3nh xung \u0111\u1ed9t t\u00edn hi\u1ec7u t\u1ea1o b\u1eb1ng k\u1ef9 thu\u1eadt chuy\u1ec3n phong c\u00e1ch. H\u1ecd l\u1ea5y h\u00ecnh d\u1ea1ng c\u1ee7a m\u1ed9t v\u1eadt, v\u00ed d\u1ee5 con m\u00e8o, r\u1ed3i ph\u1ee7 l\u00ean k\u1ebft c\u1ea5u c\u1ee7a m\u1ed9t v\u1eadt kh\u00e1c, v\u00ed d\u1ee5 da voi. M\u1ed9t ng\u01b0\u1eddi nh\u00ecn \u1ea3nh n\u00e0y v\u1eabn tr\u1ea3 l\u1eddi con m\u00e8o, v\u00ec con ng\u01b0\u1eddi ph\u00e2n lo\u1ea1i theo h\u00ecnh d\u1ea1ng.<\/p>\n\n\n\n

Ng\u01b0\u1ee3c l\u1ea1i, m\u1ed9t CNN hu\u1ea5n luy\u1ec7n tr\u00ean ImageNet ti\u00eau chu\u1ea9n c\u00f3 xu h\u01b0\u1edbng tr\u1ea3 l\u1eddi con voi, cho th\u1ea5y k\u1ebft c\u1ea5u ch\u1ee9 kh\u00f4ng ph\u1ea3i h\u00ecnh d\u1ea1ng \u0111ang d\u1eabn d\u1eaft d\u1ef1 \u0111o\u00e1n. Qua nhi\u1ec1u \u1ea3nh nh\u01b0 v\u1eady v\u00e0 nhi\u1ec1u ki\u1ebfn tr\u00fac n\u1ed5i ti\u1ebfng nh\u01b0 ResNet, VGG, GoogLeNet v\u00e0 AlexNet, m\u00f4 th\u1ee9c n\u00e0y nh\u1ea5t qu\u00e1n. Thi\u00ean ki\u1ebfn k\u1ebft c\u1ea5u v\u00ec th\u1ebf kh\u00f4ng ph\u1ea3i n\u00e9t ri\u00eang c\u1ee7a m\u1ed9t m\u00f4 h\u00ecnh m\u00e0 l\u00e0 \u0111\u1eb7c t\u00ednh chung c\u1ee7a c\u00e1ch c\u00e1c m\u1ea1ng h\u1ecdc t\u1eeb ImageNet.<\/p>\n\n\n\n

V\u00ec sao thi\u00ean ki\u1ebfn k\u1ebft c\u1ea5u quan tr\u1ecdng v\u1edbi AI \u0111\u00e1ng tin c\u1eady?<\/h2>\n\n\n\n

Thi\u00ean ki\u1ebfn k\u1ebft c\u1ea5u kh\u00f4ng ch\u1ec9 l\u00e0 chuy\u1ec7n h\u1ecdc thu\u1eadt. M\u1ed9t m\u00f4 h\u00ecnh d\u1ef1a v\u00e0o k\u1ebft c\u1ea5u s\u1ebd mong manh trong th\u1ef1c t\u1ebf. Thay \u0111\u1ed5i th\u1ed1ng k\u00ea b\u1ec1 m\u1eb7t c\u1ee7a \u1ea3nh, qua nhi\u1ec5u, m\u1edd, th\u1eddi ti\u1ebft, n\u00e9n hay \u0111\u1ed5i phong c\u00e1ch, m\u1ed9t m\u1ea1ng ph\u1ee5 thu\u1ed9c k\u1ebft c\u1ea5u c\u00f3 th\u1ec3 ph\u00e2n lo\u1ea1i sai m\u1ed9t v\u1eadt th\u1ec3 m\u00e0 h\u00ecnh d\u1ea1ng v\u1eabn c\u00f2n nguy\u00ean v\u1eb9n. Con ng\u01b0\u1eddi h\u1ea7u nh\u01b0 kh\u00f4ng b\u1ecb \u1ea3nh h\u01b0\u1edfng v\u00ec h\u00ecnh d\u1ea1ng \u0111\u01b0\u1ee3c gi\u1eef nguy\u00ean.<\/p>\n\n\n\n

