Search

Despite their remarkable performance, modern Diffusion Transformers are hindered by substantial resource requirements during inference, stemming from the fixed and large amount of compute needed for each denoising step. In this work, we revisit the conventional static paradigm that allocates a fixed compute budget per denoising iteration and propose a dynamic strategy instead. Our simple and sample-efficient framework enables pre-trained DiT models to be converted into \emph{flexible} ones -- dubbed FlexiDiT -- allowing them to process inputs at varying compute budgets. We demonstrate how a single \emph{flexible} model can generate images without any drop in quality, while reducing the required FLOPs by more than $40$\% compared to their static counterparts, for both class-conditioned and text-conditioned image generation. Our method is general and agnostic to input and conditioning modalities. We show how our approach can be readily extended for video generation, where FlexiDiT models generate samples with up to $75$\% less compute without compromising performance.

The performance of large language models (LLMs) is closely linked to their underlying size, leading to ever-growing networks and hence slower inference. Speculative decoding has been proposed as a technique to accelerate autoregressive generation, leveraging a fast draft model to propose candidate tokens, which are then verified in parallel based on their likelihood under the target model. While this approach guarantees to reproduce the target output, it incurs a substantial penalty: many high-quality draft tokens are rejected, even when they represent objectively valid continuations. Indeed, we show that even powerful draft models such as GPT-4o, as well as human text cannot achieve high acceptance rates under the standard verification scheme. This severely limits the speedup potential of current speculative decoding methods, as an early rejection becomes overwhelmingly likely when solely relying on alignment of draft and target. We thus ask the following question: Can we adapt verification to recognize correct, but non-aligned replies? To this end, we draw inspiration from the LLM-as-a-judge framework, which demonstrated that LLMs are able to rate answers in a versatile way. We carefully design a dataset to elicit the same capability in the target model by training a compact module on top of the embeddings to produce ``judgements" of the current continuation. We showcase our strategy on the Llama-3.1 family, where our 8b/405B-Judge achieves a speedup of 9x over Llama-405B, while maintaining its quality on a large range of benchmarks. These benefits remain present even in optimized inference frameworks, where our method reaches up to 141 tokens/s for 8B/70B-Judge and 129 tokens/s for 8B/405B on 2 and 8 H100s respectively.

The integration of Federated Learning (FL) and Multi-Task Learning (MTL) has been explored to address client heterogeneity, with Federated Multi-Task Learning (FMTL) treating each client as a distinct task. However, most existing research focuses on data heterogeneity (e.g., addressing non-IID data) rather than task heterogeneity, where clients solve fundamentally different tasks. Additionally, much of the work relies on centralized settings with a server managing the federation, leaving the more challenging domain of decentralized FMTL largely unexplored. Thus, this work bridges this gap by proposing ColNet, a framework designed for heterogeneous tasks in decentralized federated environments. ColNet divides models into the backbone and task-specific layers, forming groups of similar clients, with group leaders performing conflict-averse cross-group aggregation. A pool of experiments with different federations demonstrated ColNet outperforms the compared aggregation schemes in decentralized settings with label and task heterogeneity scenarios.

The `quantum gravity in the lab' paradigm suggests that quantum computers might shed light on quantum gravity by simulating the CFT side of the AdS/CFT correspondence and mapping the results to the AdS side. This relies on the assumption that the duality map (the `dictionary') is efficient to compute. In this work, we show that the complexity of the AdS/CFT dictionary is surprisingly subtle: there might be cases in which one can efficiently apply operators to the CFT state (a task we call 'operator reconstruction') without being able to extract basic properties of the dual bulk state such as its geometry (which we call 'geometry reconstruction'). Geometry reconstruction corresponds to the setting where we want to extract properties of a completely unknown bulk dual from a simulated CFT boundary state. We demonstrate that geometry reconstruction may be generically hard due to the connection between geometry and entanglement in holography. In particular we construct ensembles of states whose entanglement approximately obey the Ryu-Takayanagi formula for arbitrary geometries, but which are nevertheless computationally indistinguishable. This suggests that even for states with the special entanglement structure of holographic CFT states, geometry reconstruction might be hard. This result should be compared with existing evidence that operator reconstruction is generically easy in AdS/CFT. A useful analogy for the difference between these two tasks is quantum fully homomorphic encryption (FHE): this encrypts quantum states in such a way that no efficient adversary can learn properties of the state, but operators can be applied efficiently to the encrypted state. We show that quantum FHE can separate the complexity of geometry reconstruction vs operator reconstruction, which raises the question whether FHE could be a useful lens through which to view AdS/CFT., Comment: 45 pages, 9 figures

