<?xml version="1.0" encoding="UTF-8"?><rss xmlns:dc="http://purl.org/dc/elements/1.1/" xmlns:content="http://purl.org/rss/1.0/modules/content/" xmlns:atom="http://www.w3.org/2005/Atom" version="2.0" xmlns:cc="http://cyber.law.harvard.edu/rss/creativeCommonsRssModule.html">
    <channel>
        <title><![CDATA[Stories by Hieu T. Tran on Medium]]></title>
        <description><![CDATA[Stories by Hieu T. Tran on Medium]]></description>
        <link>https://medium.com/@htran_91788?source=rss-1771d052fc74------2</link>
        <image>
            <url>https://cdn-images-1.medium.com/fit/c/150/150/2*e3wX6jB8i52O4xhC7JewKw.jpeg</url>
            <title>Stories by Hieu T. Tran on Medium</title>
            <link>https://medium.com/@htran_91788?source=rss-1771d052fc74------2</link>
        </image>
        <generator>Medium</generator>
        <lastBuildDate>Sat, 21 Nov 2026 05:50:15 GMT</lastBuildDate>
        <atom:link href="https://medium.com/@htran_91788/feed" rel="self" type="application/rss+xml"/>
        <webMaster><![CDATA[yourfriends@medium.com]]></webMaster>
        <atom:link href="http://medium.superfeedr.com/" rel="hub"/>
        <item>
            <title><![CDATA[Compute Plasticity for Edge Devices and Network]]></title>
            <link>https://medium.com/@htran_91788/compute-plasticity-for-edge-devices-and-network-74b7bcf1e58b?source=rss-1771d052fc74------2</link>
            <guid isPermaLink="false">https://medium.com/p/74b7bcf1e58b</guid>
            <category><![CDATA[iot]]></category>
            <category><![CDATA[5g-wireless]]></category>
            <category><![CDATA[health]]></category>
            <category><![CDATA[edge-computing]]></category>
            <category><![CDATA[ai]]></category>
            <dc:creator><![CDATA[Hieu T. Tran]]></dc:creator>
            <pubDate>Fri, 20 Sep 2019 02:02:47 GMT</pubDate>
            <atom:updated>2019-09-24T18:07:38.602Z</atom:updated>
            <content:encoded><![CDATA[<p>At a recent investor event, Tesla CEO, Elon Musk <em>promised</em> the company will soon roll out a network of one million Robotaxies in 2020, by his estimate. Given the 2TB to 6TB of data that autonomous vehicles are expected to generate each day, the Robotaxi network will produce 730 to 2,200 exabytes of data annually, or anywhere between 35% to 110% of the estimated capacity of the world data center storage for 2020.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/0*xQo6KiqQJGfVhnSe.png" /><figcaption><strong>World data center storage capacity</strong> (credit: Statista / Tesla)</figcaption></figure><p>This perceived storage gap is contrived since data must first make it to the data center before it can be stored and analyzed.</p><p>2TB/day approximates to 185 Mbit/s, which is <strong>5–10 times higher</strong> than than the average uplink <em>broadband</em> speed in the US <strong>today</strong>. We can conclude that Robotaxies will unlikely be live streaming over wireless soon. A more practical scenario is for the taxies to upload data over high speed links when they wait at charging stations. This however would limit the <strong>time relevance</strong> of the data taxies collected to an average of 1/2 of the period between charges. This and others are trade-offs that engineers at Tesla and other AV firms must weigh into their design decisions.</p><p>In this post, we look at how <strong>mainstream compute resources</strong> and <strong>plasticity</strong> can help connected devices, the likes of Robotaxis and mobile edge compute infrastructure, to realize their full data potential.</p><h3>Compute at the Edge</h3><p>Empirically, <strong>faster</strong> or <strong>better</strong> technology don’t always become mainstream. <strong>Solaris</strong>, <strong>AIX</strong>, and <strong>HP-UX</strong> all have security, performance, and other high value intangibles that serve them well in certain vertical markets. As for mainstream compute, it is the WinTel platform that is most known for:</p><ul><li><strong>Diversity</strong> — PCs come in different form factors, shapes, and functions,</li><li><strong>Availability</strong> — You can can buy a PC and accessories anywhere,</li><li><strong>Disaggregation</strong> Many suppliers and sourcing options for components,</li><li><strong>Costs</strong> — Supply chain diversity and high volume drive down costs,</li><li><strong>Flexibility</strong> — PCs are used for just about anything</li></ul><p>A Tesla running Windows? Teslows?