Learned optimizers (LOs) can significantly reduce the wall-clock training time of neural networks, substantially reducing training costs. However, they often suffer from poor meta-generalization, especially when training networks larger than those seen during meta-training. To address this, we use the recently proposed Maximal Update Parametrization ($\mu$P), which allows zero-shot generalization of optimizer hyperparameters from smaller to larger models. We extend $\mu$P theory to learned optimizers, treating the meta-training problem as finding the learned optimizer under $\mu$P. Our evaluation shows that LOs meta-trained with $\mu$P substantially improve meta-generalization as compared to LOs trained under standard parametrization (SP). Notably, when applied to large-width models, our best $\mu$LO, trained for 103 GPU-hours, matches or exceeds the performance of VeLO, the largest publicly available learned optimizer, meta-trained with 4000 TPU-months of compute. Moreover, $\mu$LOs demonstrate better generalization than their SP counterparts to deeper networks and to much longer training horizons (25 times longer) than those seen during meta-training.

We analyze VeLO (versatile learned optimizer), the largest scale attempt to train a general purpose "foundational" optimizer to date. VeLO was trained on thousands of machine learning tasks using over 4000 TPU months with the goal of producing an optimizer capable of generalizing to new problems while being hyperparameter free, and outperforming industry standards such as Adam. We independently evaluate VeLO on the MLCommons optimizer benchmark suite. We find that, contrary to initial claims: (1) VeLO has a critical hyperparameter that needs problem-specific tuning, (2) VeLO does not necessarily outperform competitors in quality of solution found, and (3) VeLO is not faster than competing optimizers at reducing the training loss. These observations call into question VeLO's generality and the value of the investment in training it.

While deep learning models have replaced hand-designed features across many domains, these models are still trained with hand-designed optimizers. In this work, we leverage the same scaling approach behind the success of deep learning to learn versatile optimizers. We train an optimizer for deep learning which is itself a small neural network that ingests gradients and outputs parameter updates. Meta-trained with approximately four thousand TPU-months of compute on a wide variety of optimization tasks, our optimizer not only exhibits compelling performance, but optimizes in interesting and unexpected ways. It requires no hyperparameter tuning, instead automatically adapting to the specifics of the problem being optimized. We open source our learned optimizer, meta-training code, the associated train and test data, and an extensive optimizer benchmark suite with baselines at velo-code.github.io.

Due to the rapid growth of IoT and artificial intelligence, deploying neural networks on IoT devices is becoming increasingly crucial for edge intelligence. Federated learning (FL) facilitates the management of edge devices to collaboratively train a shared model while maintaining training data local and private. However, a general assumption in FL is that all edge devices are trained on the same machine learning model, which may be impractical considering diverse device capabilities. For instance, less capable devices may slow down the updating process because they struggle to handle large models appropriate for ordinary devices. In this paper, we propose a novel data-free FL method that supports heterogeneous client models by managing features and logits, called Felo; and its extension with a conditional VAE deployed in the server, called Velo. Felo averages the mid-level features and logits from the clients at the server based on their class labels to provide the average features and logits, which are utilized for further training the client models. Unlike Felo, the server has a conditional VAE in Velo, which is used for training mid-level features and generating synthetic features according to the labels. The clients optimize their models based on the synthetic features and the average logits. We conduct experiments on two datasets and show satisfactory performances of our methods compared with the state-of-the-art methods.