No abstract

With the introduction of SNIP [40], it has been demonstrated that modern neural networks can effectively be pruned before training. Yet, its sensitivity criterion has since been criticized for not propagating training signal properly or even disconnecting layers. As a remedy, GraSP [70] was introduced, compromising on simplicity. However, in this work we show that by applying the sensitivity criterion iteratively in smaller steps -still before training -we can improve its performance without difficult implementation. As such, we introduce 'SNIP-it'.We then demonstrate how it can be applied for both structured and unstructured pruning, before and/or during training, therewith achieving state-of-the-art sparsityperformance trade-offs. That is, while already providing the computational benefits of pruning in the training process from the start. Furthermore, we evaluate our methods on robustness to overfitting, disconnection and adversarial attacks as well.Preprint. Under review.

Large-scale pretrained models, especially those trained from vision-language data have demonstrated the tremendous value that can be gained from both larger training datasets and models. Thus, in order to benefit from these developments, there is renewed interest in transfer learning and adapting models from large-scale general pretraining to particular downstream tasks. However, the continuously increasing size of the models means that even the classic approach of finetuning is becoming infeasible for all but big institutions. Prompt leaning has emerged as a flexible way to adapt models by solely learning additional inputs to a model that is kept frozen, but so far performances remained inferior to finetuning. To address this, we propose the Prompt Generation Network (PGN) that generates input-dependent prompts by sampling from a learned library of tokens. We show the PGN is effective in adapting pretrained models to various new datasets. It surpasses previous prompt-learning methods by a large margin and even fullfinetuning on 5 out of 12 datasets while requiring 100x less parameters. PGN can even be used for training and inferring on multiple datasets simultaneously and learns to allocate tokens between domains. Given these findings, we conclude that PGN is a viable and scalable approach for downstream adaptation of frozen models. Code is available at https://github.com/jochemloedeman/PGN.

No abstract

Instance-level contrastive learning techniques, which rely on data augmentation and a contrastive loss function, have found great success in the domain of visual representation learning. They are not suitable for exploiting the rich dynamical structure of video however, as operations are done on many augmented instances. In this paper we propose "Video Cross-Stream Prototypical Contrasting", a novel method which predicts consistent prototype assignments from both RGB and optical flow views, operating on sets of samples. Specifically, we alternate the optimization process; while optimizing one of the streams, all views are mapped to one set of stream prototype vectors. Each of the assignments is predicted with all views except the one matching the prediction, pushing representations closer to their assigned prototypes. As a result, more efficient video embeddings with ingrained motion information are learned, without the explicit need for optical flow computation during inference. We obtain state-of-the-art results on nearestneighbour video retrieval and action recognition, outperforming previous best by +3.2% on UCF101 using the S3D backbone (90.5% Top-1 acc), and by +7.2% on UCF101 and +15.1% on HMDB51 using the R(2+1)D backbone. 1 * Work is done during an internship at BrainCreators, Amsterdam.