</p><p>Perhaps not, but the Robotaxi does need <strong><em>a platform</em></strong> to offload the massive amount of data it generates. At minimum this platform is an onboard computer that can process and act on time-sensitive data in near real-time while storing and forwarding the remaining. In concert, multiple edge controllers can account for the availability of bandwidth and local compute / storage to orchestrate how much and what data to process. One or more policies can drive optimization to align with costs saving, profit, and utilization objectives.</p><p>Connected and mobile systems like this may vet well in controlled settings (e.g. in the Tesla parking lot or conferences). At scale however, careful designs, mainstream platforms, and proven tools are needed to harness the full value of collected data. In this case, that from <strong><em>one million</em></strong> Robotaxies roaming the Eastern seaboard.</p><p>We consider mainstream compute to offer advantages beyond the costs saving of utilizing readily available of-the-shelf hardware. The diversity of available tools and software makes application development easier and faster. There are also more programmers available with experience in mainstream software system.</p><p>And an <strong>AppStore-like programmable platform</strong> in which programmers willingly invest time and money to build apps for an infinitesimally small, but non-zero, chance of positive returns is slightly <strong>uber-chic.</strong></p><blockquote>ETSI incubates and hosts a number of proposed Proof of Concepts (PoCs) for Multi-access Edge Computing (MEC). One in particular: <strong>PoC#12 — <em>MEC enabled OTT business</em></strong> resonates well here. Imagine a third-party service could subsidize for riders’ fees if the service is allowed to track and confirm that they participate in brand-engagement activities. The <em>service</em> can also spy and snitch on its passengers, but give them paid options for opting out (e.g “<strong>RiskARide.com</strong>”).</blockquote><h3>The Radio Access Network — RAN</h3><p>As an increasingly common part of the urban landscape, the base stations masqueraded as conspicuous building fixtures or <em>phony trees</em> atop outdoor equipment rooms are parts of the radio access network (RAN). Within each base station are one or more tower-mounted antennae (RRUs) connected to the baseband units (BBUs) below.</p><p><strong>The roles of the RRUs in the base station are:</strong></p><ul><li>Transmit and receive radio signal to and from RF devices (phones, IoT products, Robotaxies, …),</li><li>Format and transport digital radio signal to the baseband processing (BBU) and networking unit over the physical (fiber/copper) link,</li></ul><p><strong>And the BBU:</strong></p><ul><li>Convert radio signal into IP packets and forward to the packet core network for processing.</li></ul><p>Functionally, base stations are similar to the home WiFI access points, except larger.</p><h4>C-RAN — Deconstructing the Base Station</h4><p>Bandwidth usage demand of base stations varies with the time of day, seasonality, and their respective location. As a fixed asset, the RRU and BBU of a given base station are underutilized during off-peak hours. The C-RAN architecture aims at putting underutilized resources to use by deconstructing the relationship between RRU and BBU.</p><p>C-RAN decouples RRUs from BBUs to allow:</p><ul><li><strong>Pooling of BBUs</strong> to minimize baseband resources needed per deployment.</li><li>Deployment of idle BBUs to service antennae in areas with peak demand.</li><li><strong>Flexible build out</strong> of RRUs with less footprint to improve coverage.</li><li><strong>Replacement</strong> of specialized equipment with standard servers <strong>with virtualization</strong>.</li></ul><p>These capabilities translate to a <strong>reduction in CapEx and OpEx</strong>, <strong>higher efficiency,</strong> and <strong>better agility</strong>.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/652/1*SxvSYw1Dl8CDkPLOaaAxzQ.png" /><figcaption><strong>C-RAN</strong> (credit: Researchgate.net)</figcaption></figure><p>In this diagram, the RRUs connects to a pool of BBUs over the fronthaul network. This network uses either the CPRI (Common Public Radio Interface) or OBSAI (Open Base Station Architecture Initiative) standard implemented by most vendors to transmit digital radio packets over fiber or copper.