Search

As machine intelligence evolves, the need to test and compare the problem-solving abilities of different AI models grows. However, current benchmarks are often simplistic, allowing models to perform uniformly well and making it difficult to distinguish their capabilities. Additionally, benchmarks typically rely on static question-answer pairs that the models might memorize or guess. To address these limitations, we introduce Dynamic Intelligence Assessment (DIA), a novel methodology for testing AI models using dynamic question templates and improved metrics across multiple disciplines such as mathematics, cryptography, cybersecurity, and computer science. The accompanying dataset, DIA-Bench, contains a diverse collection of challenge templates with mutable parameters presented in various formats, including text, PDFs, compiled binaries, visual puzzles, and CTF-style cybersecurity challenges. Our framework introduces four new metrics to assess a model's reliability and confidence across multiple attempts. These metrics revealed that even simple questions are frequently answered incorrectly when posed in varying forms, highlighting significant gaps in models' reliability. Notably, API models like GPT-4o often overestimated their mathematical capabilities, while ChatGPT-4o demonstrated better performance due to effective tool usage. In self-assessment, OpenAI's o1-mini proved to have the best judgement on what tasks it should attempt to solve. We evaluated 25 state-of-the-art LLMs using DIA-Bench, showing that current models struggle with complex tasks and often display unexpectedly low confidence, even with simpler questions. The DIA framework sets a new standard for assessing not only problem-solving but also a model's adaptive intelligence and ability to assess its limitations. The dataset is publicly available on the project's page: https://github.com/DIA-Bench.

A vortices-bubble system may be depicted by some features which may lead to infer its acoustic charge and its spin angular momentum. We study the dynamics of this system within the framework of fluid physics and with the help of Maxwell's hydrodynamic equations. Since we will adopt an approach of a fluid without viscosity and without gravity, then the fluid is in steady state if also we will neglect the acoustic radiation. When we expressed the energy (the kinetic and potential energy) of the vortex-bubble system we found out that the state of bubble which is captured in the core of the vortex is more stable than it is out of the core. For the above allegations we consider that this system, namely a vortex-bubble system, is a good approach for an acoustic particle with charge and spin angular momentum. This system owns to the Acoustic World., Comment: 15 pages

In this paper we study the properties of vortexes, as systems specific to the Acoustic World, using both hydrodynamic theory and the corresponding hydrodynamic Maxwell equations. According to this study, it follows that the vortex behaves like an acoustic dipole that has intrinsic/internal angular momentum. The system of two identical vortices also has orbital angular momentum and behaves, at distances much greater than the distance between the axes of the vortices, as a single vortex. With the help of Maxwell's hydrodynamic equations for the vortex we deduced the force between two vortices and obtained the expression of the equivalent mass of the vortex and the permittivity of the electroacoustic field. We also obtained and interpreted the expression for the energy density of the acoustic field. The density and pressure variations induced by the vortex cause the change in the propagation speed of the acoustic waves and the acoustic lensing property of the vortex., Comment: 12 pages

In this report, we introduce the Gemini 1.5 family of models, representing the next generation of highly compute-efficient multimodal models capable of recalling and reasoning over fine-grained information from millions of tokens of context, including multiple long documents and hours of video and audio. The family includes two new models: (1) an updated Gemini 1.5 Pro, which exceeds the February version on the great majority of capabilities and benchmarks; (2) Gemini 1.5 Flash, a more lightweight variant designed for efficiency with minimal regression in quality. Gemini 1.5 models achieve near-perfect recall on long-context retrieval tasks across modalities, improve the state-of-the-art in long-document QA, long-video QA and long-context ASR, and match or surpass Gemini 1.0 Ultra's state-of-the-art performance across a broad set of benchmarks. Studying the limits of Gemini 1.5's long-context ability, we find continued improvement in next-token prediction and near-perfect retrieval (>99%) up to at least 10M tokens, a generational leap over existing models such as Claude 3.0 (200k) and GPT-4 Turbo (128k). Finally, we highlight real-world use cases, such as Gemini 1.5 collaborating with professionals on completing their tasks achieving 26 to 75% time savings across 10 different job categories, as well as surprising new capabilities of large language models at the frontier; when given a grammar manual for Kalamang, a language with fewer than 200 speakers worldwide, the model learns to translate English to Kalamang at a similar level to a person who learned from the same content.