We present Movie Gen, a cast of foundation models that generates high-quality, 1080p HD videos with different aspect ratios and synchronized audio. We also show additional capabilities such as precise instruction-based video editing and generation of personalized videos based on a user's image. Our models set a new state-of-the-art on multiple tasks: text-to-video synthesis, video personalization, video editing, video-to-audio generation, and text-to-audio generation. Our largest video generation model is a 30B parameter transformer trained with a maximum context length of 73K video tokens, corresponding to a generated video of 16 seconds at 16 frames-per-second. We show multiple technical innovations and simplifications on the architecture, latent spaces, training objectives and recipes, data curation, evaluation protocols, parallelization techniques, and inference optimizations that allow us to reap the benefits of scaling pre-training data, model size, and training compute for training large scale media generation models. We hope this paper helps the research community to accelerate progress and innovation in media generation models. All videos from this paper are available at https://go.fb.me/MovieGenResearchVideos.

Self-supervised learning on tabular data seeks to apply advances from natural language and image domains to the diverse domain of tables. However, current techniques often struggle with integrating multi-domain data and require data cleaning or specific structural requirements, limiting the scalability of pre-training datasets. We introduce PORTAL (Pretraining One-Row-at-a-Time for All tabLes), a framework that handles various data modalities without the need for cleaning or preprocessing. This simple yet powerful approach can be effectively pre-trained on online-collected datasets and fine-tuned to match state-of-the-art methods on complex classification and regression tasks. This work offers a practical advancement in self-supervised learning for large-scale tabular data., Comment: Accepted at Table Representation Learning Workshop at NeurIPS 2024

Dermal interstitial fluid (ISF) is a rich source of biomarkers (e.g., glucose) that can be used for continuous health monitoring with wearable sensors. Hollow microneedle devices are a promising solution to extract ISF on demand by penetrating the skin with minimal pain. However, they rely on inserting bio-incompatible materials (e.g., silicon) into individuals, limiting the application time. Here, the direct 3D printing of polymer hollow microneedles on silicon-based microfluidic devices and the successful in-vivo extraction of ISF are demonstrated. Our additive manufacturing approach enables the versatile combination of materials and rapid prototyping of microneedle geometry. After improving the design through finite element modeling, a hollow microneedle geometry was printed by two-photon polymerization and experimentally characterized with mechanical and fluidic tests. Microneedles were fabricated with high accuracy (i.e., 997 +/- 2 um) and reliably interfaced with the microfluidic chip (i.e., centerline alignment within 5% of diameter). The needles demonstrated sufficient mechanical strength (i.e., 411 +/- 3 mN per needle) to endure at least 10 consecutive insertions into simulated skin. Biocompatibility and ISF extraction were demonstrated in an in-vivo 72-hour test, showing the safety and reliability of our approach. Such a platform is promising for minimally invasive, continuous monitoring of biomarkers in ISF, aiding in medical diagnoses and personalized health treatments.

The advent of increasingly-growing virtual realities poses unprecedented opportunities and challenges to different societies. Artistic collectives are not an exception, and we here aim to put special attention into musicians. Compositions, lyrics and even show-advertisements are constituents of a message that artists transmit about their reality. As such, artistic creations are ultimately linked to feelings and emotions, with aesthetics playing a crucial role when it comes to transmit artist's intentions. In this context, we here analyze how virtual realities can help to broaden the opportunities for musicians to bridge with their audiences, by devising a dynamical fashion-design recommendation system inspired by sound stimulus. We present our first steps towards re-defining musical experiences in the metaverse, opening up alternative opportunities for artists to connect both with real and virtual (\textit{e.g.} machine-learning agents operating in the metaverse) in potentially broader ways.

Social networks shape individuals' lives, influencing everything from career paths to health. This paper presents a registry-based, multi-layer and temporal network of the entire Danish population in the years 2008-2021 (roughly 7.2 mill. individuals). Our network maps the relationships formed through family, households, neighborhoods, colleagues and classmates. We outline key properties of this multiplex network, introducing both an individual-focused perspective as well as a bipartite representation. We show how to aggregate and combine the layers, and how to efficiently compute network measures such as shortest paths in large administrative networks. Our analysis reveals how past connections reappear later in other layers, that the number of relationships aggregated over time reflects the position in the income distribution, and that we can recover canonical shortest path length distributions when appropriately weighting connections. Along with the network data, we release a Python package that uses the bipartite network representation for efficient analysis.

