Search

AI language models (LMs) show promise for biological sequence analysis. We re-engineered open-source LMs (GPT-2, BLOOM, DistilRoBERTa, ELECTRA, and Mamba, ranging from 70M to 12B parameters) for microbial sequence classification. The models achieved F1 scores up to 95 and operated 16,580x faster and at 2.9x the recall of BLASTP. They effectively classified the algal dark proteome - uncharacterized proteins comprising about 65% of total proteins - validated on new data including a new, complete Hi-C/Pacbio Chlamydomonas genome. Larger (>1B) LA4SR models reached high accuracy (F1 > 86) when trained on less than 2% of available data, rapidly achieving strong generalization capacity. High accuracy was achieved when training data had intact or scrambled terminal information, demonstrating robust generalization to incomplete sequences. Finally, we provide custom AI explainability software tools for attributing amino acid patterns to AI generative processes and interpret their outputs in evolutionary and biophysical contexts.

We study the problem of learning mixtures of Gaussians with approximate differential privacy. We prove that roughly $kd^2 + k^{1.5} d^{1.75} + k^2 d$ samples suffice to learn a mixture of $k$ arbitrary $d$-dimensional Gaussians up to low total variation distance, with differential privacy. Our work improves over the previous best result [AAL24b] (which required roughly $k^2 d^4$ samples) and is provably optimal when $d$ is much larger than $k^2$. Moreover, we give the first optimal bound for privately learning mixtures of $k$ univariate (i.e., $1$-dimensional) Gaussians. Importantly, we show that the sample complexity for privately learning mixtures of univariate Gaussians is linear in the number of components $k$, whereas the previous best sample complexity [AAL21] was quadratic in $k$. Our algorithms utilize various techniques, including the inverse sensitivity mechanism [AD20b, AD20a, HKMN23], sample compression for distributions [ABDH+20], and methods for bounding volumes of sumsets., Comment: 52 pages. To appear in Neural Information Processing Systems (NeurIPS), 2024

We demonstrate a modulation-free laser stabilization system using a cavity-coupled MZI with aided acquisition on a low-loss SiN chip, achieving more than an order-of-magnitude improvement in locking range and over 36 dB noise suppression., Comment: 3 pages, 3 figures

This paper analyses the performance of reconfigurable intelligent surface (RIS)-assisted device-to-device (D2D) communication systems, focusing on addressing co-channel interference, a prevalent issue due to the frequency reuse of sidelink in the underlay in-band D2D communications. In contrast to previous studies that either neglect interference or consider it only at the user, our research investigates a performance analysis in terms of outage probability (OP) for RIS-assisted D2D communication systems considering the presence of interference at both the user and the RIS. More specifically, we introduce a novel integral-form expression for an exact analysis of OP. Additionally, we present a new accurate approximation expression for OP, using the gamma distributions to approximate the fading of both desired and interference links, thereby yielding a closed-form expression. Nevertheless, both derived expressions, i.e., the exact integral-form and the approximate closed-form, contain special functions, such as Meijer's G-function and the parabolic cylinder function, which complicate real-time OP analysis. To circumvent this, we employ a deep neural network (DNN) for real-time OP prediction, trained with data generated by the exact expression. Moreover, we present a tight upper bound that quantifies the impact of interference on achievable diversity order and coding gain. We validate the derived expressions through Monte Carlo simulations. Our analysis reveals that while interference does not affect the system's diversity order, it significantly degrades the performance by reducing the coding gain. The results further demonstrate that increasing the number of RIS's reflecting elements is an effective strategy to mitigate the adverse effects of the interference on the system performance., Comment: 12 pages

Proto-Danksharding, proposed in Ethereum Improvement Proposal 4844 (EIP-4844), aims to incrementally improve the scalability of the Ethereum blockchain by introducing a new type of transaction known as blob-carrying transactions. These transactions incorporate binary large objects (blobs) of data that are stored off-chain but referenced and verified on-chain to ensure data availability. By decoupling data availability from transaction execution, Proto-Danksharding alleviates network congestion and reduces gas fees, laying the groundwork for future, more advanced sharding solutions. This letter provides an analytical model to derive the delay for these new transactions. We model the system as an $\mathrm{M/D}^B/1$ queue which we then find its steady state distribution through embedding a Markov chain and use of supplementary variable method. We show that transactions with more blobs but less frequent impose higher delays on the system compared to lower blobs but more frequent.