We propose MusicRL, the first music generation system finetuned from human feedback. Appreciation of text-to-music models is particularly subjective since the concept of musicality as well as the specific intention behind a caption are user-dependent (e.g. a caption such as "upbeat work-out music" can map to a retro guitar solo or a techno pop beat). Not only this makes supervised training of such models challenging, but it also calls for integrating continuous human feedback in their post-deployment finetuning. MusicRL is a pretrained autoregressive MusicLM (Agostinelli et al., 2023) model of discrete audio tokens finetuned with reinforcement learning to maximise sequence-level rewards. We design reward functions related specifically to text-adherence and audio quality with the help from selected raters, and use those to finetune MusicLM into MusicRL-R. We deploy MusicLM to users and collect a substantial dataset comprising 300,000 pairwise preferences. Using Reinforcement Learning from Human Feedback (RLHF), we train MusicRL-U, the first text-to-music model that incorporates human feedback at scale. Human evaluations show that both MusicRL-R and MusicRL-U are preferred to the baseline. Ultimately, MusicRL-RU combines the two approaches and results in the best model according to human raters. Ablation studies shed light on the musical attributes influencing human preferences, indicating that text adherence and quality only account for a part of it. This underscores the prevalence of subjectivity in musical appreciation and calls for further involvement of human listeners in the finetuning of music generation models.

The outcome of this paper was a shape of the interaction of two oscillating bubbles. This was done to express the secondary Bjerknes force using the Maxwell equations for a liquid. These subsequent equations were written for the quantities velocity, pressure, and density as deviations from steady state. Also, as it will show in the following rows, we found the expressions of the acoustic charge and of the acoustic intensity. Our results may facilitate the adoption of an approach for the interaction of two vortices and for the entrapment of a bubble by a vortex. This approach leads to a model for an acoustic charged particle which has internal angular momentum., Comment: 12, Fixed typos. Revised argument in section 3, new results

We present TokenSplit, a speech separation model that acts on discrete token sequences. The model is trained on multiple tasks simultaneously: separate and transcribe each speech source, and generate speech from text. The model operates on transcripts and audio token sequences and achieves multiple tasks through masking of inputs. The model is a sequence-to-sequence encoder-decoder model that uses the Transformer architecture. We also present a "refinement" version of the model that predicts enhanced audio tokens from the audio tokens of speech separated by a conventional separation model. Using both objective metrics and subjective MUSHRA listening tests, we show that our model achieves excellent performance in terms of separation, both with or without transcript conditioning. We also measure the automatic speech recognition (ASR) performance and provide audio samples of speech synthesis to demonstrate the additional utility of our model., Comment: INTERSPEECH 2023, project webpage with audio demos at https://google-research.github.io/sound-separation/papers/tokensplit

Numerically calculating the interaction forces between two free bubbles under the action of a background of random acoustic radiation, we highlight the contributions of radiative coefficient and absorption damping coefficient to the size of these forces.It is quantitatively demonstrated, for different radii of the oscillating bubbles, that the scattering absorption forces and the scattering scattering forces are close in magnitude.For superfluid helium, the forces change direction, oscillatingly, and the ratio of the forces is much less than one., Comment: 15 pages, 12 tables, 11 graphs

We introduce AudioPaLM, a large language model for speech understanding and generation. AudioPaLM fuses text-based and speech-based language models, PaLM-2 [Anil et al., 2023] and AudioLM [Borsos et al., 2022], into a unified multimodal architecture that can process and generate text and speech with applications including speech recognition and speech-to-speech translation. AudioPaLM inherits the capability to preserve paralinguistic information such as speaker identity and intonation from AudioLM and the linguistic knowledge present only in text large language models such as PaLM-2. We demonstrate that initializing AudioPaLM with the weights of a text-only large language model improves speech processing, successfully leveraging the larger quantity of text training data used in pretraining to assist with the speech tasks. The resulting model significantly outperforms existing systems for speech translation tasks and has the ability to perform zero-shot speech-to-text translation for many languages for which input/target language combinations were not seen in training. AudioPaLM also demonstrates features of audio language models, such as transferring a voice across languages based on a short spoken prompt. We release examples of our method at https://google-research.github.io/seanet/audiopalm/examples, Comment: Technical report