</p><p><strong>The 5G Transport Architecture:</strong></p><p>In order to meet the highly diverse needs for new services and deployment scenarios, the 5G radio interface (5G NR — New Radio) is designed to scale coverage spectrum from 700MHz to well above 28GHz in either the time or frequency domain.</p><p>To enhance flexibility and simplify virtualization of RAN functions, the transport architecture disaggregate the original 4G/LTE BBU into:</p><ul><li><strong>Central Unit (CU)</strong>,</li><li><strong>Distributed Unit (DU)</strong>, and</li><li><strong>Remote Radio Unit (RRU)</strong>.</li></ul><figure><img alt="" src="https://cdn-images-1.medium.com/max/579/1*TSrWNMyMMhztnjQmwlU7tA.png" /><figcaption><strong>The Evolution of eNB </strong>(credit: ITU)</figcaption></figure><p>Shown here is one possible deployment in which the DU, CU, and CU are separated over different wire-level interfaces. Other configurations collocating DU with RRU or all three into one gNB are also possible.</p><p>The compute and storage resources of these RAN units are designed to meet varying objectives. Generally, more offloading closer to sources (i.e. the edge devices) means less bandwidth needed to move data through the front and mid-haul to the packet core network. But offloading requires more complex design, and higher power consumption, increasing the likelihood of things breaking down. Frequent repair to remote radio heads mounted in precarious position could negatively impact worker safety and OpEx.</p><p>To help the network designers make design trade-offs, the 3GPP logical RAN architecture outlines different <strong>functional splits</strong> between the central and distributed units:</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/0*YxSCKAukdmsHp83M.png" /><figcaption><strong><em>3GPP TR 38.801 functional splits and CU and DU functions with a) high layer split, b) low layer split, or c) cascade split</em></strong><em> (credit: ITU)</em></figcaption></figure><p>We examine how compute plasticity can improve flexibility and fluidity in the implementation of these functional options.</p><h3>Compute Plasticity</h3><p>Until mid-century, scientists viewed the brain as a complex machine with specialized parts in preassigned locations to perform specific tasks. <strong>Cerebral localization</strong> reinforced this view by associating damages to parts the brain with the loss of sensory and mental functions commonly identified with those areas. Treatment or recovery are deemed improbable since damaged brain cells generally don’t regenerate.</p><blockquote>Counter to this viewpoint, the American neuroscientist Paul Bach-y-Rita showed that with sufficient (re)training, the brain flexes to compensate for sensory defects well into adulthood and throughout a person’s life.</blockquote><blockquote><strong>Neural plasticity</strong>, the ability of the nervous system to modify itself functionally and structurally, was the key to the techniques and methods devised by Bach-y-Rita to reroute brain activity in damaged areas to different locations as a means to treat patients with disabilities. Paul’s works upended long held beliefs in neuroscience, helped change people’s lives and restore their dignity.</blockquote><figure><img alt="" src="https://cdn-images-1.medium.com/max/780/0*PMvmqAxfKjABy4bz" /><figcaption><strong>Paul Bach-y-Rita</strong> (credit: UW-Madison)</figcaption></figure><p><strong>Compute plasticity</strong> likewise, is the the ability of a compute platform to flex with multiple functional requirements, from one deployment to another or at the same time. By this standard, Linux has higher plasticity than say, the Cray OS. Same with general compute by traditional ISA (x86, ARM, MIPS, PPC, …) versus FPGA.</p><blockquote>We may also be quick to incorrectly conclude that all <strong>hardware accelerators</strong> are the compute equivalence of cerebral localization. A counter example is nVidia, whose investors are sanguine about their ability to successfully transform the graphic accelerator into platform for AI/ML.</blockquote><p>We look at how compute plasticity plays a role in the mobile edge, specifically in the co-location of cell/signal processing with user processing.