\u0110i\u1ec1u n\u00e0y g\u1eafn thi\u00ean ki\u1ebfn k\u1ebft c\u1ea5u tr\u1ef1c ti\u1ebfp v\u1edbi \u0111\u1ed9 b\u1ec1n v\u1eefng. Nhi\u1ec1u th\u1ea5t b\u1ea1i \u0111\u01b0\u1ee3c ghi nh\u1eadn c\u1ee7a h\u1ec7 th\u1ed1ng th\u1ecb gi\u00e1c m\u00e1y t\u00ednh, t\u1eeb \u0111\u1ecdc sai \u1ea3nh b\u1ecb bi\u1ebfn d\u1ea1ng t\u1edbi s\u1ee5p \u0111\u1ed5 tr\u01b0\u1edbc c\u00e1c nhi\u1ec5u nh\u1ecf, \u0111\u1ec1u nh\u1ea5t qu\u00e1n v\u1edbi m\u1ed9t m\u00f4 h\u00ecnh h\u1ecdc l\u1ed1i t\u1eaft k\u1ebft c\u1ea5u thay v\u00ec c\u1ea5u tr\u00fac v\u1eadt th\u1ec3. V\u1edbi c\u00e1c \u1ee9ng d\u1ee5ng quan tr\u1ecdng nh\u01b0 \u1ea3nh y t\u1ebf, xe t\u1ef1 l\u00e1i hay ki\u1ec3m tra c\u00f4ng nghi\u1ec7p, m\u1ed9t m\u00f4 h\u00ecnh nh\u1eadn bi\u1ebft h\u00ecnh d\u1ea1ng \u0111\u00e1ng tin h\u01a1n nhi\u1ec1u.<\/p>\n\n\n\n

\"Gi\u1ea3m<\/figure>\n\n\n\n

L\u00e0m sao gi\u1ea3m thi\u00ean ki\u1ebfn k\u1ebft c\u1ea5u trong m\u00f4 h\u00ecnh?<\/h2>\n\n\n\n

\u0110\u00f3ng g\u00f3p h\u1eefu \u00edch nh\u1ea5t c\u1ee7a b\u00e0i b\u00e1o l\u00e0 m\u1ed9t gi\u1ea3i ph\u00e1p. Nh\u00f3m t\u00e1c gi\u1ea3 x\u00e2y d\u1ef1ng t\u1eadp hu\u1ea5n luy\u1ec7n m\u1edbi t\u00ean Stylized-ImageNet, t\u1ea1o ra b\u1eb1ng c\u00e1ch \u00e1p d\u1ee5ng chuy\u1ec3n phong c\u00e1ch l\u00ean m\u1ecdi \u1ea3nh ImageNet \u0111\u1ec3 k\u1ebft c\u1ea5u g\u1ed1c b\u1ecb thay b\u1eb1ng phong c\u00e1ch ngh\u1ec7 thu\u1eadt ng\u1eabu nhi\u00ean. V\u00ec k\u1ebft c\u1ea5u b\u1ecb ng\u1eabu nhi\u00ean h\u00f3a, n\u00f3 kh\u00f4ng c\u00f2n l\u00e0 t\u00edn hi\u1ec7u \u0111\u00e1ng tin, n\u00ean m\u1ea1ng bu\u1ed9c ph\u1ea3i d\u1ef1a v\u00e0o h\u00ecnh d\u1ea1ng \u0111\u1ec3 nh\u1eadn di\u1ec7n.<\/p>\n\n\n\n

K\u1ebft qu\u1ea3 r\u1ea5t \u1ea5n t\u01b0\u1ee3ng. M\u1ed9t m\u1ea1ng ResNet-50 hu\u1ea5n luy\u1ec7n tr\u00ean Stylized-ImageNet ph\u00e1t tri\u1ec3n thi\u00ean ki\u1ebfn h\u00ecnh d\u1ea1ng m\u1ea1nh h\u01a1n h\u1eb3n, g\u1ea7n v\u1edbi c\u1ea3m nh\u1eadn c\u1ee7a con ng\u01b0\u1eddi. N\u00f3 c\u0169ng b\u1ec1n v\u1eefng h\u01a1n tr\u01b0\u1edbc nhi\u1ec1u lo\u1ea1i bi\u1ebfn d\u1ea1ng v\u00e0 h\u01b0 h\u1ea1i \u1ea3nh ch\u01b0a t\u1eebng th\u1ea5y khi hu\u1ea5n luy\u1ec7n, v\u00e0 c\u00f2n c\u1ea3i thi\u1ec7n ph\u00e1t hi\u1ec7n v\u1eadt th\u1ec3 khi d\u00f9ng l\u00e0m m\u1ea1ng n\u1ec1n. Gi\u1ea3m thi\u00ean ki\u1ebfn k\u1ebft c\u1ea5u, do \u0111\u00f3, kh\u00f4ng l\u00e0m m\u1ea5t \u0111\u1ed9 ch\u00ednh x\u00e1c m\u00e0 t\u0103ng kh\u1ea3 n\u0103ng ch\u1ed1ng ch\u1ecbu.<\/p>\n\n\n\n