Recurrent neural networks (RNNs) are a widely used tool for sequential data analysis, however, they are still often seen as black boxes of computation. Understanding the functional principles of these networks is critical to developing ideal model architectures and optimization strategies. Previous studies typically only emphasize the network representation post-training, overlooking their evolution process throughout training. Here, we present Multiway Multislice PHATE (MM-PHATE), a novel method for visualizing the evolution of RNNs' hidden states. MM-PHATE is a graph-based embedding using structured kernels across the multiple dimensions spanned by RNNs: time, training epoch, and units. We demonstrate on various datasets that MM-PHATE uniquely preserves hidden representation community structure among units and identifies information processing and compression phases during training. The embedding allows users to look under the hood of RNNs across training and provides an intuitive and comprehensive strategy to understanding the network's internal dynamics and draw conclusions, e.g., on why and how one model outperforms another or how a specific architecture might impact an RNN's learning ability.

The advent of Artificial Intelligence is expected to imply profound changes in the short-term. It is therefore imperative for Academia, and particularly for the Computer Science scope, to develop cross-disciplinary tools that bond AI developments to their social dimension. To this aim, we introduce the Challenge-Device-Synthesis methodology (CDS), in which a specific challenge is presented to the students of AI, who are required to develop a device as a solution for the challenge. The device becomes the object of study for the different dimensions of social transformation, and the conclusions addressed by the students during the discussion around the device are presented in a synthesis piece in the shape of a 10-page scientific paper. The latter is evaluated taking into account both the depth of analysis and the level to which it genuinely reflects the social transformations associated with the proposed AI-based device. We provide data obtained during the pilot for the implementation phase of CDS within the subject of Social Innovation, a 6-ECTS subject from the 6th semester of the Degree of Artificial Intelligence, UAB-Barcelona. We provide details on temporalisation, task distribution, methodological tools used and assessment delivery procedure, as well as qualitative analysis of the results obtained., Comment: accepted as contribution for EDULEARN24 - 16th annual International Conference on Education and New Learning Technologies

Deep reinforcement learning agents are prone to goal misalignments. The black-box nature of their policies hinders the detection and correction of such misalignments, and the trust necessary for real-world deployment. So far, solutions learning interpretable policies are inefficient or require many human priors. We propose INTERPRETER, a fast distillation method producing INTerpretable Editable tRee Programs for ReinforcEmenT lEaRning. We empirically demonstrate that INTERPRETER compact tree programs match oracles across a diverse set of sequential decision tasks and evaluate the impact of our design choices on interpretability and performances. We show that our policies can be interpreted and edited to correct misalignments on Atari games and to explain real farming strategies.

We investigate the properties of Floquet states in the vicinity of a conical intersection of quasienergies and work out the consequences of the underlying spatio-temporal symmetries for a driven two-level system coupled to an ohmic heat bath. We find that on manifolds with constant quasienergy splitting, the mean energies of the Floquet states are continuously interchanged. In the presence of dissipation, the parameter dependence of the stationary populations generally resembles that of the mean energies. In turn, the mean energies are an indicator for the qualitative behavior of the density operator in the long-time limit. A further consequence of the symmetries is that for specific driving parameters, the stationary state may be fully mixed even at arbitrarily low temperatures. For large driving frequencies, such states with maximal entropy are found in the whole vicinity of the intersection, which can be explained by a chirality emerging in this limit. Analytical results beyond a high-frequency approximation are illustrated by numerical data., Comment: 7 pages, 6 figures

Image classification based on over-parametrized convolutional neural networks with a global average-pooling layer is considered. The weights of the network are learned by gradient descent. A bound on the rate of convergence of the difference between the misclassification risk of the newly introduced convolutional neural network estimate and the minimal possible value is derived.

Diffusion models are a powerful generative framework, but come with expensive inference. Existing acceleration methods often compromise image quality or fail under complex conditioning when operating in an extremely low-step regime. In this work, we propose a novel distillation framework tailored to enable high-fidelity, diverse sample generation using just one to three steps. Our approach comprises three key components: (i) Backward Distillation, which mitigates training-inference discrepancies by calibrating the student on its own backward trajectory; (ii) Shifted Reconstruction Loss that dynamically adapts knowledge transfer based on the current time step; and (iii) Noise Correction, an inference-time technique that enhances sample quality by addressing singularities in noise prediction. Through extensive experiments, we demonstrate that our method outperforms existing competitors in quantitative metrics and human evaluations. Remarkably, it achieves performance comparable to the teacher model using only three denoising steps, enabling efficient high-quality generation.

