<?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 Ilya Kavalerov on Medium]]></title>
        <description><![CDATA[Stories by Ilya Kavalerov on Medium]]></description>
        <link>https://medium.com/@ilyakavalerov?source=rss-5a97b4d337------2</link>
        <image>
            <url>https://cdn-images-1.medium.com/fit/c/150/150/0*dIN-bjfLSs0wi71K</url>
            <title>Stories by Ilya Kavalerov on Medium</title>
            <link>https://medium.com/@ilyakavalerov?source=rss-5a97b4d337------2</link>
        </image>
        <generator>Medium</generator>
        <lastBuildDate>Tue, 17 Nov 2026 04:32:06 GMT</lastBuildDate>
        <atom:link href="https://medium.com/@ilyakavalerov/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[Greek before Homer — How Linear B was cracked]]></title>
            <link>https://medium.com/@ilyakavalerov/greek-before-homer-how-linear-b-was-cracked-8c6fc2e7b193?source=rss-5a97b4d337------2</link>
            <guid isPermaLink="false">https://medium.com/p/8c6fc2e7b193</guid>
            <category><![CDATA[ancient-history]]></category>
            <category><![CDATA[cryptography]]></category>
            <category><![CDATA[language]]></category>
            <category><![CDATA[greek]]></category>
            <category><![CDATA[greece]]></category>
            <dc:creator><![CDATA[Ilya Kavalerov]]></dc:creator>
            <pubDate>Tue, 11 Aug 2020 15:40:27 GMT</pubDate>
            <atom:updated>2020-08-11T17:06:48.780Z</atom:updated>
            <content:encoded><![CDATA[<h3>Greek before Homer — How Linear B was cracked</h3><p>(A summary of “The Riddle of the Labyrinth” by Margalit Fox)</p><p>Greek 500 years before Homer! And Homer’s tales were told orally for 400 years before being written! It took until after WWII, 50 years after the language was dug up, to be able to read it. Today Linear B is in unicode, and downloadable as a <a href="http://xahlee.info/comp/unicode_linear_b.html">font</a>, a pdf of pronunciations is <a href="https://www.unicode.org/charts/PDF/U10000.pdf">here</a>.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*cI0XpS_G-1w1GeIpDdbD7A.jpeg" /><figcaption>Linear B for: i-ri-a-de; o-di-se, i.e. Iliad; Odyssey (my guess spelling, and the semicolon is modern of course).</figcaption></figure><p>Given a collection of alien symbols, sometimes appearing consecutively as words, how can you read it? In the case of Linear B these symbols were on clay tablets dug up at Knossos in Crete at the turn of the century, dated to 1400 BC.</p><h3>Structure of the language by Alice Kober</h3><p>The first task was to make a table of all unique single characters. This led to the hypothesis of a syllabary (each symbol is a vowel or a consonant-vowel combination) from the number of characters present. The language had too few symbols to be logographic like Chinese, and too many to be alphabetic like English. Working in the mid-century, Alice Kober then made tables of words with the same first character, and tables with the same last character, and sorted each by prefix and suffix. This allowed her to see the word roots and suffixes of the language. Simultaneously Kober saw evidence of an inflected language. Most simply it was observed that the symbol for “total”, known from context of the lists, appeared to have gender (for example for counts of male and female horses). Kober grouped some words into tables where the columns were words with the same prefix of 2–3 symbols and the rows were sorted by the suffixes of 1–2 letters (referred to as Kober’s triplets). She saw evidence of 3 unknown cases, and bridging characters (characters that were neither roots nor suffixes). This suggesting some symbols like the bridges shared consonants or vowels, so she made tables of consonants vs vowels (CV). Her abstract CV tables showed lines of symbols that started with the same consonant but varied in the vowel along rows. She did all this without ever making an assumption of how anything was pronounced, or a link to another language.</p><p>We are left to wonder how much further Alice Kober would have gotten with her clean fundamental approach with few assumptions. How much further could the decipherment have gone without a link to another language? Or would Kober’s next step have been to make such an assumptive leap? We will never know, because Kober died soon after these discoveries while still working furiously on the decipherment.