V\u1edbi ng\u01b0\u1eddi l\u00e0m th\u1ef1c t\u1ebf, b\u00e0i h\u1ecdc l\u00e0 m\u1ed9t b\u1ed9 c\u00f4ng c\u1ee5 c\u1ee5 th\u1ec3. T\u0103ng c\u01b0\u1eddng d\u1eef li\u1ec7u thay \u0111\u1ed5i k\u1ebft c\u1ea5u v\u00e0 phong c\u00e1ch, hu\u1ea5n luy\u1ec7n tr\u00ean d\u1eef li\u1ec7u \u0111\u00e3 ng\u1eabu nhi\u00ean h\u00f3a k\u1ebft c\u1ea5u, v\u00e0 \u0111\u00e1nh gi\u00e1 tr\u00ean c\u00e1c ph\u00e9p th\u1eed xung \u0111\u1ed9t t\u00edn hi\u1ec7u v\u00e0 h\u01b0 h\u1ea1i \u1ea3nh \u0111\u1ec1u gi\u00fap ph\u00e1t hi\u1ec7n v\u00e0 gi\u1ea3m thi\u00ean ki\u1ebfn k\u1ebft c\u1ea5u. B\u00e0i h\u1ecdc r\u1ed9ng h\u01a1n l\u00e0 m\u00f4 h\u00ecnh h\u1ecdc \u0111i\u1ec1u m\u00e0 d\u1eef li\u1ec7u l\u00e0m cho d\u1ec5, n\u00ean \u0111\u1ecbnh h\u00ecnh d\u1eef li\u1ec7u th\u01b0\u1eddng m\u1ea1nh h\u01a1n tinh ch\u1ec9nh ki\u1ebfn tr\u00fac.<\/p>\n\n\n\n

Thi\u00ean ki\u1ebfn k\u1ebft c\u1ea5u d\u1ea1y g\u00ec v\u1ec1 ch\u1ea5t l\u01b0\u1ee3ng d\u1eef li\u1ec7u?<\/h2>\n\n\n\n

B\u00e0i h\u1ecdc s\u00e2u nh\u1ea5t c\u1ee7a thi\u00ean ki\u1ebfn k\u1ebft c\u1ea5u l\u00e0 v\u1ec1 d\u1eef li\u1ec7u, kh\u00f4ng ph\u1ea3i m\u00f4 h\u00ecnh. M\u1ea1ng l\u00e0 c\u1ed7 m\u00e1y t\u00ecm l\u1ed1i t\u1eaft hi\u1ec7u qu\u1ea3. N\u1ebfu d\u1eef li\u1ec7u hu\u1ea5n luy\u1ec7n cho ph\u00e9p k\u1ebft c\u1ea5u gi\u1ea3i quy\u1ebft b\u00e0i to\u00e1n, m\u00f4 h\u00ecnh s\u1ebd d\u00f9ng k\u1ebft c\u1ea5u, ngay c\u1ea3 khi ta mong n\u00f3 h\u1ecdc h\u00ecnh d\u1ea1ng. \u0110\u00e2y l\u00e0 m\u1ed9t tr\u01b0\u1eddng h\u1ee3p c\u1ee7a h\u1ecdc l\u1ed1i t\u1eaft, khi m\u00f4 h\u00ecnh b\u00e1m v\u00e0o t\u00edn hi\u1ec7u b\u1ec1 m\u1eb7t t\u01b0\u01a1ng quan v\u1edbi nh\u00e3n thay v\u00ec kh\u00e1i ni\u1ec7m ta th\u1ef1c s\u1ef1 quan t\u00e2m.<\/p>\n\n\n\n

V\u00ec v\u1eady, c\u00e1c t\u1eadp d\u1eef li\u1ec7u \u0111\u01b0\u1ee3c tuy\u1ec3n ch\u1ecdn k\u1ef9, m\u00f4 t\u1ea3 r\u00f5 v\u00e0 c\u1ed1 \u00fd \u0111a d\u1ea1ng r\u1ea5t quan tr\u1ecdng. Hi\u1ec3u ph\u00e2n ph\u1ed1i c\u1ee7a m\u1ed9t t\u1eadp d\u1eef li\u1ec7u, c\u00e1c thi\u00ean l\u1ec7ch v\u00e0 \u0111i\u1ec3m m\u00f9 c\u1ee7a n\u00f3 l\u00e0 \u0111i\u1ec1u ki\u1ec7n ti\u00ean quy\u1ebft \u0111\u1ec3 x\u00e2y m\u00f4 h\u00ecnh kh\u00e1i qu\u00e1t t\u1ed1t. \u0110\u00e2y ch\u00ednh l\u00e0 l\u0129nh v\u1ef1c DataCore t\u1eadp trung. B\u1ea1n c\u00f3 th\u1ec3 tham kh\u1ea3o th\u00eam b\u00e0i v\u1ec1 b\u00f9ng n\u1ed5 trung t\u00e2m d\u1eef li\u1ec7u AI t\u1ea1i Vi\u1ec7t Nam<\/a> v\u00e0 kh\u00e1m ph\u00e1 n\u1ec1n t\u1ea3ng d\u1eef li\u1ec7u DataCore<\/a> v\u1edbi c\u00e1c t\u1eadp d\u1eef li\u1ec7u chu\u1ea9n h\u00f3a, d\u1ec5 \u0111\u1ecdc b\u1eb1ng m\u00e1y.<\/p>\n\n\n\n