Using the computational resources of an untrusted third party to crack a password hash can pose a high number of privacy and security risks. The act of revealing the hash digest could in itself negatively impact both the data subject who created the password, and the data controller who stores the hash digest. This paper solves this currently open problem by presenting a Privacy-Preserving Password Cracking protocol (3PC), that prevents the third party cracking server from learning any useful information about the hash digest, or the recovered cleartext. This is achieved by a tailored anonymity set of decoy hashes, based on the concept of predicate encryption, where we extend the definition of a predicate function, to evaluate the output of a one way hash function. The protocol allows the client to maintain plausible deniability where the real choice of hash digest cannot be proved, even by the client itself. The probabilistic information the server obtains during the cracking process can be calculated and minimized to a desired level. While in theory cracking a larger set of hashes would decrease computational speed, the 3PC protocol provides constant-time lookup on an arbitrary list size, bounded by the input/output operation per second (IOPS) capabilities of the third party server, thereby allowing the protocol to scale efficiently. We demonstrate these claims both theoretically and in practice, with a real-life use case implemented on an FPGA architecture.

We present SoundStorm, a model for efficient, non-autoregressive audio generation. SoundStorm receives as input the semantic tokens of AudioLM, and relies on bidirectional attention and confidence-based parallel decoding to generate the tokens of a neural audio codec. Compared to the autoregressive generation approach of AudioLM, our model produces audio of the same quality and with higher consistency in voice and acoustic conditions, while being two orders of magnitude faster. SoundStorm generates 30 seconds of audio in 0.5 seconds on a TPU-v4. We demonstrate the ability of our model to scale audio generation to longer sequences by synthesizing high-quality, natural dialogue segments, given a transcript annotated with speaker turns and a short prompt with the speakers' voices.

Realistic credit risk assessment, the estimation of losses from counterparty's failure, is central for the financial stability. Credit risk models focus on the financial conditions of borrowers and only marginally consider other risks from the real economy, supply chains in particular. Recent pandemics, geopolitical instabilities, and natural disasters demonstrated that supply chain shocks do contribute to large financial losses. Based on a unique nation-wide micro-dataset, containing practically all supply chain relations of all Hungarian firms, together with their bank loans, we estimate how firm-failures affect the supply chain network, leading to potentially additional firm defaults and additional financial losses. Within a multi-layer network framework we define a financial systemic risk index (FSRI) for every firm, quantifying these expected financial losses caused by its own- and all the secondary defaulting loans caused by supply chain network (SCN) shock propagation. We find a small fraction of firms carrying substantial financial systemic risk, affecting up to 16% of the banking system's overall equity. These losses are predominantly caused by SCN contagion. For every bank we calculate the expected loss (EL), value at risk (VaR) and expected shortfall (ES), with and without accounting for SCN contagion. We find that SCN contagion amplifies the EL, VaR, and ES by a factor of 4.3, 4.5, and 3.2, respectively. These findings indicate that for a more complete picture of financial stability and realistic credit risk assessment, SCN contagion needs to be considered. This newly quantified contagion channel is of potential relevance for regulators' future systemic risk assessments.

