Reducing the Energy Consumption of sEMG-Based Gesture Recognition at the Edge Using Transformers and Dynamic Inference.

Hand gesture recognition applications based on surface electromiographic (sEMG) signals can benefit from on-device execution to achieve faster and more predictable response times and higher energy efficiency. However, deploying state-of-the-art deep learning (DL) models for this task on memory-constrained and battery-operated edge devices, such as wearables, requires a careful optimization process, both at design time, with an appropriate tuning of the DL models' architectures, and at execution time, where the execution of large and computationally complex models should be avoided unless strictly needed. In this work, we pursue both optimization targets, proposing a novel gesture recognition system that improves upon the state-of-the-art models both in terms of accuracy and efficiency. At the level of DL model architecture, we apply for the first time tiny transformer models (which we call bioformers) to sEMG-based gesture recognition. Through an extensive architecture exploration, we show that our most accurate bioformer achieves a higher classification accuracy on the popular Non-Invasive Adaptive hand Prosthetics Database 6 (Ninapro DB6) dataset compared to the state-of-the-art convolutional neural network (CNN) TEMPONet (+3.1%). When deployed on the RISC-V-based low-power system-on-chip (SoC) GAP8, bioformers that outperform TEMPONet in accuracy consume 7.8×-44.5× less energy per inference. At runtime, we propose a three-level dynamic inference approach that combines a shallow classifier, i.e., a random forest (RF) implementing a simple "rest detector" with two bioformers of different accuracy and complexity, which are sequentially applied to each new input, stopping the classification early for "easy" data. With this mechanism, we obtain a flexible inference system, capable of working in many different operating points in terms of accuracy and average energy consumption. On GAP8, we obtain a further 1.03×-1.35× energy reduction compared to static bioformers at iso-accuracy.

Spike encoding techniques for IoT time-varying signals benchmarked on a neuromorphic classification task.

Spiking Neural Networks (SNNs), known for their potential to enable low energy consumption and computational cost, can bring significant advantages to the realm of embedded machine learning for edge applications. However, input coming from standard digital sensors must be encoded into spike trains before it can be elaborated with neuromorphic computing technologies. We present here a detailed comparison of available spike encoding techniques for the translation of time-varying signals into the event-based signal domain, tested on two different datasets both acquired through commercially available digital devices: the Free Spoken Digit dataset (FSD), consisting of 8-kHz audio files, and the WISDM dataset, composed of 20-Hz recordings of human activity through mobile and wearable inertial sensors. We propose a complete pipeline to benchmark these encoding techniques by performing time-dependent signal classification through a Spiking Convolutional Neural Network (sCNN), including a signal preprocessing step consisting of a bank of filters inspired by the human cochlea, feature extraction by production of a sonogram, transfer learning via an equivalent ANN, and model compression schemes aimed at resource optimization. The resulting performance comparison and analysis provides a powerful practical tool, empowering developers to select the most suitable coding method based on the type of data and the desired processing algorithms, and further expands the applicability of neuromorphic computational paradigms to embedded sensor systems widely employed in the IoT and industrial domains.

Q-PPG: Energy-Efficient PPG-Based Heart Rate Monitoring on Wearable Devices.

Hearth Rate (HR) monitoring is increasingly performed in wrist-worn devices using low-cost photoplethysmography (PPG) sensors. However, Motion Artifacts (MAs) caused by movements of the subject's arm affect the performance of PPG-based HR tracking. This is typically addressed coupling the PPG signal with acceleration measurements from an inertial sensor. Unfortunately, most standard approaches of this kind rely on hand-tuned parameters, which impair their generalization capabilities and their applicability to real data in the field. In contrast, methods based on deep learning, despite their better generalization, are considered to be too complex to deploy on wearable devices. In this work, we tackle these limitations, proposing a design space exploration methodology to automatically generate a rich family of deep Temporal Convolutional Networks (TCNs) for HR monitoring, all derived from a single "seed" model. Our flow involves a cascade of two Neural Architecture Search (NAS) tools and a hardware-friendly quantizer, whose combination yields both highly accurate and extremely lightweight models. When tested on the PPG-Dalia dataset, our most accurate model sets a new state-of-the-art in Mean Absolute Error. Furthermore, we deploy our TCNs on an embedded platform featuring a STM32WB55 microcontroller, demonstrating their suitability for real-time execution. Our most accurate quantized network achieves 4.41 Beats Per Minute (BPM) of Mean Absolute Error (MAE), with an energy consumption of 47.65 mJ and a memory footprint of 412 kB. At the same time, the smallest network that obtains a MAE 8 BPM, among those generated by our flow, has a memory footprint of 1.9 kB and consumes just 1.79 mJ per inference.

