Eeshan Gunesh Dhekane

Jason Ramapuram, Federico Danieli, Eeshan Dhekane et al. · apple-ml, berkeley

Attention is a key part of the transformer architecture. It is a sequence-to-sequence mapping that transforms each sequence element into a weighted sum of values. The weights are typically obtained as the softmax of dot products between keys and queries. Recent work has explored alternatives to softmax attention in transformers, such as ReLU and sigmoid activations. In this work, we revisit sigmoid attention and conduct an in-depth theoretical and empirical analysis. Theoretically, we prove that transformers with sigmoid attention are universal function approximators and benefit from improved regularity compared to softmax attention. Through detailed empirical analysis, we identify stabilization of large initial attention norms during the early stages of training as a crucial factor for the successful training of models with sigmoid attention, outperforming prior attempts. We also introduce FLASHSIGMOID, a hardware-aware and memory-efficient implementation of sigmoid attention yielding a 17% inference kernel speed-up over FLASHATTENTION2 on H100 GPUs. Experiments across language, vision, and speech show that properly normalized sigmoid attention matches the strong performance of softmax attention on a wide range of domains and scales, which previous attempts at sigmoid attention were unable to fully achieve. Our work unifies prior art and establishes best practices for sigmoid attention as a drop-in softmax replacement in transformers.

Dan Busbridge, Jason Ramapuram, Pierre Ablin et al. · apple-ml, berkeley

Preserving training dynamics across batch sizes is an important tool for practical machine learning as it enables the trade-off between batch size and wall-clock time. This trade-off is typically enabled by a scaling rule, for example, in stochastic gradient descent, one should scale the learning rate linearly with the batch size. Another important machine learning tool is the model EMA, a functional copy of a target model, whose parameters move towards those of its target model according to an Exponential Moving Average (EMA) at a rate parameterized by a momentum hyperparameter. This model EMA can improve the robustness and generalization of supervised learning, stabilize pseudo-labeling, and provide a learning signal for Self-Supervised Learning (SSL). Prior works have not considered the optimization of the model EMA when performing scaling, leading to different training dynamics across batch sizes and lower model performance. In this work, we provide a scaling rule for optimization in the presence of a model EMA and demonstrate the rule's validity across a range of architectures, optimizers, and data modalities. We also show the rule's validity where the model EMA contributes to the optimization of the target model, enabling us to train EMA-based pseudo-labeling and SSL methods at small and large batch sizes. For SSL, we enable training of BYOL up to batch size 24,576 without sacrificing performance, a 6$\times$ wall-clock time reduction under idealized hardware settings.

Andrius Ovsianas, Jason Ramapuram, Dan Busbridge et al. · apple-ml, berkeley

Self-supervised representation learning (SSL) methods provide an effective label-free initial condition for fine-tuning downstream tasks. However, in numerous realistic scenarios, the downstream task might be biased with respect to the target label distribution. This in turn moves the learned fine-tuned model posterior away from the initial (label) bias-free self-supervised model posterior. In this work, we re-interpret SSL fine-tuning under the lens of Bayesian continual learning and consider regularization through the Elastic Weight Consolidation (EWC) framework. We demonstrate that self-regularization against an initial SSL backbone improves worst sub-group performance in Waterbirds by 5% and Celeb-A by 2% when using the ViT-B/16 architecture. Furthermore, to help simplify the use of EWC with SSL, we pre-compute and publicly release the Fisher Information Matrix (FIM), evaluated with 10,000 ImageNet-1K variates evaluated on large modern SSL architectures including ViT-B/16 and ResNet50 trained with DINO.

Abhinav Moudgil, Ningyuan Huang, Eeshan Gunesh Dhekane et al. · apple-ml, mila

State Space Models (SSMs) such as Mamba have become a popular alternative to Transformer models, due to their reduced memory consumption and higher throughput at generation compared to their Attention-based counterparts. On the other hand, the community has built up a considerable body of knowledge on how to train Transformers, and many pretrained Transformer models are readily available. To facilitate the adoption of SSMs while leveraging existing pretrained Transformers, we aim to identify an effective recipe to distill an Attention-based model into a Mamba-like architecture. In prior work on cross-architecture distillation, however, it has been shown that a naïve distillation procedure from Transformers to Mamba fails to preserve the original teacher performance, a limitation often overcome with hybrid solutions combining Attention and SSM blocks. The key argument from our work is that, by equipping Mamba with a principled initialization, we can recover an overall better recipe for cross-architectural distillation. To this end, we propose a principled two-stage approach: first, we distill knowledge from a traditional Transformer into a linearized version of Attention, using an adaptation of the kernel trick. Then, we distill the linearized version into an adapted Mamba model that does not use any Attention block. Overall, the distilled Mamba model is able to preserve the original Pythia-1B Transformer performance in downstream tasks, maintaining a perplexity of 14.11 close to the teacher's 13.86. To show the efficacy of our recipe, we conduct thorough ablations at 1B scale with 10B tokens varying sequence mixer architecture, scaling analysis on model sizes and total distillation tokens, and a sensitivity analysis on tokens allocation between stages.