We introduce LMCodec, a causal neural speech codec that provides high quality audio at very low bitrates. The backbone of the system is a causal convolutional codec that encodes audio into a hierarchy of coarse-to-fine tokens using residual vector quantization. LMCodec trains a Transformer language model to predict the fine tokens from the coarse ones in a generative fashion, allowing for the transmission of fewer codes. A second Transformer predicts the uncertainty of the next codes given the past transmitted codes, and is used to perform conditional entropy coding. A MUSHRA subjective test was conducted and shows that the quality is comparable to reference codecs at higher bitrates. Example audio is available at https://mjenrungrot.github.io/chrome-media-audio-papers/publications/lmcodec., Comment: 5 pages, accepted to ICASSP 2023, project page: https://mjenrungrot.github.io/chrome-media-audio-papers/publications/lmcodec

To estimate the reaction of economies to political interventions or external disturbances, input-output (IO) tables -- constructed by aggregating data into industrial sectors -- are extensively used. However, economic growth, robustness, and resilience crucially depend on the detailed structure of non-aggregated firm-level production networks (FPNs). Due to non-availability of data little is known about how much aggregated sector-based and detailed firm-level-based model-predictions differ. Using a nearly complete nationwide FPN, containing 243,399 Hungarian firms with 1,104,141 supplier-buyer-relations we self-consistently compare production losses on the aggregated industry-level production network (IPN) and the granular FPN. For this we model the propagation of shocks of the same size on both, the IPN and FPN, where the latter captures relevant heterogeneities within industries. In a COVID-19 inspired scenario we model the shock based on detailed firm-level data during the early pandemic. We find that using IPNs instead of FPNs leads to errors up to 37% in the estimation of economic losses, demonstrating a natural limitation of industry-level IO-models in predicting economic outcomes. We ascribe the large discrepancy to the significant heterogeneity of firms within industries: we find that firms within one sector only sell 23.5% to and buy 19.3% from the same industries on average, emphasizing the strong limitations of industrial sectors for representing the firms they include. Similar error-levels are expected when estimating economic growth, CO2 emissions, and the impact of policy interventions with industry-level IO models. Granular data is key for reasonable predictions of dynamical economic systems.

Urgently needed carbon emissions reductions might lead to strict command-and-control decarbonization strategies with potentially negative economic consequences. Analysing the entire firm-level production network of a European economy, we have explored how the worst outcomes of such approaches can be avoided. We compared the systemic relevance of every firm in Hungary with its annual CO2 emissions to identify optimal emission-reducing strategies with a minimum of additional unemployment and economic losses. Setting specific reduction targets, we studied various decarbonization scenarios and quantified their economic consequences. We determined that for an emissions reduction of 20%, the most effective strategy leads to losses of about 2% of jobs and 2% of economic output. In contrast, a naive scenario targeting the largest emitters first results in 28% job losses and 33% output reduction for the same target. This demonstrates that it is possible to use firm-level production networks to design highly effective decarbonization strategies that practically preserve employment and economic output., Comment: Nature Sustainability (2024)

We introduce SPEAR-TTS, a multi-speaker text-to-speech (TTS) system that can be trained with minimal supervision. By combining two types of discrete speech representations, we cast TTS as a composition of two sequence-to-sequence tasks: from text to high-level semantic tokens (akin to "reading") and from semantic tokens to low-level acoustic tokens ("speaking"). Decoupling these two tasks enables training of the "speaking" module using abundant audio-only data, and unlocks the highly efficient combination of pretraining and backtranslation to reduce the need for parallel data when training the "reading" component. To control the speaker identity, we adopt example prompting, which allows SPEAR-TTS to generalize to unseen speakers using only a short sample of 3 seconds, without any explicit speaker representation or speaker-id labels. Our experiments demonstrate that SPEAR-TTS achieves a character error rate that is competitive with state-of-the-art methods using only 15 minutes of parallel data, while matching ground-truth speech in terms of naturalness and acoustic quality, as measured in subjective tests.

We introduce MusicLM, a model generating high-fidelity music from text descriptions such as "a calming violin melody backed by a distorted guitar riff". MusicLM casts the process of conditional music generation as a hierarchical sequence-to-sequence modeling task, and it generates music at 24 kHz that remains consistent over several minutes. Our experiments show that MusicLM outperforms previous systems both in audio quality and adherence to the text description. Moreover, we demonstrate that MusicLM can be conditioned on both text and a melody in that it can transform whistled and hummed melodies according to the style described in a text caption. To support future research, we publicly release MusicCaps, a dataset composed of 5.5k music-text pairs, with rich text descriptions provided by human experts., Comment: Supplementary material at https://google-research.github.io/seanet/musiclm/examples and https://kaggle.com/datasets/googleai/musiccaps