\"Ch\u1ea5t<\/figure>\n\n\n\n
\"Nh\u1eadn<\/figure>\n\n\n\n

H\u00ecnh d\u1ea1ng, k\u1ebft c\u1ea5u v\u00e0 c\u00e1ch con ng\u01b0\u1eddi th\u1ef1c s\u1ef1 nh\u00ecn<\/h2>\n\n\n\n

Nhi\u1ec1u th\u1eadp k\u1ef7 nghi\u00ean c\u1ee9u tri gi\u00e1c cho th\u1ea5y con ng\u01b0\u1eddi nh\u1eadn di\u1ec7n v\u1eadt th\u1ec3 ph\u1ea7n l\u1edbn qua h\u00ecnh kh\u1ed1i t\u1ed5ng th\u1ec3 v\u00e0 c\u00e1ch s\u1eafp x\u1ebfp c\u00e1c b\u1ed9 ph\u1eadn. M\u1ed9t \u0111\u1ee9a tr\u1ebb c\u00f3 th\u1ec3 nh\u1eadn ra chi\u1ebfc gh\u1ebf v\u1ebd b\u1eb1ng n\u00e9t vi\u1ec1n \u0111\u01a1n gi\u1ea3n, kh\u00f4ng m\u00e0u, kh\u00f4ng chi ti\u1ebft b\u1ec1 m\u1eb7t, v\u00ec ch\u1ec9 ri\u00eang h\u00ecnh d\u1ea1ng \u0111\u00e3 mang danh t\u00ednh. B\u1ec1 m\u1eb7t c\u00f3 \u00edch nh\u01b0ng th\u1ee9 y\u1ebfu so v\u1edbi c\u1ea5u tr\u00fac.<\/p>\n\n\n\n

Tr\u01b0\u1edbc nghi\u00ean c\u1ee9u n\u00e0y, nhi\u1ec1u ng\u01b0\u1eddi gi\u1ea3 \u0111\u1ecbnh m\u1ea1ng s\u00e2u c\u0169ng h\u1ed9i t\u1ee5 v\u1ec1 \u0111i\u1ec1u t\u01b0\u01a1ng t\u1ef1, theo tr\u1ef1c gi\u00e1c r\u1eb1ng l\u1edbp \u0111\u1ea7u b\u1eaft c\u1ea1nh, l\u1edbp gi\u1eefa gh\u00e9p b\u1ed9 ph\u1eadn, l\u1edbp cu\u1ed1i bi\u1ec3u di\u1ec5n v\u1eadt th\u1ec3. C\u00e1c th\u00ed nghi\u1ec7m xung \u0111\u1ed9t t\u00edn hi\u1ec7u \u0111\u00e3 ch\u1ecdc th\u1ee7ng tr\u1ef1c gi\u00e1c \u0111\u00f3. Khi l\u1ed1i d\u1ec5 d\u1eabn t\u1edbi nh\u00e3n \u0111\u00fang \u0111i qua k\u1ebft c\u1ea5u, m\u1ea1ng ch\u1ecdn l\u1ed1i \u0111\u00f3, v\u00e0 thi\u00ean ki\u1ebfn k\u1ebft c\u1ea5u cho th\u1ea5y \u0111\u1ed9 ch\u00ednh x\u00e1c cao tr\u00ean benchmark c\u00f3 th\u1ec3 t\u1ed3n t\u1ea1i song song v\u1edbi m\u1ed9t bi\u1ec3u di\u1ec5n kh\u00e1c h\u1eb3n tri gi\u00e1c con ng\u01b0\u1eddi.<\/p>\n\n\n\n

H\u1ecdc l\u1ed1i t\u1eaft v\u00e0 m\u00f4 th\u1ee9c r\u1ed9ng h\u01a1n<\/h2>\n\n\n\n