</p><h3>Definitive reading of the language by Michael Ventris</h3><p>Although a similar language Cypriot was known by many people attempting to decipher Linear B, no one had meaningfully applied it until Michael Ventris soon after Kober’s death. Cypriot is a language that was 1000 years newer than Linear B. It has many symbols similar in appearance to the Linear B symbols, and was deciphered in the 1870s. The Crypriot syllabary was used to write Greek after the Hellenization of Cyprus, and a bilingual inscription of Cypriot and Phoenician led to the decipherment of the Cypriot syllabary. The overlapping characters of Cypriot and Linear B gave Ventris hypotheses for the sounds of the symbols of Linear B. Overlapping is an exaggeration: sometimes the characters were merely visually similar.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/370/1*qw3Q_BEPk4l4xx4ubdGYQw.png" /><figcaption>Cypriot na ti syllables ; Linear B na ti syllables.</figcaption></figure><p>The sound correspondence that Ventris thought Cypriot suggested were possible to check with three methods:<br>1. CV grids like those Kober made: symbols along columns had to start with the same consonant pronunciation<br>2. Proper names of places/kings (emphasized in the tablets and repeated many times) had to have a pronunciation of a well known word known in history, like Knossos (the first location where Linear B tablets were discovered) and Minos (King where first tablets were discovered)<br>3. Nouns like “horses”, known from neighboring logograms, had to match the Greek pronunciation</p><p>The most definitive verification was a “tripod” tablet that plotted jugs and cups with various amounts of handles and legs. It had both logograms of the objects with a written description. Plugging in the Cypriot inspired syllabary into this tablet and reading the Greek descriptions of “tripod cauldrons of Cretan workmanship” while looking at an image of a three legged cauldron was the definitive evidence that the language had been cracked.</p><h3>Conclusion</h3><p>The biggest impression from the book is just how much the computer has changed the world. The scientific community of the mid-century deciphering the language as described in the book appears very different from today’s scientific community. Science is much more analytical today, and much of the painstaking work done by hand of the 1950s has been replaced by milliseconds of cycles on a CPU.</p><p>As Fox describes, much of the groundwork for cracking the language was laid by Kober. What distinguished Kober’s approach from her contemporaries is how modern and scientific she was. Many, Ventris included, had the strategy of picking an already known language, and making up a theory of pronounciation until they could read that language in the tablets without too many errors. For Ventris, before deciding the tablets were Greek, he was convinced they were Etruscan but was unable to show it decisively. Other scholars, like Bedřich Hrozný, believed they had already “read” the tablets long before Ventris cracked the code! But their logic convinced few other than themselves. Kober was on the contrary much more modern and analytical, refusing to make assumptions about previous writing systems or pronounciation. In 2020 it is definitely striking to read that other scholars did not share her careful, deliberate, and “from first principles” attitude. Ultimately it is her approach that led to the tooling that verified the Cypriot leap that Ventris tried.</p><p>Another striking detail is how much a computer would have helped! There is a lot to be said for optical character recognition from computer vision, which could have accelerated labeling of the characters. But there was a paucity of data in any case: because of standoffishness and wartime and post-war scarcity, difficult travel, and difficult copy-making, the records of Linear B were scattered in small groups, and there was not a big data repository of it waiting to be read. The biggest amount of work was rather sorting the words after they had been copied. Kober spent many consecutive weeks at a time making notecards of various orderings of symbols all to be able to organize and sort the words. It is incredible to think that after an initial transcription of symbols, any sorting methodology of words would take only milliseconds on a modern $100 netbook, while Kober spent tens of weeks cumulatively doing this with great devotion and care.</p><p>Margalit Fox’s book tells a great story of the individuals who participated in cracking the code, and there is a great many dramas not mentioned here making up the history of the deciphering. Read the book for more details and excellent images and comparisons with other decipherments and codes!</p><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=8c6fc2e7b193" width="1" height="1" alt="">]]></content:encoded>
