<?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 Ayesha Bajwa on Medium]]></title>
        <description><![CDATA[Stories by Ayesha Bajwa on Medium]]></description>
        <link>https://medium.com/@ayesharbajwa?source=rss-3902f5af4e8a------2</link>
        <image>
            <url>https://cdn-images-1.medium.com/fit/c/150/150/1*aSfEilW6oNUlxod57k5XMg.jpeg</url>
            <title>Stories by Ayesha Bajwa on Medium</title>
            <link>https://medium.com/@ayesharbajwa?source=rss-3902f5af4e8a------2</link>
        </image>
        <generator>Medium</generator>
        <lastBuildDate>Tue, 17 Nov 2026 15:11:32 GMT</lastBuildDate>
        <atom:link href="https://medium.com/@ayesharbajwa/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[Reflections on Cuba, and America]]></title>
            <link>https://medium.com/@ayesharbajwa/reflections-on-cuba-and-america-60fbf39b996b?source=rss-3902f5af4e8a------2</link>
            <guid isPermaLink="false">https://medium.com/p/60fbf39b996b</guid>
            <category><![CDATA[us-cuba-relations]]></category>
            <category><![CDATA[imperialism]]></category>
            <category><![CDATA[socialism]]></category>
            <category><![CDATA[cuba]]></category>
            <category><![CDATA[nationalism]]></category>
            <dc:creator><![CDATA[Ayesha Bajwa]]></dc:creator>
            <pubDate>Sat, 09 Mar 2019 03:40:29 GMT</pubDate>
            <atom:updated>2019-03-09T22:45:07.321Z</atom:updated>
            <content:encoded><![CDATA[<figure><img alt="" src="https://cdn-images-1.medium.com/max/489/1*GTMwqdqBm6FvNrOpm4wn1Q.jpeg" /></figure><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*zm81q4mnCVx1asOqIpBhGA.jpeg" /></figure><p><em>A recent trip to Cuba gave me some perspectives on American history and politics. In writing this, my main conclusion has been that there’s always another layer of historical nuance to peel back, another qualifier that brings us closer to the truth of what happened and why. I’ve had to be more cursory than I’d like, as a result! All photographs are my own.</em></p><p>When I was in high school, I somewhat incessantly drew connections to colonialism in history classes. I wasn’t trying to be obnoxious — it was simply strange to me that my classmates spoke mainly of colonialism with respect to the thirteen colonies and the Boston Tea Party, as if it were a relic from the time of hoop skirts and tricorn hats, an outdated concept irrelevant to our current world. I was acutely aware that Indian independence happened in 1947, recently enough that the generation of my own parents’ parents could tell me about it firsthand. I knew that the 20th century military struggles were not just World Wars and Cold War proxy wars, but also successful independence movements in Africa and Asia, harbingers of the demise of the British and other European empires nearly two centuries after our own Declaration of Independence. I had also read that the rapid economic rise of India and China, in this new age of globalization, had been inevitable and simply delayed some decades by colonizers [1]. Surely colonialism was the most important force dictating recent world history!</p><p>I understood America not as an old-school colonizer but as a militarily powerful empire that used limited direct force, such as the CIA, to protect a central economic and political interest: preventing the spread of communism. The language of political ideas we used in <a href="https://en.wikipedia.org/wiki/AP_United_States_History">APUSH</a> sounded distinct from that used to describe the older empires: we called it <em>manifest destiny</em> instead of conquest, <em>American exceptionalism</em> instead of superiority, <em>spheres of influence</em> and <a href="https://en.wikipedia.org/wiki/Open_Door_Policy">Open Door</a> instead of access and control, the <em>Monroe Doctrine</em> instead of dictating future colonization of the Americas. In retrospect, these differences were superficial, merely reflecting geopolitics before WWII and American hegemony — back when we needed to tread lightly. This was not a history consistent with the ideals of freedom and opportunity that immigrant families like mine seek in America. Yet, I couldn’t deny that I, <a href="https://www.nytimes.com/2018/12/21/opinion/sunday/immigration-border-policy-citizenship.html">an American citizen by birth</a>, was in many ways a beneficiary of our history as a global bully.</p><p>As a hegemon, we modern Americans have the privilege of remaining ignorant of many of our nation’s past actions abroad, since the consequences may not surface in our daily lives. The fascinating history of Cuba, the island nation of 12 million approximately 100 miles off the Florida coast, provides an interesting lens by which to view our country.