Thi\u00ean ki\u1ebfn k\u1ebft c\u1ea5u l\u00e0 m\u1ed9t v\u00ed d\u1ee5 c\u1ee7a hi\u1ec7n t\u01b0\u1ee3ng r\u1ed9ng h\u01a1n g\u1ecdi l\u00e0 h\u1ecdc l\u1ed1i t\u1eaft. H\u1ebft tr\u01b0\u1eddng h\u1ee3p n\u00e0y t\u1edbi tr\u01b0\u1eddng h\u1ee3p kh\u00e1c, m\u00f4 h\u00ecnh ph\u00e1t hi\u1ec7n c\u00e1c t\u01b0\u01a1ng quan b\u1ec1 m\u1eb7t t\u00ecnh c\u1edd d\u1ef1 \u0111o\u00e1n \u0111\u01b0\u1ee3c nh\u00e3n tr\u00ean ph\u00e2n ph\u1ed1i hu\u1ea5n luy\u1ec7n nh\u01b0ng th\u1ea5t b\u1ea1i khi t\u01b0\u01a1ng quan \u0111\u00f3 v\u1ee1. M\u1ed9t b\u1ed9 ph\u00e2n lo\u1ea1i vi\u00eam ph\u1ed5i b\u00e1m v\u00e0o lo\u1ea1i m\u00e1y ch\u1ee5p, m\u1ed9t b\u1ed9 ph\u00e1t hi\u1ec7n \u0111\u1ed9ng v\u1eadt b\u00e1m v\u00e0o n\u1ec1n \u1ea3nh, \u0111\u1ec1u theo c\u00f9ng logic nh\u01b0 m\u1ea1ng b\u00e1m v\u00e0o k\u1ebft c\u1ea5u.<\/p>\n\n\n\n

Nh\u00ecn c\u00e1c tr\u01b0\u1eddng h\u1ee3p n\u00e0y nh\u01b0 m\u1ed9t h\u1ecd gi\u00fap l\u00e0m r\u00f5 v\u1ea5n \u0111\u1ec1. V\u1ea5n \u0111\u1ec1 hi\u1ebfm khi l\u00e0 m\u00f4 h\u00ecnh h\u1ecfng; v\u1ea5n \u0111\u1ec1 l\u00e0 d\u1eef li\u1ec7u \u0111\u00e3 m\u1eddi g\u1ecdi m\u1ed9t l\u1ed1i d\u1ec5 h\u01a1n t\u1edbi \u0111\u00e1p \u00e1n so v\u1edbi kh\u00e1i ni\u1ec7m ta mu\u1ed1n. C\u00e1ch s\u1eeda c\u0169ng t\u1eeb \u0111\u00f3: n\u1ebfu mu\u1ed1n m\u00f4 h\u00ecnh h\u1ecdc kh\u00e1i ni\u1ec7m, h\u00e3y x\u00e2y d\u1eef li\u1ec7u v\u00e0 ph\u00e9p \u0111\u00e1nh gi\u00e1 n\u01a1i l\u1ed1i t\u1eaft kh\u00f4ng c\u00f2n hi\u1ec7u qu\u1ea3, bu\u1ed9c m\u1ea1ng v\u00e0o t\u00edn hi\u1ec7u kh\u00f3 h\u01a1n nh\u01b0ng kh\u00e1i qu\u00e1t h\u01a1n.<\/p>\n\n\n\n

Quy tr\u00ecnh th\u1ef1c t\u1ebf cho c\u00e1c nh\u00f3m k\u1ef9 thu\u1eadt<\/h2>\n\n\n\n

Bi\u1ebfn nghi\u00ean c\u1ee9u th\u00e0nh th\u1ef1c h\u00e0nh theo m\u1ed9t tr\u00ecnh t\u1ef1 r\u00f5 r\u00e0ng. Tr\u01b0\u1edbc h\u1ebft, d\u00f2 t\u00ecm thi\u00ean ki\u1ebfn k\u1ebft c\u1ea5u: ch\u1ea1y th\u1eed xung \u0111\u1ed9t t\u00edn hi\u1ec7u v\u00e0 \u0111\u00e1nh gi\u00e1 tr\u00ean \u1ea3nh b\u1ecb h\u01b0 h\u1ea1i ho\u1eb7c ngo\u00e0i ph\u00e2n ph\u1ed1i \u0111\u1ec3 xem hi\u1ec7u n\u0103ng c\u00f3 s\u1ee5p khi th\u1ed1ng k\u00ea b\u1ec1 m\u1eb7t thay \u0111\u1ed5i hay kh\u00f4ng. K\u1ebf \u0111\u1ebfn, n\u1ebfu c\u00f3, h\u00e3y can thi\u1ec7p v\u00e0o d\u1eef li\u1ec7u b\u1eb1ng t\u0103ng c\u01b0\u1eddng phong c\u00e1ch v\u00e0 k\u1ebft c\u1ea5u ho\u1eb7c hu\u1ea5n luy\u1ec7n stylized.<\/p>\n\n\n\n