Jason Ramapuram, Eeshan Gunesh Dhekane, Amitis Shidani et al.

Speech-only spoken language models (SLMs) lag behind text and text-speech models in performance, with recent discrete autoregressive (AR) SLMs indicating significant computational and data demands to match text models. Since discretizing continuous speech for AR creates bottlenecks, we explore whether continuous diffusion (CD) SLM is more viable. To quantify the SLMs linguistic quality, we introduce the phoneme Jensen-Shannon divergence (pJSD) metric. Our analysis reveals CD SLMs, mirroring AR behavior, exhibit scaling laws for validation loss and pJSD, and show optimal token-to-parameter ratios decreasing as compute scales. However, for the latter, loss becomes insensitive to choice of data and model sizes, showing potential for fast inference. Scaling CD SLMs to 16B parameters with tens of millions of hours of conversational data enables generation of emotive, prosodic, multi-speaker, multilingual speech, though achieving long-form coherence remains a significant challenge.

Amitis Shidani, Devon Hjelm, Jason Ramapuram et al. · apple-ml, berkeley

Contrastive learning typically matches pairs of related views among a number of unrelated negative views. Views can be generated (e.g. by augmentations) or be observed. We investigate matching when there are more than two related views which we call poly-view tasks, and derive new representation learning objectives using information maximization and sufficient statistics. We show that with unlimited computation, one should maximize the number of related views, and with a fixed compute budget, it is beneficial to decrease the number of unique samples whilst increasing the number of views of those samples. In particular, poly-view contrastive models trained for 128 epochs with batch size 256 outperform SimCLR trained for 1024 epochs at batch size 4096 on ImageNet1k, challenging the belief that contrastive models require large batch sizes and many training epochs.

Melika Ayoughi, Samira Abnar, Chen Huang et al. · apple-ml

The composition of objects and their parts, along with object-object positional relationships, provides a rich source of information for representation learning. Hence, spatial-aware pretext tasks have been actively explored in self-supervised learning. Existing works commonly start from a grid structure, where the goal of the pretext task involves predicting the absolute position index of patches within a fixed grid. However, grid-based approaches fall short of capturing the fluid and continuous nature of real-world object compositions. We introduce PART, a self-supervised learning approach that leverages continuous relative transformations between off-grid patches to overcome these limitations. By modeling how parts relate to each other in a continuous space, PART learns the relative composition of images-an off-grid structural relative positioning process that generalizes beyond occlusions and deformations. In tasks requiring precise spatial understanding such as object detection and time series prediction, PART outperforms strong grid-based methods like MAE and DropPos, while also maintaining competitive performance on global classification tasks with minimal hyperparameter tuning. By breaking free from grid constraints, PART opens up an exciting new trajectory for universal self-supervised pretraining across diverse datatypes-from natural images to EEG signals-with promising potential in video, medical imaging, and audio.

Maureen de Seyssel, Eeshan Gunesh Dhekane · apple-ml

Speech foundation models have recently achieved remarkable capabilities across a wide range of tasks. However, their evaluation remains disjointed across tasks and model types. Different models excel at distinct aspects of speech processing and thus require different evaluation protocols. This paper proposes a unified taxonomy that addresses the question: Which evaluation is appropriate for which model? The taxonomy defines three orthogonal axes: the \textbf{evaluation aspect} being measured, the model capabilities required to attempt the task, and the task or protocol requirements needed to perform it. We classify a broad set of existing evaluations and benchmarks along these axes, spanning areas such as representation learning, speech generation, and interactive dialogue. By mapping each evaluation to the capabilities a model exposes (e.g., speech generation, real-time processing) and to its methodological demands (e.g., fine-tuning data, human judgment), the taxonomy provides a principled framework for aligning models with suitable evaluation methods. It also reveals systematic gaps, such as limited coverage of prosody, interaction, or reasoning, that highlight priorities for future benchmark design. Overall, this work offers a conceptual foundation and practical guide for selecting, interpreting, and extending evaluations of speech models.