</p><p>Cuban independence, according to current Cuban history, was won in 1959. After a series of unsuccessful independence struggles against the Spanish, it was considered achieved after the defeat of <a href="https://www.globalresearch.ca/cuba-pre-1959-the-rise-and-fall-of-a-u-s-backed-dictator-with-links-to-the-mob/5464738">US-backed dictator Fulgencio Batista</a>, who had suspended the 1940 Constitution and allowed inequality to increase, with most of the prominent sugarcane industry being owned by Americans. Batista seized power through a 1952 coup, sensing he would lose an upcoming election to Fidel Castro. After years of declining popularity and increasing support for Castro’s rebel movement, <em>Movimiento 26 de Julio</em>, even after its disastrously unsuccessful attempt to seize the barracks in Santiago de Cuba in 1953, Batista finally lost US backing in December 1958 and fled to the Dominican Republic. The rebels came out of hiding in the Sierra Maestra mountains by Santiago, and Fidel Castro proclaimed victory from the town hall balcony in the main square on New Year’s Day, 1959.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*QJr-G0PIOPi-qTXKL-Un2w.jpeg" /><figcaption>Carlos Manuel de Céspedes Park, Santiago de Cuba. On the right, the town hall balcony Fidel Castro used to announce the success of the Cuban Revolution in 1959.</figcaption></figure><p>Unlike <a href="https://archive.nytimes.com/www.nytimes.com/library/world/mideast/041600iran-coup-timeline.html?mcubz=3">Iran</a> or <a href="https://www.nytimes.com/2017/10/14/world/americas/chile-coup-cia-museum.html">Chile</a>, Cuba maintained independence from American influence. With the rise of socialist thought and the US unable to install another puppet dictator after the failure of Batista, Cuba fell into the Soviet camp. It didn’t happen immediately. Castro was a larger-than-life figure who received a warm welcome from the people when he <a href="https://www.history.com/news/fidel-castros-wild-new-york-visit-55-years-ago">visited New York</a> in the spring of 1959. He returned in January 1961 for the UN General Assembly and opted to stay in Harlem, among a majority black population at a hotel popular among black politicians. He met with a variety of prominent political and cultural figures, including Malcolm X, Nikita Khrushchev, Langston Hughes, and Allen Ginsberg. At the UN, he gave a four and a half hour speech condemning US foreign policy in Latin America and Africa and explicitly aligned Cuba with the Soviets and Khrushchev. Cuba’s aware and deliberate resistance to becoming another one of our <a href="https://en.wikipedia.org/wiki/Banana_republic">banana republics</a> [2] or puppet nations, and in particular the nationalization of formerly private, often foreign-dominated industries, meant that the US and Cuba were both politically and economically at odds. The export <a href="https://en.wikipedia.org/wiki/United_States_embargo_against_Cuba">embargo</a> began in 1960 after the oil refineries were nationalized and the US cut off diplomatic relations in 1961. The actions of the American government towards Cuba represented the views of the individuals in power, but Castro’s trips to New York remind us that he had ideological sympathizers in the US.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*nQRHma0_p5PZxhXqDZR_oQ.jpeg" /></figure><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*kdG0stjJxxcv5UBnjer2NA.jpeg" /></figure><p>Today, nearly every Cuban town seems to have signboards with government-approved slogans or quotes from Fidel Castro, Raúl Castro, Ernesto “Che” Guevara, or Camilo Cienfuegos. They are certainly nationalistic and perhaps propagandistic, often mentioning <em>lucha</em> (struggle) or <em>trabajo</em> (work), and may cause some Americans to triumphantly exclaim, “Communism!”</p><p>What both the signs and the American seem to miss is how many core socialist ideas grew out of a dream of social justice, a desire to reduce inequality, and a forceful anticolonial sentiment. We don’t have to accept or endorse the political philosophies of figures like Fidel Castro and Che Guevara for us to consider their ideas as a counterpoint, to contrast our own weaknesses as a nation, or to illuminate the darker parts of our history. Any story about the past is inherently political from the perspective of the facts and narratives included and those left out. (I mean, good luck getting 100% representation).</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*4Bkhk48_kpDGOq48_XMZjg.jpeg" /></figure><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*fSIYMa6JzkLkw-I2FzLbug.jpeg" /></figure><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*dP59elCRZOnmz4Lp3G7RQg.jpeg" /></figure><figure><img alt="" src="https://cdn-images-1.medium.com/max/488/1*DFSFOz0ZfuZREMLKCyNsMA.jpeg" /><figcaption>Nationalistic imagery featuring Fidel Castro and Che Guevara, leaders of the Cuban Revolution (Habana, Santiago de Cuba, Camagüey).