We consider the problem of private density estimation for mixtures of unrestricted high dimensional Gaussians in the agnostic setting. We prove the first upper bound on the sample complexity of this problem. Previously, private learnability of high dimensional GMMs was only known in the realizable setting [Afzali et al., 2024]. To prove our result, we exploit the notion of $\textit{list global stability}$ [Ghazi et al., 2021b,a] that was originally introduced in the context of private supervised learning. We define an agnostic variant of this definition, showing that its existence is sufficient for agnostic private density estimation. We then construct an agnostic list globally stable learner for GMMs.

Blood flow reconstruction in the vasculature is important for many clinical applications. However, in clinical settings, the available data are often quite limited. For instance, Transcranial Doppler ultrasound (TCD) is a noninvasive clinical tool that is commonly used in the clinical settings to measure blood velocity waveform at several locations on brain's vasculature. This amount of data is grossly insufficient for training machine learning surrogate models, such as deep neural networks or Gaussian process regression. In this work, we propose a Gaussian process regression approach based on physics-informed kernels, enabling near-real-time reconstruction of blood flow in data-poor regimes. We introduce a novel methodology to reconstruct the kernel within the vascular network, which is a non-Euclidean space. The proposed kernel encodes both spatiotemporal and vessel-to-vessel correlations, thus enabling blood flow reconstruction in vessels that lack direct measurements. We demonstrate that any prediction made with the proposed kernel satisfies the conservation of mass principle. The kernel is constructed by running stochastic one-dimensional blood flow simulations, where the stochasticity captures the epistemic uncertainties, such as lack of knowledge about boundary conditions and uncertainties in vasculature geometries. We demonstrate the performance of the model on three test cases, namely, a simple Y-shaped bifurcation, abdominal aorta, and the Circle of Willis in the brain., Comment: 20 pages, 9 figures

The significant advancements in integrated photonics have enabled high-speed and energy efficient systems for various applications from data communications and high-performance computing, to medical diagnosis, sensing and ranging. However, data storage in these systems has been dominated by electronic memories which necessitates signal conversion between optical and electrical as well as analog and digital domains, and data movement between processor and memory that reduce the speed and energy efficiency. To date, a scalable optical memory with optical control has remained an open problem. Here we report an integrated photonic set-reset latch as a fundamental optical static memory unit based on universal optical logic gates. While the proposed memory is compatible with different photonic platforms, its functionality is demonstrated on a programmable silicon photonic chip as a proof of concept. Optical set, reset, and complementary outputs, scalability to a large number of memory units via the independent latch supply light, and compatibility with different photonic platforms enable more efficient and lower latency optical processing systems.

Acousto-optic integration offers numerous applications including low-loss microwave signal processing, nonreciprocal light propagation, frequency comb generation, and broadband acousto-optic modulation. State-of-the-art acousto-optic systems are mainly implemented entirely using in-house fabrication processes, which despite excellent performance typically suffer from low yield and are not compatible with mass production through foundry processes. Here, we demonstrate a highly efficient wideband acousto-optic modulator (AOM) implemented on a silicon photonics foundry process enabling high-yield low-cost mass production of AOMs with other photonic and electronic devices on the same substrate. In the reported structure, a 150 ${\mu}$m long AlScN-based acoustic transducer launches surface acoustic waves (SAW), which modulate the light passing through a silicon optical waveguide. A modulation efficiency of -18.3 dB over a bandwidth of 112 MHz is achieved, which to our knowledge is the highest reported efficiency and bandwidth combination among silicon based AOMs, resulting in about an order of magnitude $BW(V_{\pi}L)^{-1}$ figure-of-merit improvement compared to the state-of-the-art CMOS compatible AOMs. The monolithically integrated acousto-optic platform developed in this work will pave the way for low-cost, miniature microwave filters, true time delays, frequency combs, and other signal processors with the advanced functionality offered by foundry-integrated photonic circuits.