Th\u1ee9 ba, \u0111\u00e1nh gi\u00e1 l\u1ea1i tr\u00ean c\u00f9ng c\u00e1c ph\u00e9p th\u1eed v\u00e0 x\u00e1c nh\u1eadn \u0111\u1ed9 b\u1ec1n v\u1eefng c\u1ea3i thi\u1ec7n m\u00e0 kh\u00f4ng m\u1ea5t \u0111\u1ed9 ch\u00ednh x\u00e1c tr\u00ean d\u1eef li\u1ec7u s\u1ea1ch. Th\u1ee9 t\u01b0, ghi l\u1ea1i t\u00e0i li\u1ec7u v\u1ec1 t\u1eadp d\u1eef li\u1ec7u v\u00e0 c\u00e1c ph\u00e9p th\u1eed \u0111\u1ec3 nh\u00f3m sau k\u1ebf th\u1eeba tri th\u1ee9c. Xuy\u00ean su\u1ed1t, ch\u1ee7 \u0111\u1ec1 l\u1eb7p l\u1ea1i l\u00e0 \u0111o l\u01b0\u1eddng: b\u1ea1n kh\u00f4ng th\u1ec3 qu\u1ea3n l\u00fd m\u1ed9t thi\u00ean l\u1ec7ch m\u00e0 m\u00ecnh ch\u01b0a bao gi\u1edd ki\u1ec3m tra, v\u00e0 m\u1ed9t con s\u1ed1 ch\u00ednh x\u00e1c duy nh\u1ea5t s\u1ebd vui v\u1ebb che gi\u1ea5u n\u00f3.<\/p>\n\n\n\n

Gi\u1edbi h\u1ea1n v\u00e0 \u0111i\u1ec1u di\u1ec5n ra sau \u0111\u00f3<\/h2>\n\n\n\n

Nh\u01b0 m\u1ecdi nghi\u00ean c\u1ee9u, c\u00f4ng tr\u00ecnh n\u00e0y c\u00f3 ranh gi\u1edbi. Ph\u01b0\u01a1ng ph\u00e1p xung \u0111\u1ed9t t\u00edn hi\u1ec7u d\u00f9ng \u1ea3nh nh\u00e2n t\u1ea1o, v\u00e0 c\u1ea3nh th\u1ef1c t\u1ebf hi\u1ebfm khi \u0111\u1eb7t h\u00ecnh d\u1ea1ng ch\u1ecdi k\u1ebft c\u1ea5u s\u1ea1ch s\u1ebd nh\u01b0 v\u1eady, n\u00ean c\u00e1c con s\u1ed1 m\u00f4 t\u1ea3 m\u1ed9t ph\u00e9p d\u00f2 c\u00f3 ki\u1ec3m so\u00e1t ch\u1ee9 kh\u00f4ng ph\u1ea3i hi\u1ec7u n\u0103ng th\u01b0\u1eddng ng\u00e0y. C\u00f4ng tr\u00ecnh c\u0169ng t\u1eadp trung v\u00e0o m\u1ea1ng t\u00edch ch\u1eadp hu\u1ea5n luy\u1ec7n tr\u00ean ImageNet.<\/p>\n\n\n\n

D\u00f9 v\u1eady, h\u01b0\u1edbng c\u1ee7a ph\u00e1t hi\u1ec7n v\u1eabn v\u1eefng v\u00e0 \u1ea3nh h\u01b0\u1edfng nhi\u1ec1u nghi\u00ean c\u1ee9u v\u1ec1 sau, g\u1ed3m c\u00e1ch c\u00e1c ki\u1ebfn tr\u00fac m\u1edbi nh\u01b0 vision transformer c\u00e2n b\u1eb1ng h\u00ecnh d\u1ea1ng v\u00e0 k\u1ebft c\u1ea5u. Con s\u1ed1 c\u1ee5 th\u1ec3 thay \u0111\u1ed5i theo t\u1eebng nghi\u00ean c\u1ee9u, nh\u01b0ng th\u00f4ng \u0111i\u1ec7p trung t\u00e2m b\u1ec1n v\u1eefng: thi\u00ean ki\u1ebfn quy n\u1ea1p c\u1ee7a m\u00f4 h\u00ecnh th\u1ecb gi\u00e1c b\u1ecb \u0111\u1ecbnh h\u00ecnh m\u1ea1nh b\u1edfi d\u1eef li\u1ec7u v\u00e0 quy tr\u00ecnh hu\u1ea5n luy\u1ec7n, v\u00e0 c\u00f3 th\u1ec3 \u0111\u01b0\u1ee3c l\u00e1i m\u1ed9t c\u00e1ch ch\u1ee7 \u0111\u00edch. \u0110\u00f3 l\u00e0 m\u1eb7t l\u1ea1c quan c\u1ee7a c\u00e2u chuy\u1ec7n thi\u00ean ki\u1ebfn k\u1ebft c\u1ea5u.<\/p>\n\n\n\n

Thi\u00ean ki\u1ebfn k\u1ebft c\u1ea5u: c\u00e2u h\u1ecfi th\u01b0\u1eddng g\u1eb7p<\/h2>\n\n\n\n

Thi\u00ean ki\u1ebfn k\u1ebft c\u1ea5u l\u00e0 g\u00ec, n\u00f3i \u0111\u01a1n gi\u1ea3n?<\/h3>\n\n\n\n