</figcaption></figure><p>To understand modern Cuba, we should consider the history of Spanish colonization and influence on the island, beginning with Columbus in 1492 and the immediate demise of native populations from disease and conquest. For three centuries, Cuba remained pastoralist with a modest economy and served as a stopping point for Spanish ships. In the 1800s, Cuba became a major sugar producer. The slave trade increased as large plantations became the norm, and a growing upper class remained loyal to the Spanish crown. Taxation and Creole-Spanish conflict led to attempted independence wars in 1868 and 1895 (when Martí was killed) which, along with US victory in the Spanish-American war, weakened Spain’s grip on the island. American influence increased even as Cuba established nationhood in 1902. Evidence of this history includes military forts in port cities, the remains of slave-labor plantations, the railroads built specifically for sugarcane transport, the magnificent houses of an affluent ruling class (many of whom fled in 1959), the population of Afro-Cubans who have ancestors who were brought to the Caribbean as slaves, Catholicism and the <a href="https://en.wikipedia.org/wiki/Yoruba_people">Yoruba religion</a> of Nigerian origin, and the Spanish language.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*4rTkY_XODMapb5S5qTEwOw.jpeg" /></figure><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*EPFQjaMvZ42Cdja9a9wa-w.jpeg" /></figure><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*-xO4w5wiLLSAAcwLAUl22g.jpeg" /></figure><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*6pT4uXOufUbIs0uKWWI2Bg.jpeg" /></figure><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*46q4cMyeqdbfBQv1Eq9LKQ.jpeg" /></figure><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*cVudkBjw_9FeVB_1N91EVg.jpeg" /><figcaption>Spanish influences in Cuba, in the forms of military constructions like forts and magazines, a tower formerly used to monitor slaves in the sugarcane fields, architecture and interior decoration, Catholicism, and flamenco.</figcaption></figure><p>Spanish colonization and American influence did shape a unifying nationalism that paved the way for the 1959 revolution, but it also created a political environment critical of capitalism and sensitive to issues of social justice, egalitarianism, and equal opportunity. As was the case in many post-colonial nations [3], Cuban socialism grew out of the struggle for independence and the rejection of colonial norms. Che Guevara preserved a popular account of his personal political awakening in <a href="https://en.wikipedia.org/wiki/The_Motorcycle_Diaries_(book)">Motorcycle Diaries</a>, a memoir of his travels throughout Latin America that details the desperate social and economic inequality which inspired the young Argentine doctor’s political ideas.</p><p>(A side note to Americans: It is interesting to note that right-wingers like <a href="https://www.csmonitor.com/World/Africa/Africa-Monitor/2010/0916/Newt-Gingrich-dissertation-on-Congo-sheds-light-on-his-jab-that-Obama-is-anticolonial">Newt Gingrich</a>, <a href="https://www.thecrimson.com/column/nelsons-rum/article/2015/2/27/anticolonialism-dirty-word-giuliani/">Rudy Giuliani</a>, and <a href="https://www.npr.org/2012/09/01/160389675/right-wing-filmmaker-obamas-an-anti-colonialist">Dinesh D’Souza</a> have repeatedly attacked Barack Obama for being an “anti-colonialist.” Why would Americans speak ill of anticolonialism? Weren’t we once a colony? Well, yes, but those declaring independence from Britain were themselves successful colonizers of indigenous lands. These days, institutions such as Harvard still take academics such as <a href="https://www.lrb.co.uk/v33/n21/pankaj-mishra/watch-this-man">Niall Ferguson</a> seriously, and MIT <a href="https://thetech.com/2019/02/14/celebrating-war-criminals">invites Henry Kissinger</a> to campus. Mainstream politicians invoke the founding fathers but ignore the fact that we were formerly ruled by the British and had a war over it, not to mention a centuries-long commitment to more glaring evils like slavery — here’s <a href="https://2jc18v1irh0441xjkl3blrt4-wpengine.netdna-ssl.com/wp-content/uploads/2014/11/Slavery-Continuum.jpg">a reminder</a> of how deep that one goes. Call me crazy for thinking racism a likely factor in this selective history. Surely neocolonialists will Make America Great Again, even if Obama <em>is</em> black and doesn’t think our country should rule the world!)