Underwater Unmanned Vehicles (UUVs) have to constantly compensate for the external disturbing forces acting on their body. Adaptive Control theory is commonly used there to grant the control law some flexibility in its response to process variation. Today, learning-based (LB) adaptive methods are leading the field where model-based control structures are combined with deep model-free learning algorithms. This work proposes experiments and metrics to empirically study the stability of such a controller. We perform this stability analysis on a LB adaptive control system whose adaptive parameters are determined using a Cross-Entropy Deep Learning method.

Embracing the pursuit of intrinsically explainable reinforcement learning raises crucial questions: what distinguishes explainability from interpretability? Should explainable and interpretable agents be developed outside of domains where transparency is imperative? What advantages do interpretable policies offer over neural networks? How can we rigorously define and measure interpretability in policies, without user studies? What reinforcement learning paradigms,are the most suited to develop interpretable agents? Can Markov Decision Processes integrate interpretable state representations? In addition to motivate an Interpretable RL community centered around the aforementioned questions, we propose the first venue dedicated to Interpretable RL: the InterpPol Workshop.

Image classification from independent and identically distributed random variables is considered. Image classifiers are defined which are based on a linear combination of deep convolutional networks with max-pooling layer. Here all the weights are learned by stochastic gradient descent. A general result is presented which shows that the image classifiers are able to approximate the best possible deep convolutional network. In case that the a posteriori probability satisfies a suitable hierarchical composition model it is shown that the corresponding deep convolutional neural network image classifier achieves a rate of convergence which is independent of the dimension of the images., Comment: arXiv admin note: text overlap with arXiv:2312.17007

We investigate the link between quantum position-verification (QPV) and holography established in [MPS19] using holographic quantum error correcting codes as toy models. By inserting the "temporal" scaling of the AdS metric by hand via the bulk Hamiltonian interaction strength, we recover a toy model with consistent causality structure. This leads to an interesting implication between two topics in quantum information: if position-based verification is secure against attacks with small entanglement then there are new fundamental lower bounds for resources required for one Hamiltonian to simulate another., Comment: (v2: updated presentation)

One of the most recent and fascinating breakthroughs in artificial intelligence is ChatGPT, a chatbot which can simulate human conversation. ChatGPT is an instance of GPT4, which is a language model based on generative gredictive gransformers. So if one wants to study from a theoretical point of view, how powerful such artificial intelligence can be, one approach is to consider transformer networks and to study which problems one can solve with these networks theoretically. Here it is not only important what kind of models these network can approximate, or how they can generalize their knowledge learned by choosing the best possible approximation to a concrete data set, but also how well optimization of such transformer network based on concrete data set works. In this article we consider all these three different aspects simultaneously and show a theoretical upper bound on the missclassification probability of a transformer network fitted to the observed data. For simplicity we focus in this context on transformer encoder networks which can be applied to define an estimate in the context of a classification problem involving natural language.

In this study, we propose two novel input processing paradigms for novel view synthesis (NVS) methods based on layered scene representations that significantly improve their runtime without compromising quality. Our approach identifies and mitigates the two most time-consuming aspects of traditional pipelines: building and processing the so-called plane sweep volume (PSV), which is a high-dimensional tensor of planar re-projections of the input camera views. In particular, we propose processing this tensor in parallel groups for improved compute efficiency as well as super-sampling adjacent input planes to generate denser, and hence more accurate scene representation. The proposed enhancements offer significant flexibility, allowing for a balance between performance and speed, thus making substantial steps toward real-time applications. Furthermore, they are very general in the sense that any PSV-based method can make use of them, including methods that employ multiplane images, multisphere images, and layered depth images. In a comprehensive set of experiments, we demonstrate that our proposed paradigms enable the design of an NVS method that achieves state-of-the-art on public benchmarks while being up to $50x$ faster than existing state-of-the-art methods. It also beats the current forerunner in terms of speed by over $3x$, while achieving significantly better rendering quality.

