Nous utilisons des témoins pour analyser le trafic et l’utilisation de notre site web, afin de personnaliser votre expérience. Vous pouvez désactiver ces technologies à tout moment, mais cela peut restreindre certaines fonctionnalités du site. Consultez notre Politique de protection de la vie privée pour en savoir plus.
Paramètre des cookies
Vous pouvez activer et désactiver les types de cookies que vous souhaitez accepter. Cependant certains choix que vous ferez pourraient affecter les services proposés sur nos sites (ex : suggestions, annonces personnalisées, etc.).
Cookies essentiels
Ces cookies sont nécessaires au fonctionnement du site et ne peuvent être désactivés. (Toujours actif)
Cookies analyse
Acceptez-vous l'utilisation de cookies pour mesurer l'audience de nos sites ?
Multimedia Player
Acceptez-vous l'utilisation de cookies pour afficher et vous permettre de regarder les contenus vidéo hébergés par nos partenaires (YouTube, etc.) ?
Publications
Gradient descent induces alignment between weights and the pre-activation tangents for deep non-linear networks
Understanding the mechanisms through which neural networks extract statistics from input-label pairs is one of the most important unsolved p… (voir plus)roblems in supervised learning. Prior works have identified that the gram matrices of the weights in trained neural networks of general architectures are proportional to the average gradient outer product of the model, in a statement known as the Neural Feature Ansatz (NFA). However, the reason these quantities become correlated during training is poorly understood. In this work, we clarify the nature of this correlation and explain its emergence at early training times. We identify that the NFA is equivalent to alignment between the left singular structure of the weight matrices and the newly defined pre-activation tangent kernel. We identify a centering of the NFA that isolates this alignment and is robust to initialization scale. We show that, through this centering, the speed of NFA development can be predicted analytically in terms of simple statistics of the inputs and labels.
We seek to shed light on language model (LM) saturation from the perspective of learning dynamics.
To this end, we define a decomposition o… (voir plus)f the cross-entropy gradient, which forms a shared low-dimensional basis for analyzing the training dynamics of models across scales.
Intuitively, this decomposition consists of attractive and repulsive components that increase the logit of the correct class and decrease the logits of incorrect classes, respectively.
Our analysis in this subspace reveals a phenomenon we term \textit{gradient dissent}, characterized by gradient components becoming systematically opposed such that loss cannot be improved along one component without being degraded along the other.
Notably, we find that complete opposition, which we term \textit{total dissent}, reliably occurs in tandem with the saturation of smaller LMs.
Based on these results, we hypothesize that gradient dissent can provide a useful foundation for better understanding and mitigating saturation.
Across the primate cortex, neurons that perform similar functions tend to be spatially grouped together. In high-level visual cortex, this w… (voir plus)idely observed biological rule manifests itself as a modular organization of neuronal clusters, each tuned to a specific object category. The tendency toward short connections is one of the most widely accepted views of why such an organization exists in the brains of many animals. Yet, how such a feat is implemented at the neural level remains unclear.
Here, using artificial deep neural networks as test beds, we demonstrate that a topographical organization similar to that in the primary, intermediate, and high-level human visual cortex emerges when units in these models are laterally connected and their weight parameters are tuned by top-down credit assignment. Importantly, the emergence of the modular organization in the absence of explicit topography-inducing learning rules and objectives questions their necessity and suggests that local lateral connectivity alone may be sufficient for the formation of the topographic organization across the cortex.
The issue of bias in Machine Learning (ML) models is a significant challenge for the machine learning community. Real-world biases can be em… (voir plus)bedded in the data used to train models, and prior studies have shown that ML models can learn and even amplify these biases. This can result in unfair treatment of individuals based on their inherent characteristics or sensitive attributes such as gender, race, or age. Ensuring fairness is crucial with the increasing use of ML models in high-stakes scenarios and has gained significant attention from researchers in recent years. However, the challenge of ensuring fairness becomes much greater when the assumption of full access to sensitive attributes does not hold. The settings where the hypothesis does not hold include cases where (1) only limited or noisy demographic information is available or (2) demographic information is entirely unobserved due to privacy restrictions. This survey reviews recent research efforts to enforce fairness when sensitive attributes are missing. We propose a taxonomy of existing works and, more importantly, highlight current challenges and future research directions to stimulate research in ML fairness in the setting of missing sensitive attributes.
Neural network training begins with a chaotic phase in which the network is sensitive to small perturbations, such as those caused by stocha… (voir plus)stic gradient descent (SGD). This sensitivity can cause identically initialized networks to diverge both in parameter space and functional similarity.
However, the exact degree to which networks are sensitive to perturbation, and the sensitivity of networks as they transition out of the chaotic phase, is unclear.
To address this uncertainty, we apply a controlled perturbation at a single point in training time and measure its effect on otherwise identical training trajectories.
We find that both the
Discrete audio tokens have recently gained attention for their potential to bridge the gap between audio and language processing. Ideal audi… (voir plus)o tokens must preserve content, paralinguistic elements, speaker identity, and many other audio details. Current audio tokenization methods fall into two categories: Semantic tokens, acquired through quantization of Self-Supervised Learning (SSL) models, and Neural compression-based tokens (codecs). Although previous studies have benchmarked codec models to identify optimal configurations, the ideal setup for quantizing pretrained SSL models remains unclear. This paper explores the optimal configuration of semantic tokens across discriminative and generative tasks. We propose a scalable solution to train a universal vocoder across multiple SSL layers. Furthermore, an attention mechanism is employed to identify task-specific influential layers, enhancing the adaptability and performance of semantic tokens in diverse audio applications.