This paper presents exact formulas for the probability distribution function (PDF) and moment generating function (MGF) of the sum-product of statistically independent but not necessarily identically distributed (i.n.i.d.) Nakagami-$m$ random variables (RVs) in terms of Meijer's G-function. Additionally, exact series representations are also derived for the sum of double-Nakagami RVs, providing useful insights on the trade-off between accuracy and computational cost. Simple asymptotic analytical expressions are provided to gain further insight into the derived formula, and the achievable diversity order is obtained. The suggested statistical properties are proved to be a highly useful tool for modeling parallel cascaded Nakagami-$m$ fading channels. The application of these new results is illustrated by deriving exact expressions and simple tight upper bounds for the outage probability (OP) and average symbol error rate (ASER) of several binary and multilevel modulation signals in intelligent reflecting surfaces (IRSs)-assisted communication systems operating over Nakagami-$m$ fading channels. It is demonstrated that the new asymptotic expression is highly accurate and can be extended to encompass a wider range of scenarios. To validate the theoretical frameworks and formulations, Monte-Carlo simulation results are presented. Additionally, supplementary simulations are provided to compare the derived results with two common types of approximations available in the literature, namely the central limit theorem (CLT) and gamma distribution., Comment: 11 pages

We study the problem of hypothesis selection under the constraint of local differential privacy. Given a class $\mathcal{F}$ of $k$ distributions and a set of i.i.d. samples from an unknown distribution $h$, the goal of hypothesis selection is to pick a distribution $\hat{f}$ whose total variation distance to $h$ is comparable with the best distribution in $\mathcal{F}$ (with high probability). We devise an $\varepsilon$-locally-differentially-private ($\varepsilon$-LDP) algorithm that uses $\Theta\left(\frac{k}{\alpha^2\min \{\varepsilon^2,1\}}\right)$ samples to guarantee that $d_{TV}(h,\hat{f})\leq \alpha + 9 \min_{f\in \mathcal{F}}d_{TV}(h,f)$ with high probability. This sample complexity is optimal for $\varepsilon<1$, matching the lower bound of Gopi et al. (2020). All previously known algorithms for this problem required $\Omega\left(\frac{k\log k}{\alpha^2\min \{ \varepsilon^2 ,1\}} \right)$ samples to work. Moreover, our result demonstrates the power of interaction for $\varepsilon$-LDP hypothesis selection. Namely, it breaks the known lower bound of $\Omega\left(\frac{k\log k}{\alpha^2\min \{ \varepsilon^2 ,1\}} \right)$ for the sample complexity of non-interactive hypothesis selection. Our algorithm breaks this barrier using only $\Theta(\log \log k)$ rounds of interaction. To prove our results, we define the notion of \emph{critical queries} for a Statistical Query Algorithm (SQA) which may be of independent interest. Informally, an SQA is said to use a small number of critical queries if its success relies on the accuracy of only a small number of queries it asks. We then design an LDP algorithm that uses a smaller number of critical queries.

During collaboration in XR (eXtended Reality), users typically share and interact with virtual objects in a common, shared virtual environment. Specifically, collaboration among users in Mixed Reality (MR) requires knowing their position, movement, and understanding of the visual scene surrounding their physical environments. Otherwise, one user could move an important virtual object to a position blocked by the physical environment for others. However, even for a single physical environment, 3D reconstruction takes a long time and the produced 3D data is typically very large in size. Also, these large amounts of 3D data take a long time to be streamed to receivers making real-time updates on the rendered scene challenging. Furthermore, many collaboration systems in MR require multiple devices, which take up space and make setup difficult. To address these challenges, in this paper, we describe a single-device system called Collaborative Adaptive Mixed Reality Environment (CAMRE). We build CAMRE using the scene understanding capabilities of HoloLens 2 devices to create shared MR virtual environments for each connected user and demonstrate using a Leader-Follower(s) paradigm: faster reconstruction and scene update times due to smaller data. Consequently, multiple users can receive shared, synchronized, and close-to-real-time latency virtual scenes from a chosen Leader, based on their physical position and movement. We also illustrate other expanded features of CAMRE MR virtual environment such as navigation using a real-time virtual mini-map and X-ray vision for handling adaptive wall opacity. We share several experimental results that evaluate the performance of CAMRE in terms of the network latency in sharing virtual objects and other capabilities.