Following some previous papers, we continue in this paper to give proof of the analogy between the acoustic world and the electromagnetic world. Hence we derive the expressions for: the scattering-scattering force (the electro-acoustic force) between the bubbles found in the inner cluster and the bubble found out of the cluster, the scattering-absorption force (the gravito-acoustic forces) between the bubbles found in the inner cluster, the gravito-acoustic forces between the cluster as a whole and an outer bubble, the temperature corresponding to the translational motion of the bubbles in the inner cluster and the average pressure of the acoustic radiation. The results of our calculus led us to find out that the absolute value of the average pressure of the acoustic radiation around a bubble is equal to the density of the energy of the electro-acoustic field, Eq. (10). This is the most important result of this paper. These densities are the densities of the energy corresponding to the oscillation of the bubble (the liquid around the bubble) that are involved in the interaction phenomenon between two oscillating bubbles. We have also demonstrated that the gravito-acoustic forces, at resonance, between the bubbles inside the cluster, generated by the absorption of energy in the bubbles, are proportional to the square of the virtual masses of the bubbles, Eqs. (29) and (30)., Comment: 19

We introduce AudioLM, a framework for high-quality audio generation with long-term consistency. AudioLM maps the input audio to a sequence of discrete tokens and casts audio generation as a language modeling task in this representation space. We show how existing audio tokenizers provide different trade-offs between reconstruction quality and long-term structure, and we propose a hybrid tokenization scheme to achieve both objectives. Namely, we leverage the discretized activations of a masked language model pre-trained on audio to capture long-term structure and the discrete codes produced by a neural audio codec to achieve high-quality synthesis. By training on large corpora of raw audio waveforms, AudioLM learns to generate natural and coherent continuations given short prompts. When trained on speech, and without any transcript or annotation, AudioLM generates syntactically and semantically plausible speech continuations while also maintaining speaker identity and prosody for unseen speakers. Furthermore, we demonstrate how our approach extends beyond speech by generating coherent piano music continuations, despite being trained without any symbolic representation of music.

In this paper, we deduce the expression of the gravito-acoustic force between two oscillating bubbles using the hypothesis that this type of force is a force of scattering-absorption of the energy of excitatory waves. The expression of the gravito-acoustic force at resonance highlights the dependence of this force on the product of the virtual masses of the two bubbles and on an acoustic gravitational constant. The acoustic gravitational constant depends on the absorption damping coefficient. We may say also that the expression of the acoustic gravitational constant is analogous to the expression of the gravitational constant in the electromagnetic world, that one obtained in the Einstein-Sciama model and the Dirac-Eddington large numbers hypothesis. The results obtained for this type of phenomenon in the acoustic world support the similarity between the acoustic world and the electromagnetic world., Comment: 9 pages

We present a method to separate speech signals from noisy environments in the embedding space of a neural audio codec. We introduce a new training procedure that allows our model to produce structured encodings of audio waveforms given by embedding vectors, where one part of the embedding vector represents the speech signal, and the rest represent the environment. We achieve this by partitioning the embeddings of different input waveforms and training the model to faithfully reconstruct audio from mixed partitions, thereby ensuring each partition encodes a separate audio attribute. As use cases, we demonstrate the separation of speech from background noise or from reverberation characteristics. Our method also allows for targeted adjustments of the audio output characteristics., Comment: Accepted at ICASSP 2023

We propose SpeechPainter, a model for filling in gaps of up to one second in speech samples by leveraging an auxiliary textual input. We demonstrate that the model performs speech inpainting with the appropriate content, while maintaining speaker identity, prosody and recording environment conditions, and generalizing to unseen speakers. Our approach significantly outperforms baselines constructed using adaptive TTS, as judged by human raters in side-by-side preference and MOS tests., Comment: Submitted to Interspeech 2022