        </item>
        <item>
            <title><![CDATA[“The Man Who Solved the Market”: And the solution was… HMMs and regression]]></title>
            <link>https://medium.com/@ilyakavalerov/the-man-who-solved-the-market-and-the-solution-was-hmms-and-regression-dd60cea5a6d7?source=rss-5a97b4d337------2</link>
            <guid isPermaLink="false">https://medium.com/p/dd60cea5a6d7</guid>
            <category><![CDATA[stock-market]]></category>
            <category><![CDATA[data-science]]></category>
            <category><![CDATA[finance]]></category>
            <category><![CDATA[trading]]></category>
            <category><![CDATA[machine-learning]]></category>
            <dc:creator><![CDATA[Ilya Kavalerov]]></dc:creator>
            <pubDate>Sun, 05 Jan 2020 03:44:48 GMT</pubDate>
            <atom:updated>2020-01-05T03:44:48.935Z</atom:updated>
            <content:encoded><![CDATA[<p>I just finished reading Gregory Zuckerman’s fantastic biography of Jim Simons: “<a href="https://www.goodreads.com/en/book/show/43889703">The man who solved the market</a>.” It covers some of the many unique characters at RenTech up to the present, and has the most to say about their 1980s–2000s trading strategies. The best answers I was able to get from the book on what their solution to the market was:</p><ol><li>Markov chains, HMM models</li><li>Linear factor models, followed by high dimensional kernel regression methods</li><li>Big monolithic models and lots of data</li><li>A well orchestrated engineering effort</li></ol><p>The history in the book can be summed up in the following chart, which I made by grep-ing the ebook for years:</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*xrIG0J4s-J-5A_MBpYh6_g.png" /><figcaption>Year v.s. the cumulative count of years in the main text (before the politics discussion starts midway though Chap 14). The steeper the line the more mentions that time period had by the author. The vertical markers from left to right show: Baum is immersed in traditional trading. Ax moves west as Axcom and hires Carmona. Berlekamp takes over Axcom/Nova. Laufer’s programs achieve great successes for Nova. Brown &amp; Mercer’s stock trading team earns 2x the profit of the bond, commodity, currency team.</figcaption></figure><p>The steeper the line the more the author mentioned that time period. The majority of the history is about the bond, commodity, currency team (started as Axcom with Ax’s Markov models, taken over by Berlekamp, improved by Laufer, and eventually taken over by Brown &amp; Mercer). Within that steep part is the development of the stock team (which uses Frey’s factor models). The flatter bits like after 2010: not much is written about the fund. And the earlier years with Baum before Ax are dominated by traditional trading.</p><p>The book is highly engaging, and contains a lot more than what I highlight here. A lot of the strategies mentioned in the book can be divided into trending and mean reverting strategies: when is the price rise/dip just the beginning, and when is the price about to bounce back to what it should be according to historical factors. But I wouldn’t say that it seems like there is one big idea behind the fund. Its success appears to be a result of a huge software engineering endeavor that processes a lot of data, looks at a lot of correlations, and makes a lot of trades. In this regime, you only need to be right slightly more than half the time to make money, as the book’s characters say.</p><h3>Early traditional days</h3><p>One of the first employees of Jim Simons (JS) was Leonard Baum, co-inventor of the well known Baum-Welch algorithm for estimating the transition matrix in HMMs with EM. In an early paper “<a href="https://web.archive.org/web/20131005001241/http://projecteuclid.org/DPubS?verb=Display&amp;version=1.0&amp;service=UI&amp;handle=euclid.aoms%2F1177697196&amp;page=record">A Maximization Technique Occurring in the Statistical Analysis of Probabilistic Functions of Markov Chains</a>” that introduces this idea you can find JS as a reference for the never published “Probabilistic models for stock market behavior” which is cited as an application. Interestingly, Baum preferred trading traditionally with JS and not much was done quantitatively.