Simultaneous Weight and Architecture Optimization for Neural Networks

Neural networks are trained by choosing an architecture and training the parameters. The choice of architecture is often by trial and error or with Neural Architecture Search (NAS) methods. While NAS provides some automation, it often relies on discrete steps that optimize the architecture and then train the parameters. We introduce a novel neural network training framework that fundamentally transforms the process by learning architecture and parameters simultaneously with gradient descent. With the appropriate setting of the loss function, it can discover sparse and compact neural networks for given datasets. Central to our approach is a multi-scale encoder-decoder, in which the encoder embeds pairs of neural networks with similar functionalities close to each other (irrespective of their architectures and weights). To train a neural network with a given dataset, we randomly sample a neural network embedding in the embedding space and then perform gradient descent using our custom loss function, which incorporates a sparsity penalty to encourage compactness. The decoder generates a neural network corresponding to the embedding. Experiments demonstrate that our framework can discover sparse and compact neural networks maintaining a high performance.

Task-Agnostic Federated Learning

In the realm of medical imaging, leveraging large-scale datasets from various institutions is crucial for developing precise deep learning models, yet privacy concerns frequently impede data sharing. federated learning (FL) emerges as a prominent solution for preserving privacy while facilitating collaborative learning. However, its application in real-world scenarios faces several obstacles, such as task & data heterogeneity, label scarcity, non-identically distributed (non-IID) data, computational vaiation, etc. In real-world, medical institutions may not want to disclose their tasks to FL server and generalization challenge of out-of-network institutions with un-seen task want to join the on-going federated system. This study address task-agnostic and generalization problem on un-seen tasks by adapting self-supervised FL framework. Utilizing Vision Transformer (ViT) as consensus feature encoder for self-supervised pre-training, no initial labels required, the framework enabling effective representation learning across diverse datasets and tasks. Our extensive evaluations, using various real-world non-IID medical imaging datasets, validate our approach's efficacy, retaining 90\% of F1 accuracy with only 5\% of the training data typically required for centralized approaches and exhibiting superior adaptability to out-of-distribution task. The result indicate that federated learning architecture can be a potential approach toward multi-task foundation modeling.

Global Prediction of COVID-19 Variant Emergence Using Dynamics-Informed Graph Neural Networks

During the COVID-19 pandemic, a major driver of new surges has been the emergence of new variants. When a new variant emerges in one or more countries, other nations monitor its spread in preparation for its potential arrival. The impact of the variant and the timing of epidemic peaks in a country highly depend on when the variant arrives. The current methods for predicting the spread of new variants rely on statistical modeling, however, these methods work only when the new variant has already arrived in the region of interest and has a significant prevalence. The question arises: Can we predict when (and if) a variant that exists elsewhere will arrive in a given country and reach a certain prevalence? We propose a variant-dynamics-informed Graph Neural Network (GNN) approach. First, We derive the dynamics of variant prevalence across pairs of regions (countries) that applies to a large class of epidemic models. The dynamics suggest that ratios of variant proportions lead to simpler patterns. Therefore, we use ratios of variant proportions along with some parameters estimated from the dynamics as features in a GNN. We develop a benchmarking tool to evaluate variant emergence prediction over 87 countries and 36 variants. We leverage this tool to compare our GNN-based approach against our dynamics-only model and a number of machine learning models. Results show that the proposed dynamics-informed GNN method retrospectively outperforms all the baselines, including the currently pervasive framework of Physics-Informed Neural Networks (PINNs) that incorporates the dynamics in the loss function.

Reuse Kernels or Activations? A Flexible Dataflow for Low-latency Spectral CNN Acceleration

Spectral-domain CNNs have been shown to be more efficient than traditional spatial CNNs in terms of reducing computation complexity. However they come with a `kernel explosion' problem that, even after compression (pruning), imposes a high memory burden and off-chip bandwidth requirement for kernel access. This creates a performance gap between the potential acceleration offered by compression and actual FPGA implementation performance, especially for low-latency CNN inference. In this paper, we develop a principled approach to overcoming this performance gap and designing a low-latency, low-bandwidth, spectral sparse CNN accelerator on FPGAs. First, we analyze the bandwidth-storage tradeoff of sparse convolutional layers and locate communication bottlenecks. We then develop a dataflow for flexibly optimizing data reuse in different layers to minimize off-chip communication. Finally, we propose a novel scheduling algorithm to optimally schedule the on-chip memory access of multiple sparse kernels and minimize read conflicts. On a state-of-the-art FPGA platform, our design reduces data transfers by 42\% with DSP utilization up to 90\% and achieves inference latency of 9 ms for VGG16, compared to the baseline state-of-the-art latency of 68 ms.

Acoustic-to-articulatory inversion for dysarthric speech: Are pre-trained self-supervised representations favorable?

$ $Acoustic-to-articulatory inversion (AAI) involves mapping from the acoustic space to the articulatory space. Signal-processing features like the MFCCs, have been widely used for the AAI task. For subjects with dysarthric speech, AAI is challenging because of an imprecise and indistinct pronunciation. In this work, we perform AAI for dysarthric speech using representations from pre-trained self-supervised learning (SSL) models. We demonstrate the impact of different pre-trained features on this challenging AAI task, at low-resource conditions. In addition, we also condition x-vectors to the extracted SSL features to train a BLSTM network. In the seen case, we experiment with three AAI training schemes (subject-specific, pooled, and fine-tuned). The results, consistent across training schemes, reveal that DeCoAR, in the fine-tuned scheme, achieves a relative improvement of the Pearson Correlation Coefficient (CC) by ${\sim}$1.81\% and ${\sim}$4.56\% for healthy controls and patients, respectively, over MFCCs. In the unseen case, we observe similar average trends for different SSL features. Overall, SSL networks like wav2vec, APC, and DeCoAR, which are trained with feature reconstruction or future timestep prediction tasks, perform well in predicting dysarthric articulatory trajectories.