We study the problem of estimating mixtures of Gaussians under the constraint of differential privacy (DP). Our main result is that $\text{poly}(k,d,1/\alpha,1/\varepsilon,\log(1/\delta))$ samples are sufficient to estimate a mixture of $k$ Gaussians in $\mathbb{R}^d$ up to total variation distance $\alpha$ while satisfying $(\varepsilon, \delta)$-DP. This is the first finite sample complexity upper bound for the problem that does not make any structural assumptions on the GMMs. To solve the problem, we devise a new framework which may be useful for other tasks. On a high level, we show that if a class of distributions (such as Gaussians) is (1) list decodable and (2) admits a "locally small'' cover (Bun et al., 2021) with respect to total variation distance, then the class of its mixtures is privately learnable. The proof circumvents a known barrier indicating that, unlike Gaussians, GMMs do not admit a locally small cover (Aden-Ali et al., 2021b).

Despite the long history of research on slotted ALOHA, the exact analysis of the average delay is still in question as the performance of each node is coupled with the activity of other nodes. In this paper, we consider a network comprised of two half-duplex transmitter nodes with asynchronous arrival traffic that follow the slotted ALOHA protocol. We propose a new queueing theoretic model based on the state-dependent queues to analyze the network. In addition, we derive the exact values of delay and stability region for each node. The numerical results demonstrate the accuracy of our proposed model., Comment: 5 pages, 5 figures

A variety of topological Hamiltonians have been demonstrated in photonic platforms, leading to fundamental discoveries and enhanced robustness in applications such as lasing, sensing, and quantum technologies. To date, each topological photonic platform implements a specific type of Hamiltonian with inexistent or limited reconfigurability. Here, we propose and demonstrate different topological models by using the same reprogrammable integrated photonics platform, consisting of a hexagonal mesh of silicon Mach-Zehnder interferometers with phase-shifters. We specifically demonstrate a one-dimensional Su-Schrieffer-Heeger Hamiltonian supporting a localized topological edge mode and a higher-order topological insulator based on a two-dimensional breathing Kagome Hamiltonian with three corner states. These results highlight a nearly universal platform for topological models that may fast-track research progress toward applications of topological photonics and other coupled systems.

Gradient descent-based backpropagation training is widely used in many neural network systems. However, photonic implementation of such method is not straightforward mainly since having both the nonlinear activation function and its gradient using standard integrated photonic components is challenging. Here, we demonstrate the realization of two commonly used neural nonlinear activation functions and their gradients on a silicon photonic platform. Our method leverages the nonlinear electro-optic response of a micro-disk modulator. As a proof of concept, the experimental results are incorporated into a neural network simulation platform to classify MNIST handwritten digits dataset where we classification accuracies of more than 97\% are achieved that are on par with those of ideal nonlinearities and gradients.

We consider the class of noisy multi-layered sigmoid recurrent neural networks with $w$ (unbounded) weights for classification of sequences of length $T$, where independent noise distributed according to $\mathcal{N}(0,\sigma^2)$ is added to the output of each neuron in the network. Our main result shows that the sample complexity of PAC learning this class can be bounded by $O (w\log(T/\sigma))$. For the non-noisy version of the same class (i.e., $\sigma=0$), we prove a lower bound of $\Omega (wT)$ for the sample complexity. Our results indicate an exponential gap in the dependence of sample complexity on $T$ for noisy versus non-noisy networks. Moreover, given the mild logarithmic dependence of the upper bound on $1/\sigma$, this gap still holds even for numerically negligible values of $\sigma$., Comment: arXiv admin note: text overlap with arXiv:2206.07199