</p><p>In 1959, Cuba began a transformation from a third world to a <a href="https://www.nationsonline.org/oneworld/third_world_countries.htm">second world</a> country. Almost immediately, a <a href="https://en.wikipedia.org/wiki/Cuban_Literacy_Campaign">successful literacy movement</a> engaged an estimated million Cubans and raised the literacy rate to 96%. In the years that followed, Cuba nationalized its industries, established socialized healthcare and ration ships, and exerted its influence abroad in the struggle against empire. In the post-WWII and Cold War era, this meant siding with Soviet-backed independence movements and countering American-backed ones (recall that the British and French colonial empires had been on the Allied side). The most notable participation in a liberation war was Cuba’s presence in Angola from 1975 through 1991. Castro’s government made the argument that for Afro-Cuban soldiers, who knew their ancestors had been brought across the Atlantic on slave ships, the fight could be personal as well as ideological; however, <a href="https://www.historynet.com/cuban-fighters-angolas-civil-war.htm">race relations in Cuba weren’t perfect and the war in Angola had high political stakes</a>. Cuban forces were also involved in the Nicaraguan Revolution. In both Angola and Nicaragua, Cuba aligned itself with leftist or Marxist movements, while the US aligned with rightist counterrevolutionaries.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/884/1*HGUdaoKqjiyYkOz8mBnvcg.jpeg" /><figcaption>A museum display about the <a href="https://en.wikipedia.org/wiki/Cuban_Literacy_Campaign">Cuban Literacy Campaign</a>.</figcaption></figure><p>Cuba’s implementation of socialism did not create utopia. A planned economy could be difficult to maintain. The government restricted Cubans from traveling outside the country by boat and did not allow foreign media. The revolution created both political and economic refugees in the decades following. The fall of the USSR and the loss of Soviet economic subsidies in the early 1990s marked the beginning of Cuba’s “special period.” During this time, the Cuban government imposed extreme austerity measures and staunchly refused to reform the economy or political structure. Many Cubans attempted to leave to other countries for better economic circumstances, and illegal migration to the US peaked.</p><p>Cuba today remains a poor country by most economic standards, despite high accessibility to free healthcare and education. Imports are still difficult because of the US embargo. It is also somewhat incredible that it maintains the practice of exporting doctors and nurses to other countries — though it is sometimes <a href="https://www.theguardian.com/global-development/2018/nov/23/brazil-fears-it-cant-fill-abrupt-vacancies-after-cuban-doctors-withdraw">met with political resistance, as in Brazil today</a>. It’s a complex arrangement: the positions are usually paid for by the host government negotiating a deal with the Cuban government, which keeps a majority of the pay, so the program is a major source of income for Cuba (in fact, medical staff are <a href="http://time.com/5467742/cuba-doctors-export-brazil/">Cuba’s main export</a>). Even so, these doctors and nurses often earn more than they would at home and play a remarkable humanitarian role in improving healthcare access. Tourism, though it currently brings in less revenue than the export of medical staff, is another large source of income. There are other economic norms in Cuba that will seem unfair to most Americans; for example, even private cacao farms must simply give a large percentage of their harvest to the government, which uses different agricultural processes from the smaller farms.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*iF8mCdWopSCD9KBTwsO3ZQ.jpeg" /><figcaption>A national pharmacy (for Cubans only) in Baracoa.</figcaption></figure><p>Nicholas Kristof <a href="https://www.nytimes.com/2019/01/23/opinion/cuba-embargo.html">writes</a>, “Cuba is neither the demonic tyranny conjured by some conservatives nor the heroic worker paradise romanticized by some on the left. It’s simply a tired little country, no threat to anyone, with impressive health care and education but a repressive police state and a dysfunctional economy.” Though I can’t comment on the role of the police, I saw evidence of the other points during my visit, particularly accessibility to healthcare and rations, a difficulty with importing everyday goods, and a lack of transportation infrastructure. I’m also aware of the brain drain problem, and the fact that many Cuban students have to look beyond the island for further educational and economic opportunities — the spotty Internet infrastructure may have a lot to do with it. This need to seek education elsewhere seems especially true for <a href="https://www.miamiherald.com/news/local/education/article220200665.html">STEM students</a> who want to go beyond theory and into application. So, while there is a history of strong social programs and educational access for the population in general, those who excel in their fields might find opportunities lacking. If a doctor’s pay is low, is that person’s value to society being deflated?