A Deep Learning Approach to the Screening of Oncogenic Gene Fusions in Humans.

Gene fusions have a very important role in the study of cancer development. In this regard, predicting the probability of protein fusion transcripts of developing into a cancer is a very challenging and yet not fully explored research problem. To this date, all the available approaches in literature try to explain the oncogenic potential of gene fusions based on protein domain analysis, that is cancer-specific and not easy to adapt to newly developed information. In our work, we choose the raw protein sequences as the input baseline, and propose the use of deep learning, and more specifically Convolutional Neural Networks, to infer the oncogenity probability score of gene fusion transcripts and to group them into a number of categories (e.g., oncogenic/not oncogenic). This is an inherently flexible methodology that, unlike previous approaches, can be re-trained with very less efforts on newly available data (for example, from a different cancer). Based on experimental results on a large dataset of pre-annotated gene fusions, our method is able to predict the oncogenity potential of gene fusion transcripts with accuracy of about 72%, which increases to 86% if we consider the only instances that are classified with a high confidence level.

LAPSE: Low-Overhead Adaptive Power Saving and Contrast Enhancement for OLEDs.

Organic Light Emitting Diode (OLED) display panels are becoming increasingly popular especially in mobile devices; one of the key characteristics of these panels is that their power consumption strongly depends on the displayed image. In this paper we propose LAPSE, a new methodology to concurrently reduce the energy consumed by an OLED display and enhance the contrast of the displayed image, that relies on image-specific pixel-by-pixel transformations. Unlike previous approaches, LAPSE focuses specifically on reducing the overheads required to implement the transformation at runtime. To this end, we propose a transformation that can be executed in real time, either in software, with low time overhead, or in a hardware accelerator with a small area and low energy budget. Despite the significant reduction in complexity, we obtain comparable results to those achieved with more complex approaches in terms of power saving and image quality. Moreover, our method allows to easily explore the full quality-versus-power tradeoff by acting on a few basic parameters; thus, it enables the runtime selection among multiple display quality settings, according to the status of the system.

VDJSeq-Solver: in silico V(D)J recombination detection tool.

In this paper we present VDJSeq-Solver, a methodology and tool to identify clonal lymphocyte populations from paired-end RNA Sequencing reads derived from the sequencing of mRNA neoplastic cells. The tool detects the main clone that characterises the tissue of interest by recognizing the most abundant V(D)J rearrangement among the existing ones in the sample under study. The exact sequence of the clone identified is capable of accounting for the modifications introduced by the enzymatic processes. The proposed tool overcomes limitations of currently available lymphocyte rearrangements recognition methods, working on a single sequence at a time, that are not applicable to high-throughput sequencing data. In this work, VDJSeq-Solver has been applied to correctly detect the main clone and identify its sequence on five Mantle Cell Lymphoma samples; then the tool has been tested on twelve Diffuse Large B-Cell Lymphoma samples. In order to comply with the privacy, ethics and intellectual property policies of the University Hospital and the University of Verona, data is available upon request to supporto.utenti@ateneo.univr.it after signing a mandatory Materials Transfer Agreement. VDJSeq-Solver JAVA/Perl/Bash software implementation is free and available at http://eda.polito.it/VDJSeq-Solver/.

