FOSTER: Feature Boosting and Compression for Class-Incremental Learning

Adaptive Aggregation Networks for Class-Incremental Learning

“…Parameter-isolation-based methods increase the model parameters in each new incremental phase, to prevent knowledge forgetting caused by parameter overwritten. Some of them [19,38,44,47,48] proposed to progressively expand the size of the neural network to learn new coming data. Others [1,22,26,49] froze a part of network parameters (to maintain the old class knowledge) to alleviate the problem of knowledge overwriting.…”

“…A straightforward way to retain old class knowledge is keeping around a few old class exemplars in the memory and using them to re-train the model in subsequent phases. The number of exemplars is usually limited, e.g., 5 ∼ 20 exemplars per class [13,18,26,28,37,44,46,48,50], as the total memory in CIL strictly budgeted, e.g., 2k exemplars.…”

“…In particular, for each phase, we perform a local BOP with two steps to tune the parameters of the CIM model: 1) a temporary model is trained with the compressed exemplars as input; and 2) a validation loss on the uncom-pressed new data is computed and the gradients are backpropagated to optimize the parameters of CIM. To evaluate CIM, we conduct extensive experiments by plugging it in recent CIL methods, 3 LUCIR [18], DER [48], and FOSTER [44], on three high-resolution benchmarks, Food-101 [3], ImageNet-100 [18], and ImageNet-1000 [10]. We find that using the compressed exemplars by CIM brings consistent and significant improvements, e.g., 4.2% and 4.8% higher than the SOTA method FOSTER [44], respectively, in the 5-phase and 10-phase settings of ImageNet-1000, with a total memory budget for 5k exemplars.…”

“…To evaluate CIM, we conduct extensive experiments by plugging it in recent CIL methods, 3 LUCIR [18], DER [48], and FOSTER [44], on three high-resolution benchmarks, Food-101 [3], ImageNet-100 [18], and ImageNet-1000 [10]. We find that using the compressed exemplars by CIM brings consistent and significant improvements, e.g., 4.2% and 4.8% higher than the SOTA method FOSTER [44], respectively, in the 5-phase and 10-phase settings of ImageNet-1000, with a total memory budget for 5k exemplars.…”

Benchmarking Deepart Detection

“…To address the ODD benchmark, we revisit the state-of-the-art networks like vision transformer (ViT) Dosovitskiy et al (2020) with either fine-tuning or prompt-tuning strategies. For the first CDD benchmark that allows for the rehearsal on early encountered deeparts and LAION-5B conarts, we propose a sharing scheme to enable the failed general rehearsal-based methods like Foster Wang et al (2022a) to work again. For the second CDD benchmark that merely allows for the rehearsal on LAION conarts, we suggest a similar sharing method to fix the malfunctioned state-of-the-art exemplar-free methods like S-Prompts Wang et al (2022b).…”