Thi\u00ean ki\u1ebfn k\u1ebft c\u1ea5u l\u00e0 xu h\u01b0\u1edbng m\u00f4 h\u00ecnh \u1ea3nh nh\u1eadn di\u1ec7n v\u1eadt th\u1ec3 theo hoa v\u0103n b\u1ec1 m\u1eb7t thay v\u00ec h\u00ecnh d\u1ea1ng t\u1ed5ng th\u1ec3. Khi h\u00ecnh d\u1ea1ng v\u00e0 k\u1ebft c\u1ea5u b\u1ea5t \u0111\u1ed3ng, m\u00f4 h\u00ecnh thi\u00ean ki\u1ebfn k\u1ebft c\u1ea5u \u0111i theo k\u1ebft c\u1ea5u, c\u00f2n con ng\u01b0\u1eddi \u0111i theo h\u00ecnh d\u1ea1ng.<\/p>\n\n\n\n

C\u00f3 ph\u1ea3i m\u1ecdi CNN \u0111\u1ec1u c\u00f3 thi\u00ean ki\u1ebfn k\u1ebft c\u1ea5u?<\/h3>\n\n\n\n

Nghi\u00ean c\u1ee9u t\u00ecm th\u1ea5y thi\u00ean ki\u1ebfn k\u1ebft c\u1ea5u m\u1ea1nh \u1edf nhi\u1ec1u ki\u1ebfn tr\u00fac ti\u00eau chu\u1ea9n hu\u1ea5n luy\u1ec7n tr\u00ean ImageNet, n\u00ean \u0111\u00e2y l\u00e0 \u0111\u1eb7c t\u00ednh chung c\u1ee7a c\u00e1ch hu\u1ea5n luy\u1ec7n \u0111i\u1ec3n h\u00ecnh. M\u1ee9c \u0111\u1ed9 c\u00f3 th\u1ec3 thay \u0111\u1ed5i b\u1eb1ng c\u00e1ch \u0111\u1ed5i d\u1eef li\u1ec7u hu\u1ea5n luy\u1ec7n.<\/p>\n\n\n\n

\u0110\u1ecdc nghi\u00ean c\u1ee9u g\u1ed1c \u1edf \u0111\u00e2u?<\/h3>\n\n\n\n

Ph\u00e1t hi\u1ec7n \u0111\u1ebfn t\u1eeb Geirhos v\u00e0 c\u1ed9ng s\u1ef1, \"ImageNet-trained CNNs are Biased Towards Texture; Increasing Shape Bias Improves Accuracy and Robustness\", c\u00f3 tr\u00ean arXiv<\/a>.<\/p>\n\n\n\n

B\u00e0i vi\u1ebft t\u00f3m t\u1eaft nghi\u00ean c\u1ee9u h\u1ecdc s\u00e2u \u0111\u00e3 c\u00f4ng b\u1ed1 cho m\u1ee5c \u0111\u00edch gi\u00e1o d\u1ee5c. C\u00e1c ph\u00e1t hi\u1ec7n \u0111\u01b0\u1ee3c ghi nh\u1eadn cho nh\u00f3m t\u00e1c gi\u1ea3 g\u1ed1c n\u00eau tr\u00ean.<\/em><\/p>\n\n\n\n

Thi\u00ean ki\u1ebfn k\u1ebft c\u1ea5u c\u00f3 xu\u1ea5t hi\u1ec7n \u1edf c\u00e1c ki\u1ebfn tr\u00fac m\u1edbi nh\u01b0 Vision Transformer kh\u00f4ng?<\/strong> C\u00f3, nh\u01b0ng \u1edf m\u1ee9c \u0111\u1ed9 kh\u00e1c nhau. C\u00e1c nghi\u00ean c\u1ee9u ti\u1ebfp n\u1ed1i Geirhos 2018 cho th\u1ea5y Vision Transformer th\u01b0\u1eddng thi\u00ean v\u1ec1 h\u00ecnh d\u1ea1ng nhi\u1ec1u h\u01a1n m\u1ea1ng t\u00edch ch\u1eadp truy\u1ec1n th\u1ed1ng, song m\u1ee9c \u0111\u1ed9 ph\u1ee5 thu\u1ed9c k\u1ebft c\u1ea5u v\u1eabn thay \u0111\u1ed5i theo d\u1eef li\u1ec7u hu\u1ea5n luy\u1ec7n v\u00e0 ph\u01b0\u01a1ng ph\u00e1p t\u0103ng c\u01b0\u1eddng d\u1eef li\u1ec7u. N\u00f3i c\u00e1ch kh\u00e1c, thi\u00ean ki\u1ebfn k\u1ebft c\u1ea5u l\u00e0 thu\u1ed9c t\u00ednh c\u1ee7a c\u1ea3 ki\u1ebfn tr\u00fac l\u1eabn d\u1eef li\u1ec7u, n\u00ean m\u1ecdi quy tr\u00ecnh \u0111\u00e1nh gi\u00e1 m\u00f4 h\u00ecnh th\u1ecb gi\u00e1c m\u00e1y t\u00ednh v\u1eabn c\u1ea7n b\u00e0i ki\u1ec3m tra xung \u0111\u1ed9t h\u00ecnh d\u1ea1ng v\u00e0 k\u1ebft c\u1ea5u, b\u1ea5t k\u1ec3 ki\u1ebfn tr\u00fac n\u00e0o \u0111ang \u0111\u01b0\u1ee3c s\u1eed d\u1ee5ng.<\/p>","protected":false},"excerpt":{"rendered":"