This paper presents a comprehensive study on the role of Classifier-Free Guidance (CFG) in text-conditioned diffusion models from the perspective of inference efficiency. In particular, we relax the default choice of applying CFG in all diffusion steps and instead search for efficient guidance policies. We formulate the discovery of such policies in the differentiable Neural Architecture Search framework. Our findings suggest that the denoising steps proposed by CFG become increasingly aligned with simple conditional steps, which renders the extra neural network evaluation of CFG redundant, especially in the second half of the denoising process. Building upon this insight, we propose "Adaptive Guidance" (AG), an efficient variant of CFG, that adaptively omits network evaluations when the denoising process displays convergence. Our experiments demonstrate that AG preserves CFG's image quality while reducing computation by 25%. Thus, AG constitutes a plug-and-play alternative to Guidance Distillation, achieving 50% of the speed-ups of the latter while being training-free and retaining the capacity to handle negative prompts. Finally, we uncover further redundancies of CFG in the first half of the diffusion process, showing that entire neural function evaluations can be replaced by simple affine transformations of past score estimates. This method, termed LinearAG, offers even cheaper inference at the cost of deviating from the baseline model. Our findings provide insights into the efficiency of the conditional denoising process that contribute to more practical and swift deployment of text-conditioned diffusion models.

Diffusion models have recently revolutionized the field of image synthesis due to their ability to generate photorealistic images. However, one of the major drawbacks of diffusion models is that the image generation process is costly. A large image-to-image network has to be applied many times to iteratively refine an image from random noise. While many recent works propose techniques to reduce the number of required steps, they generally treat the underlying denoising network as a black box. In this work, we investigate the behavior of the layers within the network and find that 1) the layers' output changes smoothly over time, 2) the layers show distinct patterns of change, and 3) the change from step to step is often very small. We hypothesize that many layer computations in the denoising network are redundant. Leveraging this, we introduce block caching, in which we reuse outputs from layer blocks of previous steps to speed up inference. Furthermore, we propose a technique to automatically determine caching schedules based on each block's changes over timesteps. In our experiments, we show through FID, human evaluation and qualitative analysis that Block Caching allows to generate images with higher visual quality at the same computational cost. We demonstrate this for different state-of-the-art models (LDM and EMU) and solvers (DDIM and DPM)., Comment: Project page: https://fwmb.github.io/blockcaching/

We study the complexity of a classic problem in computational topology, the homology problem: given a description of some space $X$ and an integer $k$, decide if $X$ contains a $k$-dimensional hole. The setting and statement of the homology problem are completely classical, yet we find that the complexity is characterized by quantum complexity classes. Our result can be seen as an aspect of a connection between homology and supersymmetric quantum mechanics. We consider clique complexes, motivated by the practical application of topological data analysis (TDA). The clique complex of a graph is the simplicial complex formed by declaring every $k+1$-clique in the graph to be a $k$-simplex. Our main result is that deciding whether the clique complex of a weighted graph has a hole or not, given a suitable promise on the gap, is $\text{QMA}_1$-hard and contained in $\text{QMA}$. Our main innovation is a technique to lower bound the eigenvalues of the combinatorial Laplacian operator. For this, we invoke a tool from algebraic topology known as \emph{spectral sequences}. In particular, we exploit a connection between spectral sequences and Hodge theory. Spectral sequences will play a role analogous to perturbation theory for combinatorial Laplacians. In addition, we develop the simplicial surgery technique used in prior work. Our result provides some suggestion that the quantum TDA algorithm \emph{cannot} be dequantized. More broadly, we hope that our results will open up new possibilities for quantum advantage in topological data analysis.

Recent results show that estimates defined by over-parametrized deep neural networks learned by applying gradient descent to a regularized empirical $L_2$ risk are universally consistent and achieve good rates of convergence. In this paper, we show that the regularization term is not necessary to obtain similar results. In the case of a suitably chosen initialization of the network, a suitable number of gradient descent steps, and a suitable step size we show that an estimate without a regularization term is universally consistent for bounded predictor variables. Additionally, we show that if the regression function is H\"older smooth with H\"older exponent $1/2 \leq p \leq 1$, the $L_2$ error converges to zero with a convergence rate of approximately $n^{-1/(1+d)}$. Furthermore, in case of an interaction model, where the regression function consists of a sum of H\"older smooth functions with $d^*$ components, a rate of convergence is derived which does not depend on the input dimension $d$.

Spin-flips are one of the limiting factors for spin-based information processing. We demonstrate a transport approach for determining the spin-flip rates of a self-assembled InAs double quantum dot occupied by a single electron. In such devices, different Land\'e factors lead to an inhomogeneous Zeeman splitting, so that the two spin channels can never be at resonance simultaneously, leading to a spin blockade at low temperatures. This blockade is analyzed in terms of spin flips for different temperatures and magnetic fields. Our results are in good agreement with a quantum master equation that combines the dot-lead couplings with ohmic dissipation stemming from spin-flip cotunneling., Comment: 9 pages, 8 figures