Purpose: Oscillating gradient spin echo (OGSE) sequences have demonstrated an ability to probe time-dependent microstructural features, though they often suffer from low SNR due to increased echo times. In this work we introduce frequency tuned bipolar (FTB) gradients as a variation of oscillating gradients with reduced echo time and demonstrate their utility by mapping the frequency dispersion of diffusion kurtosis in human subjects. Methods: An FTB oscillating gradient waveform is presented that provides encoding of 1.5 net oscillation periods thereby reducing the echo time of the acquisition. Simulations were performed to determine an optimal protocol based on SNR of kurtosis frequency dispersion - defined as the difference in kurtosis between pulsed and oscillating gradient acquisitions. Healthy human subjects were scanned at 7T using pulsed gradient and an optimized 23 Hz FTB protocol, which featured a maximum b-value of 2500 s/mm2. In addition, to directly compare existing methods, measurements using traditional cosine OGSE were also acquired. Results: FTB oscillating gradients demonstrated equivalent frequency dependent diffusion measurements compared to cosine modulated OGSE while enabling a significant reduction in echo time. Optimization and in vivo results suggest FTB gradients provide increased SNR of kurtosis dispersion maps compared to traditional cosine OGSE. The optimized FTB gradient protocol demonstrated consistent reductions in apparent kurtosis values and increased diffusivity in generated frequency dispersion maps. Conclusion: This work presents an alternative to traditional cosine OGSE sequences enabling more time efficient acquisitions of frequency dependent diffusion quantities as demonstrated through in vivo kurtosis frequency dispersion maps.

The increasing availability of massive data sets poses a series of challenges for machine learning. Prominent among these is the need to learn models under hardware or human resource constraints. In such resource-constrained settings, a simple yet powerful approach is to operate on small subsets of the data. Coresets are weighted subsets of the data that provide approximation guarantees for the optimization objective. However, existing coreset constructions are highly model-specific and are limited to simple models such as linear regression, logistic regression, and $k$-means. In this work, we propose a generic coreset construction framework that formulates the coreset selection as a cardinality-constrained bilevel optimization problem. In contrast to existing approaches, our framework does not require model-specific adaptations and applies to any twice differentiable model, including neural networks. We show the effectiveness of our framework for a wide range of models in various settings, including training non-convex models online and batch active learning.

Crises like COVID-19 or the Japanese earthquake in 2011 exposed the fragility of corporate supply networks. The production of goods and services is a highly interdependent process and can be severely impacted by the default of critical suppliers or customers. While knowing the impact of individual companies on national economies is a prerequisite for efficient risk management, the quantitative assessment of the involved economic systemic risks (ESR) is hitherto practically non-existent, mainly because of a lack of fine-grained data in combination with coherent methods. Based on a unique value added tax dataset we derive the detailed production network of an entire country and present a novel approach for computing the ESR of all individual firms. We demonstrate that a tiny fraction (0.035%) of companies has extraordinarily high systemic risk impacting about 23% of the national economic production should any of them default. Firm size alone cannot explain the ESR of individual companies; their position in the production networks does matter substantially. If companies are ranked according to their economic systemic risk index (ESRI), firms with a rank above a characteristic value have very similar ESRI values, while for the rest the rank distribution of ESRI decays slowly as a power-law; 99.8% of all companies have an impact on less than 1% of the economy. We show that the assessment of ESR is impossible with aggregate data as used in traditional Input-Output Economics. We discuss how simple policies of introducing supply chain redundancies can reduce ESR of some extremely risky companies.

By modeling the particle as a two-dimensional oscillator with the natural angular frequency equal to the Zitterbewegung frequency, the expression of the gravitational force between two particles is obtained. Gravitational force is the effect of the absorption-scattering of the CZPF background by the oscillators. The connection between the gravitational and the electrostatic interaction is obtained., Comment: 13 pages

Assuming the charged particle to be a two-dimensional oscillator that scatters the classical background of zero-point field one can deduce the Coulomb force of the two interacting particles. The correct deduction of the force is conditioned by the equality between the natural angular frequency of the oscillator and the angular frequency of Zitterbewegung., Comment: 11 pages

A crucial aspect for the successful deployment of audio-based models "in-the-wild" is the robustness to the transformations introduced by heterogeneous acquisition conditions. In this work, we propose a method to perform one-shot microphone style transfer. Given only a few seconds of audio recorded by a target device, MicAugment identifies the transformations associated to the input acquisition pipeline and uses the learned transformations to synthesize audio as if it were recorded under the same conditions as the target audio. We show that our method can successfully apply the style transfer to real audio and that it significantly increases model robustness when used as data augmentation in the downstream tasks.