</p><h3>Quantitative Bonds, commodities, and currencies team</h3><p>James Ax, a number theorist and strong believer in Markov Chains for financial markets was JS’s first quant, and headed the effort under the name Axcom in CA. By ’86 Axcom traded 21 different futures contracts. René Carmona then joined them to try to incorporate SDEs, his field of expertise. When that failed, Carmona suggested changing their existing linear regression approaches to be nonlinear, high dimensional kernel regression methods, and having the model directly suggest buy/sell orders. These improved results on trending models.</p><p>Elwyn Berlekamp (who had worked with Kelly) took over Axcom in ’89 and brings in Henry Laufer in ‘92, they worked on mean reversion strategies and looking at correlations between time periods. They start using data in 5 minute bars, used pairs trading, and had an online learner, constantly looking for trading signals (one auspicious one they called “Henry’s signal”). By ’97 “more than half of the trading signals Simons’s team was discovering were nonintuitive” and they ignored them. But they figured out the right day and time to make their trades. Later Simons says “we’re the best at estimating the cost of a trade.” In ’98 they are the majority of the firm, with stock trading only 10% of the profits.</p><h3>Quantitative Stock team</h3><p>This team starts to make progress in ’93 when Nick Patterson contacts Brown and Mercer at IBM’s speech group (which uses algorithms like Baum-Welch in their HMMs for speech-to-text). Brown and Mercer take over Robert Frey’s factor stock trading fund named Kepler (later Nova). Frey (previously statistical arbitrage at Morgan Stanley) was identifying various independent variables for factor trading models. Brown &amp; Mercer retain Frey’s model, and elaborate it to cover real-life technicalities he was ignoring. They make an adaptive single trading system for their whole portfolio, and self-correcting for when the trades it suggests are unexecutable. The system repeated on loop several times an hour. It was a well engineered product, and usually bet on mean reversion strategies. By ’03 their profits are 2x Laufer’s team, and they work on a model to replace the futures team’s. Alexey Kononenko rises through the ranks of this team. Two members Belopolsky &amp; Volfbeyn bolt to Israel Englander’s Millennium management, allegedly with (millions of lines of) code and ideas, and have great success there (“some of the most successful traders Englander had encountered”). By 2010 the system is huge, executing thousands of simultaneous trades throughout the day, with lots of factors and interrelations.</p><h3>The models and the engineering effort</h3><p>Both HMMs and factor models are famous by now, and the author Gregory Zuckerman has good evidence that the secretive fund used these techniques. RenTech did a lot of hiring from the IBM speech team. Speech recognition relied heavily on HMMs until 2010, when Deep Learning started to overtake the field. Frey, who left to teach at Stony Brook, has a <a href="http://www.ams.sunysb.edu/~frey/QuantitativeFinance/Resources/TutorialsWorkshops/index.html">tutorial on HMMs for finance online</a>. Carmona has a book “<a href="https://books.google.com/books?id=IoS8BAAAQBAJ&amp;dq=rene+carmona+kernel&amp;source=gbs_navlinks_s">Statistical Analysis of Financial Data in R</a>” which includes chapters on using kernel regression for financial data.</p><p>A big benefit of hiring from IBM’s speech team was getting people experienced not just in HMMs but also in large software engineering projects. Mercer and Brown showed that Frey’s factor model idea was fundamentally good, and once the implementation carefully incorporated all the elaborate characteristics of real markets the strategy went from dud to stud.</p><p>HMMs, factor models, non-linear kernel regression, correlations between various time periods, these are all separately well known pieces. But the book says many times that a single model was handing out trading instructions. My first guess for how to put these different elements together would be that everything else was generating features for the HMM.