ResNets首次反超有監(jiān)督學習!DeepMind用自監(jiān)督實現(xiàn)逆襲,無需標注

??新智元報道??

??新智元報道??
編輯:好困 袁榭
【新智元導讀】近日,DeepMind又整了個新活:RELIC第二代!首次用自監(jiān)督學習實現(xiàn)了對有監(jiān)督學習的超越。莫非,今后真的不用標注數(shù)據(jù)了?



方法


for x in batch: # load a batch of B samples# Apply saliency mask and remove backgroundx_m = remove_background(x)for i in range(num_large_crops):# Select either original or background-removed# Image with probability p_mx = Bernoulli(p_m) ? x_m : x# Do large random crop and augmentxl_i = aug(crop_l(x))ol_i = f_o(xl_i)tl_i = g_t(xl_i)for i in range(num_small_crops):# Do small random crop and augmentxs_i = aug(crop_s(x))# Small crops only go through the online networkos_i = f_o(xs_i)loss = 0# Compute loss between all pairs of large cropsfor i in range(num_large_crops):for j in range(num_large_crops):loss += loss_relicv2(ol_i, tl_j, n_e)# Compute loss between small crops and large cropsfor i in range(num_small_crops):for j in range(num_large_crops):loss += loss_relicv2(os_i, tl_j, n_e)scale = (num_large_crops + num_small_crops) * num_large_cropsloss /= scale# Compute grads, update online and target networksloss.backward()update(f_o)g_t = gamma * g_t + (1 - gamma) * f_oRELICv2的偽代碼
結(jié)果






分析



結(jié)論
參考資料:
https://arxiv.org/abs/2201.05119

評論
圖片
表情