Unsupervised HEp-2 mitosis recognition in indirect immunofluorescence imaging.

Automated HEp-2 mitotic cell recognition in IIF images is an important and yet scarcely explored step in the computer-aided diagnosis of autoimmune disorders. Such step is necessary to assess the goodness of the HEp-2 samples and helps the early diagnosis of the most difficult or ambiguous cases. In this work, we propose a completely unsupervised approach for HEp-2 mitotic cell recognition that overcomes the problem of mitotic/non-mitotic class imbalance due to the limited number of mitotic cells. Our technique automatically selects a limited set of candidate cells from the HEp-2 slide and then applies a clustering algorithm to identify the mitotic ones based on their texture. Finally, a second stage of clustering discriminates between positive and negative mitoses. Experiments on public IIF images demonstrate the performance of our technique compared to previous approaches.

The Human Brain Project and neuromorphic computing.

Understanding how the brain manages billions of processing units connected via kilometers of fibers and trillions of synapses, while consuming a few tens of Watts could provide the key to a completely new category of hardware (neuromorphic computing systems). In order to achieve this, a paradigm shift for computing as a whole is needed, which will see it moving away from current "bit precise" computing models and towards new techniques that exploit the stochastic behavior of simple, reliable, very fast, lowpower computing devices embedded in intensely recursive architectures. In this paper we summarize how these objectives will be pursued in the Human Brain Project.

Gelsius: a literature-based workflow for determining quantitative associations between genes and biological processes.

An effective knowledge extraction and quantification methodology from biomedical literature would allow the researcher to organize and analyze the results of high-throughput experiments on microarrays and next-generation sequencing technologies. Despite the large amount of raw information available on the web, a tool able to extract a measure of the correlation between a list of genes and biological processes is not yet available. In this paper, we present Gelsius, a workflow that incorporates biomedical literature to quantify the correlation between genes and terms describing biological processes. To achieve this target, we build different modules focusing on query expansion and document cononicalization. In this way, we reached to improve the measurement of correlation, performed using a latent semantic analysis approach. To the best of our knowledge, this is the first complete tool able to extract a measure of genes-biological processes correlation from literature. We demonstrate the effectiveness of the proposed workflow on six biological processes and a set of genes, by showing that correlation results for known relationships are in accordance with definitions of gene functions provided by NCI Thesaurus. On the other side, the tool is able to propose new candidate relationships for later experimental validation. The tool is available at >http://bioeda1.polito.it:8080/medSearchServlet/.

Acceleration of coarse grain molecular dynamics on GPU architectures.

Coarse grain (CG) molecular models have been proposed to simulate complex systems with lower computational overheads and longer timescales with respect to atomistic level models. However, their acceleration on parallel architectures such as graphic processing units (GPUs) presents original challenges that must be carefully evaluated. The objective of this work is to characterize the impact of CG model features on parallel simulation performance. To achieve this, we implemented a GPU-accelerated version of a CG molecular dynamics simulator, to which we applied specific optimizations for CG models, such as dedicated data structures to handle different bead type interactions, obtaining a maximum speed-up of 14 on the NVIDIA GTX480 GPU with Fermi architecture. We provide a complete characterization and evaluation of algorithmic and simulated system features of CG models impacting the achievable speed-up and accuracy of results, using three different GPU architectures as case studies.

Computer-aided techniques for chromogenic immunohistochemistry: status and directions.

Although immunohistochemistry (IHC) is a popular imaging technique, the quantitative analysis of IHC images via computer-aided methods is an emerging field that is gaining more and more importance thanks to the new developments in digital high-throughput scanners. In this paper, we discuss the main steps of IHC and review the techniques for computer-aided chromogenic IHC analysis, including methods to determine the location of interest of the antigens and quantify their activations. Moreover, we discuss the issues arising from the standardization of the immunostaining process, that are generally overlooked by the current literature, and finally provide requirements for reliable computer-aided IHC quantification.

Bellerophontes: an RNA-Seq data analysis framework for chimeric transcripts discovery based on accurate fusion model.

