Current federated learning algorithms take tens of communication rounds transmitting unwieldy model weights under ideal circumstances and hundreds when data is poorly distributed. Inspired by recent work on dataset distillation and distributed one-shot learning, we propose Distilled One-Shot Federated Learning (DOSFL) to significantly reduce the communication cost while achieving comparable performance. In just one round, each client distills their private dataset, sends the synthetic data (e.g. images or sentences) to the server, and collectively trains a global model. The distilled data look like noise and are only useful to the specific model weights, i.e., become useless after the model updates. With this weight-less and gradient-less design, the total communication cost of DOSFL is up to three orders of magnitude less than FedAvg while preserving between 93% to 99% performance of a centralized counterpart. Afterwards, clients could switch to traditional methods such as FedAvg to finetune the last few percent to fit personalized local models with local datasets. Through comprehensive experiments, we show the accuracy and communication performance of DOSFL on both vision and language tasks with different models including CNN, LSTM, Transformer, etc. We demonstrate that an eavesdropping attacker cannot properly train a good model using the leaked distilled data, without knowing the initial model weights. DOSFL serves as an inexpensive method to quickly converge on a performant pre-trained model with less than 0.1% communication cost of traditional methods.
We study the problem of dataset distillation - creating a small set of synthetic examples capable of training a good model. In particular, we study the problem of label distillation - creating synthetic labels for a small set of real images, and show it to be more effective than the prior image-based approach to dataset distillation. Methodologically, we introduce a more robust and flexible meta-learning algorithm for distillation, as well as an effective first-order strategy based on convex optimization layers. Distilling labels with our new algorithm leads to improved results over prior image-based distillation. More importantly, it leads to clear improvements in flexibility of the distilled dataset in terms of compatibility with off-the-shelf optimizers and diverse neural architectures. Interestingly, label distillation can also be applied across datasets, for example enabling learning Japanese character recognition by training only on synthetically labeled English letters.
Facial expression plays an important role in understanding human emotions. Most recently, deep learning based methods have shown promising for facial expression recognition. However, the performance of the current state-of-the-art facial expression recognition (FER) approaches is directly related to the labeled data for training. To solve this issue, prior works employ the pretrain-and-finetune strategy, i.e., utilize a large amount of unlabeled data to pretrain the network and then finetune it by the labeled data. As the labeled data is in a small amount, the final network performance is still restricted. From a different perspective, we propose to perform omni-supervised learning to directly exploit reliable samples in a large amount of unlabeled data for network training. Particularly, a new dataset is firstly constructed using a primitive model trained on a small number of labeled samples to select samples with high confidence scores from a face dataset, i.e., MS-Celeb-1M, based on feature-wise similarity. We experimentally verify that the new dataset created in such an omni-supervised manner can significantly improve the generalization ability of the learned FER model. However, as the number of training samples grows, computational cost and training time increase dramatically. To tackle this, we propose to apply a dataset distillation strategy to compress the created dataset into several informative class-wise images, significantly improving the training efficiency. We have conducted extensive experiments on widely used benchmarks, where consistent performance gains can be achieved under various settings using the proposed framework. More importantly, the distilled dataset has shown its capabilities of boosting the performance of FER with negligible additional computational costs.
Dataset distillation is a method for reducing dataset sizes by learning a small number of synthetic samples containing all the information of a large dataset. This has several benefits like speeding up model training, reducing energy consumption, and reducing required storage space. Currently, each synthetic sample is assigned a single `hard' label, and also, dataset distillation can currently only be used with image data. We propose to simultaneously distill both images and their labels, thus assigning each synthetic sample a `soft' label (a distribution of labels). Our algorithm increases accuracy by 2-4% over the original algorithm for several image classification tasks. Using `soft' labels also enables distilled datasets to consist of fewer samples than there are classes as each sample can encode information for multiple classes. For example, training a LeNet model with 10 distilled images (one per class) results in over 96% accuracy on MNIST, and almost 92% accuracy when trained on just 5 distilled images. We also extend the dataset distillation algorithm to distill sequential datasets including texts. We demonstrate that text distillation outperforms other methods across multiple datasets. For example, models attain almost their original accuracy on the IMDB sentiment analysis task using just 20 distilled sentences. Our code can be found at $\href{https://github.com/ilia10000/dataset-distillation}{\text{https://github.com/ilia10000/dataset-distillation}}$.
Model distillation aims to distill the knowledge of a complex model into a simpler one. In this paper, we consider an alternative formulation called dataset distillation: we keep the model fixed and instead attempt to distill the knowledge from a large training dataset into a small one. The idea is to synthesize a small number of data points that do not need to come from the correct data distribution, but will, when given to the learning algorithm as training data, approximate the model trained on the original data. For example, we show that it is possible to compress 60,000 MNIST training images into just 10 synthetic distilled images (one per class) and achieve close to original performance with only a few gradient descent steps, given a fixed network initialization. We evaluate our method in various initialization settings and with different learning objectives. Experiments on multiple datasets show the advantage of our approach compared to alternative methods.