We use cookies to analyze the browsing and usage of our website and to personalize your experience. You can disable these technologies at any time, but this may limit certain functionalities of the site. Read our Privacy Policy for more information.
Setting cookies
You can enable and disable the types of cookies you wish to accept. However certain choices you make could affect the services offered on our sites (e.g. suggestions, personalised ads, etc.).
Essential cookies
These cookies are necessary for the operation of the site and cannot be deactivated. (Still active)
Analytics cookies
Do you accept the use of cookies to measure the audience of our sites?
Multimedia Player
Do you accept the use of cookies to display and allow you to watch the video content hosted by our partners (YouTube, etc.)?
Brain function represents one of the most complex systems driving our world. Decoding its signals poses significant challenges, particularly… (see more) due to the limited availability of data and the high cost of recordings. The existence of large hospital datasets and laboratory collections partially mitigates this issue. However, the lack of standardized recording protocols, varying numbers of channels, diverse setups, scenarios, and recording devices further complicate the task. This work addresses these challenges by introducing the Brain Foundation Model (BFM), a suite of open-source models trained on brain signals. These models serve as foundational tools for various types of time-series neuroimaging tasks. This work presents the first model of the BFM series, which is trained on electroencephalogram signal data. Our results demonstrate that BFM-EEG can generate signals more accurately than other models. Upon acceptance, we will release the model weights and pipeline.
This work aims to understand how, in terms of training dynamics, scaling up language model size yields predictable loss improvements. We fin… (see more)d that these improvements can be tied back to loss deceleration, an abrupt transition in the rate of loss improvement, characterized by piece-wise linear behavior in log-log space. Notably, improvements from increased model size appear to be a result of (1) improving the loss at which this transition occurs; and (2) improving the rate of loss improvement after this transition. As an explanation for the mechanism underlying this transition (and the effect of model size on loss it mediates), we propose the zero-sum learning (ZSL) hypothesis. In ZSL, per-token gradients become systematically opposed, leading to degenerate training dynamics where the model can't improve loss on one token without harming it on another; bottlenecking the overall rate at which loss can improve. We find compelling evidence of ZSL, as well as unexpected results which shed light on other factors contributing to ZSL.
Test Time Adaptation (TTA) addresses the problem of distribution shift by adapting a pretrained model to a new domain during inference. When… (see more) faced with challenging shifts, most methods collapse and perform worse than the original pretrained model. In this paper, we find that not all layers are equally receptive to the adaptation, and the layers with the most misaligned gradients often cause performance degradation. To address this, we propose GALA, a novel layer selection criterion to identify the most beneficial updates to perform during test time adaptation. This criterion can also filter out unreliable samples with noisy gradients. Its simplicity allows seamless integration with existing TTA loss functions, thereby preventing degradation and focusing adaptation on the most trainable layers. This approach also helps to regularize adaptation to preserve the pretrained features, which are crucial for handling unseen domains. Through extensive experiments, we demonstrate that the proposed layer selection framework improves the performance of existing TTA approaches across multiple datasets, domain shifts, model architectures, and TTA losses.
Test Time Adaptation (TTA) addresses the problem of distribution shift by adapting a pretrained model to a new domain during inference. When… (see more) faced with challenging shifts, most methods collapse and perform worse than the original pretrained model. In this paper, we find that not all layers are equally receptive to the adaptation, and the layers with the most misaligned gradients often cause performance degradation. To address this, we propose GALA, a novel layer selection criterion to identify the most beneficial updates to perform during test time adaptation. This criterion can also filter out unreliable samples with noisy gradients. Its simplicity allows seamless integration with existing TTA loss functions, thereby preventing degradation and focusing adaptation on the most trainable layers. This approach also helps to regularize adaptation to preserve the pretrained features, which are crucial for handling unseen domains. Through extensive experiments, we demonstrate that the proposed layer selection framework improves the performance of existing TTA approaches across multiple datasets, domain shifts, model architectures, and TTA losses.
Learning transferable representations for deep reinforcement learning (RL) is a challenging problem due to the inherent non-stationarity, di… (see more)stribution shift, and unstable training dynamics. To be useful, a transferable representation needs to be robust to such factors. In this work, we introduce a new architecture and training strategy for learning robust representations for transfer learning in RL. We propose leveraging multiple CNN encoders and training them not to specialize in areas of the state space but instead to match each other's representation. We find that learned representations transfer well across many Atari tasks, resulting in better transfer learning performance and data efficiency than training from scratch.
Random feature models are a popular approach for studying network learning that can capture important behaviors while remaining simpler than… (see more) traditional training.
Guth et al. [2024] introduced “rainbow” networks which model the distribution of trained weights as correlated random features conditioned on previous layer activity.
Sampling new weights from distributions fit to learned networks led to similar performance in entirely untrained networks, and the observed weight covariance were found to be low rank.
This provided evidence that random feature models could be extended to some networks away from initialization, but White et al. [2024] failed to replicate their results in the deeper ResNet18 architecture.
Here we ask whether the rainbow formulation can succeed in deeper networks by directly training a stochastic ensemble of random features, which we call stochastic rainbow networks.
At every gradient descent iteration, new weights are sampled for all intermediate layers and features aligned layer-wise.
We find:
(1) this approach scales to deeper models, which outperform shallow networks at large widths;
(2) ensembling multiple samples from the stochastic model is better than retraining the classifier head; and
(3) low-rank parameterization of the learnable weight covariances can approach the accuracy of full-rank networks.
This offers more evidence for rainbow and other structured random feature networks as reduced models of deep learning.
Recent advancements in vision-language models (VLMs) have been driven by contrastive models like CLIP which learn to associate visual inform… (see more)ation with their corresponding text descriptions. However, these models have limitations in understanding complex compositional scenes involving multiple objects and their spatial relationships. To address these challenges, we propose a novel approach that diverges from traditional data-centric methods of enhancing model performance with hard negatives examples. Our work instead focuses on integrating sufficient inductive biases into pre-trained CLIP-like models to improve their compositional understanding without using additional data annotations. We introduce a binding module that connects a scene graph of the text with an induced graph-like representation of the image, facilitating a structured similarity assessment. We also leverage relationships as text-conditioned visual constraints, thereby capturing the intricate interactions between objects and their contextual relationships more effectively. Our resulting model (OC-CLIP) not only enhances the performance of CLIP in multi-object compositional understanding but also paves the way for more accurate and efficient image-text matching in complex scenes.
Recent trends of larger model and larger datasets require huge amounts of computational resources, making distributed deep learning essentia… (see more)l. Data parallelism is a common approach to speed up training, but it often involves frequent communication between workers, which can be a bottleneck. In this work, we propose a method called Pseudo-Asynchronous Local SGD (PALSGD) to improve the efficiency of data-parallel training. PALSGD is a novel extension of LocalSGD (SU Stich, 2018), designed to further reduce communication frequency by introducing a pseudo-synchronization mechanism. PALSGD allows the use of longer synchronization intervals compared to standard LocalSGD. Despite the reduced communication frequency, the pseudo-synchronization approach ensures that model consistency is maintained, leading to performance results comparable to those achieved with more frequent synchronization. Furthermore, we provide a theoretical analysis of PALSGD, establishing its convergence and deriving its convergence rate. This analysis offers insights into the algorithm's behavior and performance guarantees. We evaluated PALSGD on CIFAR-10 using a CNN and GPT-NEO on TinyStories. Our results show that PALSGD achieves better performance in less time compared to existing methods like distributed data parallel (DDP), Local SGD and DiLoCo (Douillard et al. 2023).