</p><h3>Cell versus User Processing</h3><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/0*1QQqnKthQPi15Mr2.png" /><figcaption><strong><em>Cell processing in 3GPP TR 38.801 option 6 Credit: 3GPP)</em></strong></figcaption></figure><p>In 3GPP TR 38.801 option 6, the <strong>MAC and upper layers</strong> are in the Central Unit (CU), whereas the <strong>physical layer and RF</strong> are in the Distributed Unit (DU). The Small Cell Forum proposed nFAPI, which is a set of interfaces for supporting this split option to connect the MAC and PHY.</p><p>Because cell processing in 5G in a given CU/DU deployment needs to adapt to diverse transmission spectrum and domain of the RRUs, the physical layer is commonly implemented as Software Defined Radio (SDR) in FPGA.</p><p>We show a simplified view of the up and downlink between the MAC and RRH, and describe the framework we developed to implement the cell processing functions on mainstream platform.</p><h3>Service Chaining with Teranium</h3><p>Our design used an internal tools called Teranium to graphically model the relationship and dataflow among physical elements. The LTE Physical Downlink Shared Channel (PDSCH) is the physical channel that carries Downlink data or Random Access Response Message when the PDSCH is mapped to RACH channel.</p><p>The Teranium model can be bound to different backend implementations. Teranium supports L4, orchestrated VMs and containers, Amazon AWS, and Click as backends. In this case, the Teranium tools generates backend for Click. The generated code runs on a software SDN framework that we developed called Dronava.</p><h4>The Click Modular Router</h4><p>Developed by Eddie Kohler and teams at MIT/UCLA, Click is a fast, modular packet processing and analysis framework. ( <em>Meraki — now Cisco — used it to implement their wireless network products</em>). Click comes with collection of software functions for packet processing, which are encapsulated as <strong>elements</strong> and implemented in C/C++ instead of some foodie VM-bound 4GL designed for vegans.</p><p>Elements are connected / service-chained together to implement particular packet flows. The Dronava framework extends the Click router with physical radio packet processing elements and performance optimization.</p><h4>Implementing LTE PDSCH with Teranium and Dronava</h4><p>Our use-case assumes a channel with bandwidth 10MHz to transfer data with coded bit rate at least 30 Mbps (after channel coding). When integrated to the physical layer, the PDSCH channel operates with parameters scheduled by the MAC, and shares resources with other channels.</p><p>The high level steps are:</p><ul><li>The PDSCH channel service chains together the <strong>Phy5gCrc</strong>, <strong>Ph5ygSegmentation</strong>, <strong>Phy5gTurboEncoder</strong>, <strong>Phy5gTurboSubblockInterleaver</strong>, <strong>Phy5gTurboRateMatcher</strong>, <strong>Phy5gScrambler</strong>, <strong>Phy5gModulationMapper</strong>, <strong>Phy5gLayerMapping</strong>, <strong>Phy5gPrecoding</strong>, <strong>Phy5gREMapping</strong>, and <strong>Phy5gPdschProcess</strong> elements.</li><li>To transmit time-domain data, <strong>Phy5gOfdmMod</strong> and <strong>Phy5gDft</strong> are also needed.</li><li>For 5G NR PDSCH (gNB), <strong>LDPC</strong> is is used instead of <strong>Turbo code</strong> for channel coding.</li></ul><p>Lastly, <strong>Phy5gPdschProcess</strong> is a helper utility element that supplies PDSCH channel parameters, data and temporary buffers to the signal processing blocks (PHY Elements).</p><p>Below is a graphical layout of the use-case as composed by the designer tools:</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/0*ddD_VgEcoPDNn81m.png" /><figcaption><strong><em>PDSCH in Teranium</em></strong><em> (credit: Teranium/Dronava team)</em></figcaption></figure><p>Theoretically, we can use all 50 resource blocks (RBs) for PDSCH. A resource block is the smallest unit of RF resource that can be assigned to a UE.