Polina Turishcheva, Jason Ramapuram, Sinead Williamson et al. · apple-ml, berkeley

Understanding model uncertainty is important for many applications. We propose Bootstrap Your Own Variance (BYOV), combining Bootstrap Your Own Latent (BYOL), a negative-free Self-Supervised Learning (SSL) algorithm, with Bayes by Backprop (BBB), a Bayesian method for estimating model posteriors. We find that the learned predictive std of BYOV vs. a supervised BBB model is well captured by a Gaussian distribution, providing preliminary evidence that the learned parameter posterior is useful for label free uncertainty estimation. BYOV improves upon the deterministic BYOL baseline (+2.83% test ECE, +1.03% test Brier) and presents better calibration and reliability when tested with various augmentations (eg: +2.4% test ECE, +1.2% test Brier for Salt & Pepper noise).

Ankit Vani, Max Schwarzer, Yuchen Lu et al.

Although neural module networks have an architectural bias towards compositionality, they require gold standard layouts to generalize systematically in practice. When instead learning layouts and modules jointly, compositionality does not arise automatically and an explicit pressure is necessary for the emergence of layouts exhibiting the right structure. We propose to address this problem using iterated learning, a cognitive science theory of the emergence of compositional languages in nature that has primarily been applied to simple referential games in machine learning. Considering the layouts of module networks as samples from an emergent language, we use iterated learning to encourage the development of structure within this language. We show that the resulting layouts support systematic generalization in neural agents solving the more complex task of visual question-answering. Our regularized iterated learning method can outperform baselines without iterated learning on SHAPES-SyGeT (SHAPES Systematic Generalization Test), a new split of the SHAPES dataset we introduce to evaluate systematic generalization, and on CLOSURE, an extension of CLEVR also designed to test systematic generalization. We demonstrate superior performance in recovering ground-truth compositional program structure with limited supervision on both SHAPES-SyGeT and CLEVR.

Disha Shrivastava, Eeshan Gunesh Dhekane, Riashat Islam

Exploration and adaptation to new tasks in a transfer learning setup is a central challenge in reinforcement learning. In this work, we build on the idea of modeling a distribution over policies in a Bayesian deep reinforcement learning setup to propose a transfer strategy. Recent works have shown to induce diversity in the learned policies by maximizing the entropy of a distribution of policies (Bachman et al., 2018; Garnelo et al., 2018) and thus, we postulate that our proposed approach leads to faster exploration resulting in improved transfer learning. We support our hypothesis by demonstrating favorable experimental results on a variety of settings on fully-observable GridWorld and partially observable MiniGrid (Chevalier-Boisvert et al., 2018) environments.

Chin-Wei Huang, Kris Sankaran, Eeshan Dhekane et al.

Importance weighted variational inference (Burda et al., 2015) uses multiple i.i.d. samples to have a tighter variational lower bound. We believe a joint proposal has the potential of reducing the number of redundant samples, and introduce a hierarchical structure to induce correlation. The hope is that the proposals would coordinate to make up for the error made by one another to reduce the variance of the importance estimator. Theoretically, we analyze the condition under which convergence of the estimator variance can be connected to convergence of the lower bound. Empirically, we confirm that maximization of the lower bound does implicitly minimize variance. Further analysis shows that this is a result of negative correlation induced by the proposed hierarchical meta sampling scheme, and performance of inference also improves when the number of samples increases.

Gaurav Verma, Eeshan Gunesh Dhekane, Tanaya Guha

We introduce the problem of learning affective correspondence between audio (music) and visual data (images). For this task, a music clip and an image are considered similar (having true correspondence) if they have similar emotion content. In order to estimate this crossmodal, emotion-centric similarity, we propose a deep neural network architecture that learns to project the data from the two modalities to a common representation space, and performs a binary classification task of predicting the affective correspondence (true or false). To facilitate the current study, we construct a large scale database containing more than $3,500$ music clips and $85,000$ images with three emotion classes (positive, neutral, negative). The proposed approach achieves $61.67\%$ accuracy for the affective correspondence prediction task on this database, outperforming two relevant and competitive baselines. We also demonstrate that our network learns modality-specific representations of emotion (without explicitly being trained with emotion labels), which are useful for emotion recognition in individual modalities.