Active learning is an effective technique for reducing the labeling cost by improving data efficiency. In this work, we propose a novel batch acquisition strategy for active learning in the setting where the model training is performed in a semi-supervised manner. We formulate our approach as a data summarization problem via bilevel optimization, where the queried batch consists of the points that best summarize the unlabeled data pool. We show that our method is highly effective in keyword detection tasks in the regime when only few labeled samples are available.

In this paper, it is shown that in interaction with an oscillating bubbles cluster the fluid becomes inhomogeneous. The radial variation of the acoustic refractive index of the fluid generates an acoustic lens with spherical symmetry. When the focal length of the lens associated with the bubbles cluster is equal to the length of the position vector then the bubbles cluster behaves like a dumb hole. Unlike the properties of a bubble, the cluster, by compression under the action of attractive internal forces, can shrink its radius to become a dumb hole., Comment: 6 pages

Coresets are small data summaries that are sufficient for model training. They can be maintained online, enabling efficient handling of large data streams under resource constraints. However, existing constructions are limited to simple models such as k-means and logistic regression. In this work, we propose a novel coreset construction via cardinality-constrained bilevel optimization. We show how our framework can efficiently generate coresets for deep neural networks, and demonstrate its empirical benefits in continual learning and in streaming settings., Comment: NeurIPS 2020

We describe a new method for the rapid determination of the mass of particles confined in a free-space optical dipole-force trap. The technique relies on direct imaging of drop-and-restore experiments without the need for a vacuum environment. In these experiments, the trapping light is rapidly shuttered with an acousto-optic modulator causing the particle to be released from and subsequently recaptured by the trapping force. The trajectories of both the falls and restorations, imaged using a high-speed CMOS sensor, are combined to determine the particle mass. We corroborate these measurements using an analysis of position autocorrelation functions of the trapped particles. We report a statistical uncertainty of less than 2% for masses on the order of $5\times10^{-14}$ kg using a data acquisition time of approximately 90 seconds.

The aim of this paper is to investigate the coupled oscillations of multiple bubbles within a cluster. The interaction between a bubble and the other bubbles in a cluster produces an additional mass. For a fixed number of bubbles ( ) and uniformly distributed, in case of a certain value of the bubble density, we deduce the relations analogous to the Eddington relation (between the cluster radius and the bubble radius) and the Sciama relation (between the cluster radius and the gravitoacoustic radius) according to Mach's Principle., Comment: 8 pages

Recent advances in Neural Architecture Search (NAS) have produced state-of-the-art architectures on several tasks. NAS shifts the efforts of human experts from developing novel architectures directly to designing architecture search spaces and methods to explore them efficiently. The search space definition captures prior knowledge about the properties of the architectures and it is crucial for the complexity and the performance of the search algorithm. However, different search space definitions require restarting the learning process from scratch. We propose a novel agent based on the Transformer that supports joint training and efficient transfer of prior knowledge between multiple search spaces and tasks., Comment: 6th ICML Workshop on Automated Machine Learning

In this paper it was shown that, under certain restrictive conditions, Bjerknes secondary forces are attractive and proportionate to the product of the virtual masses of the two bubbles., Comment: 10 pages

Adaptive importance sampling for stochastic optimization is a promising approach that offers improved convergence through variance reduction. In this work, we propose a new framework for variance reduction that enables the use of mixtures over predefined sampling distributions, which can naturally encode prior knowledge about the data. While these sampling distributions are fixed, the mixture weights are adapted during the optimization process. We propose VRM, a novel and efficient adaptive scheme that asymptotically recovers the best mixture weights in hindsight and can also accommodate sampling distributions over sets of points. We empirically demonstrate the versatility of VRM in a range of applications.