Optical neural networks (ONNs) enable high speed parallel and energy efficient processing compared to conventional digital electronic counterparts. However, realizing large scale systems is an open problem. Among various integrated and non-integrated ONNs, free-space diffractive ONNs benefit from a large number of pixels of spatial light modulators to realize millions of neurons. However, a significant fraction of computation time and energy is consumed by the nonlinear activation function that is typically implemented using a camera sensor. Here, we propose a novel surface-normal photodetector (SNPD) with a nonlinear response to replace the camera sensor that enables about three orders of magnitude faster (5.7 us response time) and more energy efficient (less than 10 nW/pixel) response. Direct efficient vertical optical coupling, polarization insensitivity, inherent nonlinearity with no control electronics, low optical power requirements, and the possibility of implementing large scale arrays make the SNPD a promising nonlinear activation function for diffractive ONNs. To show the applicability, successful classification simulation of MNIST and Fashion MNIST datasets using the measured response of SNPD with accuracy comparable to that of an ideal ReLU function are demonstrated.

We study the problem of privately estimating the parameters of $d$-dimensional Gaussian Mixture Models (GMMs) with $k$ components. For this, we develop a technique to reduce the problem to its non-private counterpart. This allows us to privatize existing non-private algorithms in a blackbox manner, while incurring only a small overhead in the sample complexity and running time. As the main application of our framework, we develop an $(\varepsilon, \delta)$-differentially private algorithm to learn GMMs using the non-private algorithm of Moitra and Valiant [MV10] as a blackbox. Consequently, this gives the first sample complexity upper bound and first polynomial time algorithm for privately learning GMMs without any boundedness assumptions on the parameters. As part of our analysis, we prove a tight (up to a constant factor) lower bound on the total variation distance of high-dimensional Gaussians which can be of independent interest., Comment: Accepted in ICML 2023

In the recent years, various 3D mixed reality serious games have been developed for different applications such as physical training, rehabilitation, and education. Task performance in a serious game is a measurement of how efficiently and accurately users accomplish the game's objectives. Prior research includes a graph-based representation of tasks, e.g. Hierarchical Task Network (HTN), which only models a game's tasks but does not perform assessment. In this paper, we propose Assessment HTN (A-HTN), which both models the task efficiently and incorporates assessment logic for game objectives. Based on how the task performance is evaluated, A-HTN automatically performs: (a) Task-level Assessment by comparing object manipulations and (b) Action-level Assessment by comparing motion trajectories. The system can also categorize the task performance assessment into single user or multi-user based on who is being assessed. We showcase the effectiveness of the A-HTN using two 3D VR serious games: a hydrometer experiment and a multi-user chemistry experiment. The A-HTN experiments show a high correlation between instructor scores and the system generated scores indicating that the proposed A-HTN generalizes automatic assessment at par with Subject Matter Experts., Comment: 8 pages, 5 figures, 1 table

Inverse-designed Silicon photonic metastructures offer an efficient platform to perform analog computations with electromagnetic waves. However, due to computational difficulties, scaling up these metastructures to handle a large number of data channels is not trivial. Furthermore, a typical inverse-design procedure utilizes a small computational domain and therefore tends to employ resonant features to achieve its objectives. This results in structures that are narrow-bandwidth and highly sensitive to fabrication errors. Here, we employ a 2D inverse-design method based on the effective index approximation with a low-index contrast constraint. This results in compact amorphous lens systems which are generally feed-forward and low-resonance. We designed and experimentally demonstrated a vector-matrix product for a 2 x 2 and a 3 x 3 matrix. We also designed a 10 x 10 matrix using the proposed 2D computational method. These examples demonstrate that these techniques have the potential to enable larger-scale wave-based analog computing platforms., Comment: 29 pages (including the main text and the supplementary materials), 10 figures (4 for the main text and 6 for supplementary materials)