MOTIVATION: Next-generation sequencing technology allows the detection of genomic structural variations, novel genes and transcript isoforms from the analysis of high-throughput data. In this work, we propose a new framework for the detection of fusion transcripts through short paired-end reads which integrates splicing-driven alignment and abundance estimation analysis, producing a more accurate set of reads supporting the junction discovery and taking into account also not annotated transcripts. Bellerophontes performs a selection of putative junctions on the basis of a match to an accurate gene fusion model. RESULTS: We report the fusion genes discovered by the proposed framework on experimentally validated biological samples of chronic myelogenous leukemia (CML) and on public NCBI datasets, for which Bellerophontes is able to detect the exact junction sequence. With respect to state-of-art approaches, Bellerophontes detects the same experimentally validated fusions, however, it is more selective on the total number of detected fusions and provides a more accurate set of spanning reads supporting the junctions. We finally report the fusions involving non-annotated transcripts found in CML samples. AVAILABILITY AND IMPLEMENTATION: Bellerophontes JAVA/Perl/Bash software implementation is free and available at http://eda.polito.it/bellerophontes/.

miREE: miRNA recognition elements ensemble.

BACKGROUND: Computational methods for microRNA target prediction are a fundamental step to understand the miRNA role in gene regulation, a key process in molecular biology. In this paper we present miREE, a novel microRNA target prediction tool. miREE is an ensemble of two parts entailing complementary but integrated roles in the prediction. The Ab-Initio module leverages upon a genetic algorithmic approach to generate a set of candidate sites on the basis of their microRNA-mRNA duplex stability properties. Then, a Support Vector Machine (SVM) learning module evaluates the impact of microRNA recognition elements on the target gene. As a result the prediction takes into account information regarding both miRNA-target structural stability and accessibility. RESULTS: The proposed method significantly improves the state-of-the-art prediction tools in terms of accuracy with a better balance between specificity and sensitivity, as demonstrated by the experiments conducted on several large datasets across different species. miREE achieves this result by tackling two of the main challenges of current prediction tools: (1) The reduced number of false positives for the Ab-Initio part thanks to the integration of a machine learning module (2) the specificity of the machine learning part, obtained through an innovative technique for rich and representative negative records generation. The validation was conducted on experimental datasets where the miRNA:mRNA interactions had been obtained through (1) direct validation where even the binding site is provided, or through (2) indirect validation, based on gene expression variations obtained from high-throughput experiments where the specific interaction is not validated in detail and consequently the specific binding site is not provided. CONCLUSIONS: The coupling of two parts: a sensitive Ab-Initio module and a selective machine learning part capable of recognizing the false positives, leads to an improved balance between sensitivity and specificity. miREE obtains a reasonable trade-off between filtering false positives and identifying targets. miREE tool is available online at http://didattica-online.polito.it/eda/miREE/

A molecular dynamics study of a miRNA:mRNA interaction.

In this paper we present a methodology to evaluate the binding free energy of a miRNA:mRNA complex through molecular dynamics (MD)-thermodynamic integration (TI) simulations. We applied our method to the Caenorhabditis elegans let-7 miRNA:lin-41 mRNA complex-a validated miRNA:mRNA interaction-in order to estimate the energetic stability of the structure. To make the miRNA:mRNA simulation possible and realistic, the methodology introduces specific solutions to overcome some of the general challenges of nucleic acid simulations and binding free energy computations that have been discussed widely in many previous research reports. The main features of the proposed methodology are: (1) positioning of the restraints imposed on the simulations in order to guarantee complex stability; (2) optimal sampling of the phase space to achieve satisfactory accuracy in the binding energy value; (3) determination of a suitable trade-off between computational costs and accuracy of binding free energy computation by the assessment of the scalability characteristics of the parallel simulations required for the TI. The experiments carried out demonstrate that MD simulations are a viable strategy for the study of miRNA binding characteristics, opening the way to the development of new computational target prediction methods based on three-dimensional structure information.