</p><h3>Conclusion</h3><p><a href="https://www.goodreads.com/en/book/show/43889703">Gregory Zuckerman’s exciting book</a> contains much more than the trading strategies of RenTech. The stories of each of the characters and the dramas behind the fund are page-turners. As the subtitle of the book says, Jim Simons certainly did launch the quant revolution. And RenTech is having its best years right now! $63B of its $104B trading profits are since 2010, and with just 300 employees. Surely it has updated its techniques, and probably cashed in on the deep learning revolution raging since (at least) 2014!</p><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=dd60cea5a6d7" width="1" height="1" alt="">]]></content:encoded>
        </item>
        <item>
            <title><![CDATA[What’s inside the Google Coral Edge TPU? Speed Test & Teardown]]></title>
            <link>https://medium.com/@ilyakavalerov/whats-inside-the-google-coral-edge-tpu-speed-test-teardown-8673b440141d?source=rss-5a97b4d337------2</link>
            <guid isPermaLink="false">https://medium.com/p/8673b440141d</guid>
            <category><![CDATA[tpu]]></category>
            <category><![CDATA[machine-learning]]></category>
            <category><![CDATA[deep-learning]]></category>
            <category><![CDATA[accelerated-learning]]></category>
            <category><![CDATA[google]]></category>
            <dc:creator><![CDATA[Ilya Kavalerov]]></dc:creator>
            <pubDate>Tue, 17 Dec 2019 05:08:55 GMT</pubDate>
            <atom:updated>2019-12-17T07:36:38.992Z</atom:updated>
            <content:encoded><![CDATA[<p>Earlier this year Google finally released TPU hardware that you can own via their Coral brand. However these are not the beefcake cloud TPUs training networks like BigGAN at 100+ petaflop/s for a week or even the cheapest 180 TFlop/s v2 TPU that you can rent at $4.50/hour on demand. These are TPU devices meant to work “<a href="https://medium.com/tensorflow/build-ai-that-works-offline-with-coral-dev-board-edge-tpu-and-tensorflow-lite-7074a9fd0172">at the edge</a>,” that is to deliver deep learning solutions in the field on small devices without internet. So how good are the TPUs you can own?</p><p>The specs of all the <a href="https://coral.ai/products">current TPU products</a> you can buy is 4 TOPs. The units aren’t the usual 32 bit floating point because the <a href="https://www.tensorflow.org/lite">tensorflow-lite</a> that runs on these devices uses 8 bit fixed point. You also won’t find memory bandwidth like on cloud TPUs or GPUs at 600+ GB/s. Edge TPUs are connected via USB 3.0 or a single mPCIe lane (gen 2) so 640 or 500 MB/s. Of course, since there is only <a href="https://coral.ai/docs/edgetpu/compiler/#parameter-data-caching">8MB of SRAM</a> on the edge TPU this means at most 16ms are spent transferring a model to device, and in the model used in this post, it took just 10ms. You won’t miss the memory because these devices are not for from-scratch training. But you<a href="https://coral.ai/docs/edgetpu/models-intro/#transfer-learning-on-device"> can train the last layer of a model</a> on this device! The main usecase of Edge TPUs is inference, which is how I will be benchmarking.</p><p>The most self contained<a href="https://coral.ai/products/accelerator/"> USB3.0 “Coral Edge TPU” device</a> came out in March of this year, and this is what I got to play around with. It’s one of Coral’s <a href="https://coral.ai/products/#prototyping-products/">prototyping products</a>, though on paper it has the same exact performance as their mPCIe <a href="https://coral.ai/products/#production-products/">production boards</a>. USB 3.0 is easy, but if I didn’t want to deal with the external dongle, the <a href="https://coral.ai/docs/m2/datasheet/">M.2 Accelerator A+E key</a> could replace my wifi card internally:</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*WE4Dt44XAlCAPjT7_x_UgA.jpeg" /><figcaption>Left: My Wifi card in the laptop. Right the M.2 A+E accelerator that could slot right into the motherboard.</figcaption></figure><p>So what’s inside the $75 30mm x 65 mm TPU Edge Accelerator? Could it be the $35 30mm x 22mm mPCIe Accelerator with a USB adapter and a heat sink? Can we open the plastic + metal snap safely?