</p><pre><strong>Modulation symbol rate</strong> (normal CP) <br>50RBs * 14 ofdm symbols * 12k subcarriers = <strong>8.4 Msps</strong> </pre><pre><strong>Modulation symbol rate </strong>(extended CP)<strong><br></strong>50RBs * 14 ofdm symbols * 12k subcarriers = <strong>7.2 Msps</strong></pre><p>Teranium generates the following Click code to configure the DLSCH channel:</p><pre>dlsch_channel::Phy5gPdschProcess( <br>     UNITEST true,  // <br>     MCS 17,        <strong><em>// 0-9: QPSK, 10-16 16QAM, 17-28: 64 QAM</em></strong>         <br>     N_RB_DL 50,    // <strong><em>50 mean 10MHz, 100 mean 20MHz </em></strong><br>     nb_antennas_tx 1, <br>     nb_antennas_rx 1, <br>     RNTI 100,      // <strong><em>must &gt;0 frame_type 0, //0 for FDD, 1 for TDD </em></strong>      <br>     N_RB_ALLOC 40, // <strong><em>must &lt; N_RB_DL CP_type 0 //0 if normal, 1 if extended <br></em></strong>);</pre><p>And the following code to service-chain the elements together:</p><pre><strong>crc_add::Phy5gCrc</strong>(<br>     CRC_TYPE 24A,<br>     CRC_MODE ATTACH, <br>     VERBOSE true); <br><strong>byte_generator</strong>(LENGTH 16000, NUM 1, SIZE_IN_BIT true) <br>-&gt;[6]dlsch_channel[6]-&gt;crc_add <br>-&gt;[0]dlsch_channel[0]-&gt;Phy5gSegmentation <br>-&gt;[1]dlsch_channel[1]-&gt;Phy5gTurboEncoder(ARCHITECTURE SSE) <br>-&gt;[2]dlsch_channel[2]-&gt;Phy5gTurboSubblockInterleaver <br>-&gt;[3]dlsch_channel[3]-&gt;Phy5gTurboRateMatcher <br>-&gt;[4]dlsch_channel[4]-&gt;Phy5gScrambler <br>-&gt;[5]dlsch_channel[5]-&gt;Phy5gModulationMapper <br>-&gt;<strong>Phy5gLayerMapping</strong>(<br>     layermap_type 1, <br>     number_codeword 1, <br>     number_layer 1, <br>     instruction 1)<br>-&gt;<strong>Phy5gPrecoding</strong>(<br>     precoding_type 1, <br>     number_layer 1, <br>     number_antport 1, <br>     cb_index 0, <br>     instruction 1)<br>-&gt;[7]dlsch_channel[7] <br>-&gt;[0]<strong>Phy5gREMapping</strong>;</pre><p>In a complete design, PDSCH must share resources with other channels such as PCFICH, PDCCH,PHICH and PBCH as well as PSS, SSS, and the reference signals.</p><blockquote>In our unit testing, we found the channel condition to be non-optimal when modulation is hard-coded to 64 QAM, 256 QAM, or 1024 QAM. UEs would occasionally detect errors, forcing the eNB / gNB to resend. Various channel aware scheduling algorithms to coordinate transmission are essential to achieving more ideal channel condition.</blockquote><h3>The Case for Compute Plasticity</h3><p>The Radio Frequency Machine Learning System (RFMLS) funded by DARPA aims to develop foundations for applying Machine Learning to address performance and security limitations of conventional RF systems.</p><p>We found from the PDSCH use-case that <strong>channel aware scheduling</strong> is necessary to coordinate transmission. This aligns with one of the operational needs by the DoD in the RF Spectrums. The additional requirements are :</p><ul><li><strong>RF Fingerprinting</strong> — Learning to identify UEs by their unique hardware transmission fingerprint.</li><li><strong>Fingerprint Homing</strong> — Training the UEs to fine-tune / enhance their hardware fingerprint.</li><li><strong>Autonomous Configuration</strong> — Channel-aware, autonomous configuration of RF parameters.</li></ul><p>We posit that mainstream platforms with high compute plasticity are a good match for these requirements:</p><ul><li>Cell processing (RF) and user processing (ML/training) can colocate in the same compute domain, minimizing the complexity and latency of cross-domain (e.g. over PCIe) interaction.</li><li>Availability of tools and framework accelerates testing/development and increases product quality.</li><li>Diversity of hardware options to fit power, budget, and performance needs.</li></ul><h3>Conclusion</h3><p>We implemented a radio physical processing framework with mainstream software and hardware platforms and tools. This framework allows us develop, debug, and measure the performance and efficacy of new algorithm and model (e.g. scheduling) in the PHY layer. Unit and integration test are straight forward, and implementation errors are easy to identify and fix. Switching to different transmission spectrum or domain can be done in real-time simply by changing packet configuration.</p><p>We look forward to the opportunity to discuss our related works on testing, performance optimization and vertical use-cases in subsequent posts.</p><h3>Credits</h3><p>The framework, results, and findings stated in this post are the works of our exceptional and dedicated team of engineers. From left to right:</p><blockquote><strong>Quyen Vu, Nhat Le, Hieu Hoang, Trang Cao, Minh Nguyen, Tiep Nguyen, and Phu Tran.