DTW+S: Shape-based Comparison of Time-series with Ordered Local Trend

Measuring distance or similarity between time-series data is a fundamental aspect of many applications including classification and clustering. Existing measures may fail to capture similarities due to local trends (shapes) and may even produce misleading results. Our goal is to develop a measure that looks for similar trends occurring around similar times and is easily interpretable for researchers in applied domains. This is particularly useful for applications where time-series have a sequence of meaningful local trends that are ordered, such as in epidemics (a surge to an increase to a peak to a decrease). We propose a novel measure, DTW+S, which creates an interpretable "closeness-preserving" matrix representation of the time-series, where each column represents local trends, and then it applies Dynamic Time Warping to compute distances between these matrices. We present a theoretical analysis that supports the choice of this representation. We demonstrate the utility of DTW+S in ensemble building and clustering of epidemic curves. We also demonstrate that our approach results in better classification compared to Dynamic Time Warping for a class of datasets, particularly when local trends rather than scale play a decisive role.

Spatio-Temporal Attention in Multi-Granular Brain Chronnectomes for Detection of Autism Spectrum Disorder

The traditional methods for detecting autism spectrum disorder (ASD) are expensive, subjective, and time-consuming, often taking years for a diagnosis, with many children growing well into adolescence and even adulthood before finally confirming the disorder. Recently, graph-based learning techniques have demonstrated impressive results on resting-state functional magnetic resonance imaging (rs-fMRI) data from the Autism Brain Imaging Data Exchange (ABIDE). We introduce IMAGIN, a multI-granular, Multi-Atlas spatio-temporal attention Graph Isomorphism Network, which we use to learn graph representations of dynamic functional brain connectivity (chronnectome), as opposed to static connectivity (connectome). The experimental results demonstrate that IMAGIN achieves a 5-fold cross-validation accuracy of 79.25%, which surpasses the current state-of-the-art by 1.5%. In addition, analysis of the spatial and temporal attention scores provides further validation for the neural basis of autism.

Random Forest of Epidemiological Models for Influenza Forecasting

Forecasting the hospitalizations caused by the Influenza virus is vital for public health planning so that hospitals can be better prepared for an influx of patients. Many forecasting methods have been used in real-time during the Influenza seasons and submitted to the CDC for public communication. The forecasting models range from mechanistic models, and auto-regression models to machine learning models. We hypothesize that we can improve forecasting by using multiple mechanistic models to produce potential trajectories and use machine learning to learn how to combine those trajectories into an improved forecast. We propose a Tree Ensemble model design that utilizes the individual predictors of our baseline model SIkJalpha to improve its performance. Each predictor is generated by changing a set of hyper-parameters. We compare our prospective forecasts deployed for the FluSight challenge (2022) to all the other submitted approaches. Our approach is fully automated and does not require any manual tuning. We demonstrate that our Random Forest-based approach is able to improve upon the forecasts of the individual predictors in terms of mean absolute error, coverage, and weighted interval score. Our method outperforms all other models in terms of the mean absolute error and the weighted interval score based on the mean across all weekly submissions in the current season (2022). Explainability of the Random Forest (through analysis of the trees) enables us to gain insights into how it improves upon the individual predictors.

TransforMAP: Transformer for Memory Access Prediction

Data Prefetching is a technique that can hide memory latency by fetching data before it is needed by a program. Prefetching relies on accurate memory access prediction, to which task machine learning based methods are increasingly applied. Unlike previous approaches that learn from deltas or offsets and perform one access prediction, we develop TransforMAP, based on the powerful Transformer model, that can learn from the whole address space and perform multiple cache line predictions. We propose to use the binary of memory addresses as model input, which avoids information loss and saves a token table in hardware. We design a block index bitmap to collect unordered future page offsets under the current page address as learning labels. As a result, our model can learn temporal patterns as well as spatial patterns within a page. In a practical implementation, this approach has the potential to hide prediction latency because it prefetches multiple cache lines likely to be used in a long horizon. We show that our approach achieves 35.67% MPKI improvement and 20.55% IPC improvement in simulation, higher than state-of-the-art Best-Offset prefetcher and ISB prefetcher.

Fine-Grained Address Segmentation for Attention-Based Variable-Degree Prefetching

Machine learning algorithms have shown potential to improve prefetching performance by accurately predicting future memory accesses. Existing approaches are based on the modeling of text prediction, considering prefetching as a classification problem for sequence prediction. However, the vast and sparse memory address space leads to large vocabulary, which makes this modeling impractical. The number and order of outputs for multiple cache line prefetching are also fundamentally different from text prediction. We propose TransFetch, a novel way to model prefetching. To reduce vocabulary size, we use fine-grained address segmentation as input. To predict unordered sets of future addresses, we use delta bitmaps for multiple outputs. We apply an attention-based network to learn the mapping between input and output. Prediction experiments demonstrate that address segmentation achieves 26% - 36% higher F1-score than delta inputs and 15% - 24% higher F1-score than page & offset inputs for SPEC 2006, SPEC 2017, and GAP benchmarks. Simulation results show that TransFetch achieves 38.75% IPC improvement compared with no prefetching, outperforming the best-performing rule-based prefetcher BOP by 10.44%, and ML-based prefetcher Voyager by 6.64%.