Automated segmentation of cells with IHC membrane staining.

This study presents a fully automated membrane segmentation technique for immunohistochemical tissue images with membrane staining, which is a critical task in computerized immunohistochemistry (IHC). Membrane segmentation is particularly tricky in immunohistochemical tissue images because the cellular membranes are visible only in the stained tracts of the cell, while the unstained tracts are not visible. Our automated method provides accurate segmentation of the cellular membranes in the stained tracts and reconstructs the approximate location of the unstained tracts using nuclear membranes as a spatial reference. Accurate cell-by-cell membrane segmentation allows per cell morphological analysis and quantification of the target membrane proteins that is fundamental in several medical applications such as cancer characterization and classification, personalized therapy design, and for any other applications requiring cell morphology characterization. Experimental results on real datasets from different anatomical locations demonstrate the wide applicability and high accuracy of our approach in the context of IHC analysis.

TOM: a web-based integrated approach for identification of candidate disease genes.

The massive production of biological data by means of highly parallel devices like microarrays for gene expression has paved the way to new possible approaches in molecular genetics. Among them the possibility of inferring biological answers by querying large amounts of expression data. Based on this principle, we present here TOM, a web-based resource for the efficient extraction of candidate genes for hereditary diseases. The service requires the previous knowledge of at least another gene responsible for the disease and the linkage area, or else of two disease associated genetic intervals. The algorithm uses the information stored in public resources, including mapping, expression and functional databases. Given the queries, TOM will select and list one or more candidate genes. This approach allows the geneticist to bypass the costly and time consuming tracing of genetic markers through entire families and might improve the chance of identifying disease genes, particularly for rare diseases. We present here the tool and the results obtained on known benchmark and on hereditary predisposition to familial thyroid cancer. Our algorithm is available at http://www-micrel.deis.unibo.it/~tom/.

Automated DNA fragments recognition and sizing through AFM image processing.

This paper presents an automated algorithm to determine DNA fragment size from atomic force microscope images and to extract the molecular profiles. The sizing of DNA fragments is a widely used procedure for investigating the physical properties of individual or protein-bound DNA molecules. Several atomic force microscope (AFM) real and computer-generated images were tested for different pixel and fragment sizes and for different background noises. The automated approach minimizes processing time with respect to manual and semi-automated DNA sizing. Moreover, the DNA molecule profile recognition can be used to perform further structural analysis. For computer-generated images, the root mean square error incurred by the automated algorithm in the length estimation is 0.6% for a 7.8 nm image pixel size and 0.34% for a 3.9 nm image pixel size. For AFM real images we obtain a distribution of lengths with a standard deviation of 2.3% of mean and a measured average length very close to the real one, with an error around 0.33%.

Automatic intrinsic DNA curvature computation from AFM images.

Critical information on several biological processes such as DNA-protein interactions and DNA transcription can be derived from analysis of DNA curvature. Under thermal perturbation, the curvature is composed of static and dynamic contributions, thus, can be described as the sum of intrinsic curvature and a fluctuation contribution. Without considering thermal agitations, the DNA curvature is reducible to the intrinsic component, which is a function of the DNA nucleotide sequence only. In this paper, we present an automated algorithm to determine the DNA intrinsic curvature profiles and the molecular spatial orientations in Atomic Force Microscope images. The algorithm allows to reconstruct the intrinsic curvature profile by filtering the thermal contribution. It detects fragment orientation on atomic force microscope images without labels with a percentage of correct molecular-orientation detection of 96.79% in computer-generated benchmarks, for molecules with a high curvature peak. The automated algorithm reconstructs the intrinsic curvature profile of DNA molecules with a mean square error of 3.8122 x 10(-4) rads over a profile with a central peak value of 0.196 rads, and 6.1 x 10(-3) rads over a curvature profile with two symmetric peaks of about 0.08 rads. Moreover, it correctly detects the location of the peaks in the molecules with a deviation of about 1% of molecule length.