</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*_2f-TTBqGWwxsm7BYSlOlg.jpeg" /><figcaption>Front and back of the Edge TPU</figcaption></figure><p>The plastic shell can be pried off without too much struggle, thanks to the heat sink being very stiff. And the heat sink can be removed by undoing 4 screws.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*fciKiUnZJ0vcuC0iX0B5xQ.jpeg" /><figcaption>Back of the device with the plastic cover off, with and without heatsink, and heatsink by itself.</figcaption></figure><p>The answer to the first question is no, not literally, but the boards are very similar:</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/991/1*ujhL2e8RGF5g2pelXy2Hrg.jpeg" /><figcaption>Business side of the Coral Edge TPU. Right is the M2 A+E accelerator and the sole TPU chip on a penny.</figcaption></figure><p>The two chips in contact with the heatsink are probably the TPU and the memory, with the larger one being the TPU. After I put the Edge TPU back together I did a little speed testing of my own, though <a href="https://coral.ai/docs/edgetpu/benchmarks/">Google’s benchmarks</a> say between 10 and 20 x speedup over a CPU.</p><p>The <a href="https://github.com/google-coral/tflite/tree/master/python/examples/detection">Coral example for detection</a> can be run on CPU and TPU alike. The computer that I’m attaching the TPU to is my Acer Chromebook 11 running <a href="https://galliumos.org/">Galium OS 2.1</a>, which is enough like debian so that there are no problems installing tf-lite and the edge TPU runtime. The MobileNetV2 SSDLite models for detection are less than 7MB and operate on 300x300 images. On the TPU a single image runs in 20ms (+10ms model copy time on the first iteration). On my laptop’s <a href="https://www.cpubenchmark.net/cpu.php?cpu=Intel+Celeron+N2840+%40+2.16GHz&amp;id=2388">Intel Celeron 2.16GHz CPU</a> from 2014 a single image is 1500ms. On another PC with <a href="https://www.cpubenchmark.net/cpu.php?cpu=Intel+Xeon+E5-2670+v2+%40+2.50GHz&amp;id=2152">Intel Xeon 2.5 GHz CPU</a> (which cpubenchmark.net predicts is 15x faster) a single image is 130ms (at 210 watts!). The best CPU available would be 3x faster than this according to cpubenchmark.net: still slower than the edge TPU which is much cheaper in $$$ and power!</p><p>A little more exciting I think is<a href="https://github.com/ilyakava/tflite/tree/master/python/examples/detection"> detection on the live streaming camera data</a> on my laptop. On the Celeron, this example runs at less than 1 frame per second. On the TPU, it runs at 20 frames per second: realtime! And only a quarter of the time is spent on the inference on TPU, the rest is resizing images and drawing on the output which must be done on the CPU. Though I did install the max operating frequency TPU runtime, I’m not near the operating limit of the TPU, so heat is no problem: the TPU’s heatsink doesn’t burn to the touch.</p><p><a href="https://drive.google.com/open?id=14FyTKOj7FBdzOs6PQ3dt2Hjzeo42s4wf">street4.avi</a></p><p>These examples were easy to use/tweak after just a few minutes of installation on linux. All you need to get going with this device is to <a href="https://coral.ai/docs/accelerator/get-started/">apt-get install</a> the TPU runtime, pip install tf-lite, and then you can jump straight into the <a href="https://github.com/google-coral/tflite/tree/master/python/examples">examples</a>. The models in these examples were pre-compiled. A truer speed test would be to train a model from scratch, execute that on the GPU and CPU, then convert it to tf-lite, and execute that on CPU and TPU.</p><p>Overall I’m excited about this device. I think it is amazing that a 300x300 object detection network runs on a netbook from 2014 in realtime with a $75 modification ($35 if I got the M.2 board). Sure, the edge TPU won’t replace any GPUs for training, or even for model evaluation. But I see a lot of potential in the <a href="https://coral.ai/products/m2-accelerator-ae/">10 USD/TOP, 0.5 watt/TOP board</a>. This is at least 10x cheaper per OP than a GPU! And much less power hungry. For traveling demos this would be great. As long as the operations you’re interested in are in this <a href="https://coral.ai/docs/edgetpu/models-intro/#supported-operations">table</a>, it could be a useful chip to program for when low power and lightweight systems are a priority.</p><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=8673b440141d" width="1" height="1" alt="">]]></content:encoded>
        </item>
    </channel>
</rss>