</strong></blockquote><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/0*rkB1sjCHf0GmVCI0.jpg" /><figcaption>Viosoft Dronava Team</figcaption></figure><p>All errors, mistakes, and exaggerations are exclusively mine.</p><h3>References</h3><ol><li><a href="https://www.accenture.com/_acnmedia/pdf-73/accenture-autonomous-vehicles-the-race-is-on.pdf">https://www.accenture.com/_acnmedia/pdf-73/accenture-autonomous-vehicles-the-race-is-on.pdf</a></li><li><a href="https://www.vox.com/2017/9/7/16264430/fastest-broadband-speeds-ookla-city-internet-service-provider">https://www.vox.com/2017/9/7/16264430/fastest-broadband-speeds-ookla-city-internet-service-provider</a></li><li><a href="https://bigthink.com/technology-innovation/musk-Robotaxis">https://bigthink.com/technology-innovation/musk-Robotaxis</a></li><li><a href="https://www.statista.com/statistics/638593/worldwide-data-center-storage-capacity-cloud-vs-traditional/">https://www.statista.com/statistics/638593/worldwide-data-center-storage-capacity-cloud-vs-traditional/</a></li><li><a href="https://www.computerworld.com/article/3015576/2015-the-tech-year-in-cartoons.html#slide5">https://www.computerworld.com/article/3015576/2015-the-tech-year-in-cartoons.html#slide5</a></li><li><a href="https://www.5g-picture-project.eu/download/Complexity_EuCNC2019.pdf">https://www.5g-picture-project.eu/download/Complexity_EuCNC2019.pdf</a></li><li><a href="https://images.techhive.com/images/article/2015/12/cw_2015_05_klossner_windows10_onetimeuse_rsz-100633564-orig.jpg">https://images.techhive.com/images/article/2015/12/cw_2015_05_klossner_windows10_onetimeuse_rsz-100633564-orig.jpg</a></li><li><a href="https://dl.cdn-anritsu.com/images/tm/buy/en-us/promotions/bbu-use-case-gen3.png?la=en-us">https://dl.cdn-anritsu.com/images/tm/buy/en-us/promotions/bbu-use-case-gen3.png?la=en-us</a></li><li><a href="https://www.researchgate.net/publication/330659689/figure/fig1/AS:719497866530816@1548553067261/Mobile-backhaul-and-fronthaul-network-architecture-in-LTE-extracted-from-5.jpg">https://www.researchgate.net/publication/330659689/figure/fig1/AS:719497866530816@1548553067261/Mobile-backhaul-and-fronthaul-network-architecture-in-LTE-extracted-from-5.jpg</a></li><li><a href="http://5g-crosshaul.eu/wp-content/uploads/2015/05/2016_IEEECommsMag_CPRI.pdf">http://5g-crosshaul.eu/wp-content/uploads/2015/05/2016_IEEECommsMag_CPRI.pdf</a></li><li><a href="https://www.sciencedaily.com/releases/2018/04/180405223413.htm">https://www.sciencedaily.com/releases/2018/04/180405223413.htm</a></li><li><a href="https://www.ncbi.nlm.nih.gov/pmc/articles/PMC1457762/">https://www.ncbi.nlm.nih.gov/pmc/articles/PMC1457762/</a></li><li><a href="https://www.ncbi.nlm.nih.gov/pubmed/29080018">https://www.ncbi.nlm.nih.gov/pubmed/29080018</a></li><li><a href="https://www.etsi.org/images/files/ETSIWhitePapers/etsi_wp23_MEC_and_CRAN_ed1_FINAL.pdf">https://www.etsi.org/images/files/ETSIWhitePapers/etsi_wp23_MEC_and_CRAN_ed1_FINAL.pdf</a></li><li><a href="https://www.3gpp.org/ftp/Specs/archive/38_series/38.801/">https://www.3gpp.org/ftp/Specs/archive/38_series/38.801/</a></li><li><a href="https://www.itu.int/dms_pub/itu-t/opb/tut/T-TUT-HOME-2018-PDF-E.pdf">https://www.itu.int/dms_pub/itu-t/opb/tut/T-TUT-HOME-2018-PDF-E.pdf</a></li><li><a href="https://www.darpa.mil/program/radio-frequency-machine-learning-systems">https://www.darpa.mil/program/radio-frequency-machine-learning-systems</a></li></ol><figure><a href="https://usejournal.com/?utm_source=medium.com&amp;utm_medium=noteworthy_blog&amp;utm_campaign=tech&amp;utm_content=guest_post_image"><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*PYxBP_WV7LI6zc-zdUkNqQ.png" /></a></figure><p>📝 Read this story later in <a href="https://usejournal.com/?utm_source=medium.com&amp;utm_medium=noteworthy_blog&amp;utm_campaign=tech&amp;utm_content=guest_post_read_later_text">Journal</a>.</p><p>👩‍💻 Wake up every Sunday morning to the week’s most noteworthy stories in Tech waiting in your inbox. <a href="https://usejournal.com/newsletter/noteworthy-in-tech/?utm_source=medium.com&amp;utm_medium=noteworthy_blog&amp;utm_campaign=tech&amp;utm_content=guest_post_text">Read the Noteworthy in Tech newsletter</a>.</p><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=74b7bcf1e58b" width="1" height="1" alt="">]]></content:encoded>
        </item>
    </channel>
</rss>