Thi\u00ean ki\u1ebfn k\u1ebft c\u1ea5u: v\u00ec sao CNN hu\u1ea5n luy\u1ec7n tr\u00ean ImageNet nh\u1eadn di\u1ec7n theo k\u1ebft c\u1ea5u thay v\u00ec h\u00ecnh d\u1ea1ng, t\u00e1c h\u1ea1i t\u1edbi \u0111\u1ed9 b\u1ec1n v\u1eefng, v\u00e0 c\u00e1ch hu\u1ea5n luy\u1ec7n stylized t\u1ea1o thi\u00ean ki\u1ebfn h\u00ecnh d\u1ea1ng gi\u1ed1ng con ng\u01b0\u1eddi.<\/p>","protected":false},"author":19,"featured_media":0,"comment_status":"open","ping_status":"open","sticky":false,"template":"","format":"standard","meta":{"content-type":"","_uag_custom_page_level_css":"","_swt_meta_header_display":false,"_swt_meta_footer_display":false,"_swt_meta_site_title_display":false,"_swt_meta_sticky_header":false,"_swt_meta_transparent_header":false,"footnotes":""},"categories":[19,82],"tags":[965,1198,1201,1155,1204],"class_list":["post-2025","post","type-post","status-publish","format-standard","hentry","category-blog-vi","category-cong-nghe-vi","tag-dc-2026-w26-vi-vi","tag-imagenet-vi-vi","tag-mang-no-ron-vi","tag-thi-giac-may-tinh-vi","tag-thien-kien-ket-cau-vi"],"uagb_featured_image_src":{"full":false,"thumbnail":false,"medium":false,"medium_large":false,"large":false,"1536x1536":false,"2048x2048":false,"trp-custom-language-flag":false},"uagb_author_info":{"display_name":"DataCore Marketing","author_link":"https:\/\/blog.datacore.vn\/vi\/author\/datacore_marketing\/"},"uagb_comment_info":0,"uagb_excerpt":"Thi\u00ean ki\u1ebfn k\u1ebft c\u1ea5u: v\u00ec sao CNN hu\u1ea5n luy\u1ec7n tr\u00ean ImageNet nh\u1eadn di\u1ec7n theo k\u1ebft c\u1ea5u thay v\u00ec h\u00ecnh d\u1ea1ng, t\u00e1c h\u1ea1i t\u1edbi \u0111\u1ed9 b\u1ec1n v\u1eefng, v\u00e0 c\u00e1ch hu\u1ea5n luy\u1ec7n stylized t\u1ea1o thi\u00ean ki\u1ebfn h\u00ecnh d\u1ea1ng gi\u1ed1ng con ng\u01b0\u1eddi.","_links":{"self":[{"href":"https:\/\/blog.datacore.vn\/vi\/wp-json\/wp\/v2\/posts\/2025","targetHints":{"allow":["GET"]}}],"collection":[{"href":"https:\/\/blog.datacore.vn\/vi\/wp-json\/wp\/v2\/posts"}],"about":[{"href":"https:\/\/blog.datacore.vn\/vi\/wp-json\/wp\/v2\/types\/post"}],"author":[{"embeddable":true,"href":"https:\/\/blog.datacore.vn\/vi\/wp-json\/wp\/v2\/users\/19"}],"replies":[{"embeddable":true,"href":"https:\/\/blog.datacore.vn\/vi\/wp-json\/wp\/v2\/comments?post=2025"}],"version-history":[{"count":2,"href":"https:\/\/blog.datacore.vn\/vi\/wp-json\/wp\/v2\/posts\/2025\/revisions"}],"predecessor-version":[{"id":2111,"href":"https:\/\/blog.datacore.vn\/vi\/wp-json\/wp\/v2\/posts\/2025\/revisions\/2111"}],"wp:attachment":[{"href":"https:\/\/blog.datacore.vn\/vi\/wp-json\/wp\/v2\/media?parent=2025"}],"wp:term":[{"taxonomy":"category","embeddable":true,"href":"https:\/\/blog.datacore.vn\/vi\/wp-json\/wp\/v2\/categories?post=2025"},{"taxonomy":"post_tag","embeddable":true,"href":"https:\/\/blog.datacore.vn\/vi\/wp-json\/wp\/v2\/tags?post=2025"}],"curies":[{"name":"wp","href":"https:\/\/api.w.org\/{rel}","templated":true}]}}