</p><p>Cuban history at the time of Fidel Castro’s death in 2016 seems to indicate that the Revolution was neither totally successful nor completely unsuccessful. We’re likely to see further opening up of Cuba under the government of Miguel Díaz-Canel, who transitioned into power in 2018 after Raúl Castro stepped down. I personally hope this opening up includes better Internet access for regular Cuban students.</p><p>What should we make of Cuban history, in contrast to our own? While America made the capitalist bet and has been economically successful by some metrics, many would say we’ve completely failed at exporting democracy — consider our recent involvement in the Middle East — at a high military cost. One could argue that spreading American ideals was always a shoddy cover for maintaining short-term economic interests and quashing dissent. Our geopolitical alliances often favor countries that lack stated American values, such as freedom of speech or gender equality, over countries where progressive social movements have been successful. Throughout our relatively short history as an independent nation, we have dealt unsuccessfully with dictators. And through our involvement in military coups or interventions gone awry, we have often given more power to the factions within other countries that view America as evil. It was, to expand an earlier point, <a href="https://en.wikipedia.org/wiki/1953_Iranian_coup_d%27%C3%A9tat">our 1953 intervention</a> that toppled a progressive democracy in Iran and gave credibility to the religious conservatives who ultimately took over.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*EkkeXQi8iP3VvZvR6RmNsw.jpeg" /></figure><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*X73KzdCP_3yR-IWRR7kP8g.jpeg" /></figure><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*ONk5CrE6jGJ6kXkBkJCUzw.jpeg" /><figcaption>Cacao, along with coffee and coconut, is a large industry in Eastern Cuba. These regions are particularly susceptible to hurricanes and tropical storms, and are likely to face more hardship due to climate change.</figcaption></figure><p>Cuba, a country of 12 million, appears to have integrated cultural influences more than our nation of 300 million has, intentionally or not. Visitors in search of Cuban culture witness a mixture of Spanish, African, Afro-Caribbean, Soviet, and American influences, in addition to more recent globalization. Classic cars with new engines pass by houses with Spanish colonial facades and concrete block Soviet apartments. Both Christianity and the Yoruba religion are prevalent, and the <em>Casa de Cultura </em>and<em> Casa de Trova </em>in most towns showcase a wide variety of music and dance, including son, salsa, rumba, and jazz. We saw shiny new BMW and Mercedes Benz cars in seemingly rural places, and our tour bus was brought over from China on a container ship. Also, the national rum is cheap and has the same price everywhere.</p><p>American culture, to many of us, consists of disjoint pockets of varying cultural histories, religious inclinations, societal values, and political leanings. An oversimplification of these independent subcultures leads us to the rural-metropolitan “two Americas” divide that became so discussed after the 2016 election [4]. My friends who grew up in one kind of America and then migrated to another for school or work say the contrast can be jarring. It’s hard to name any single aspect of American culture that is truly ubiquitous. There’s pop culture now, sure, but is there a <em>deep </em>American cultural tradition?</p><p>If our culture is instead based on American values, we should note inherent contradictions and be wary of hypocrisy. It’s common to emphasize gratitude towards those who serve in the military, though we <a href="https://www.rollingstone.com/politics/politics-news/the-shameful-way-america-treats-its-veterans-52825/">could better support benefits for veterans</a>. In contrast to the strong historic symbolism of the American factory worker, individual protections are collateral damage to corporate aspirations <a href="https://fivethirtyeight.com/features/americans-dont-miss-manufacturing-they-miss-unions/">as unions decline</a>, and our longtime manufacturing advantage is shrinking with new technologies abroad. US universities are acclaimed, but even progressive states like California have <a href="http://www.dailycal.org/2018/08/08/2018-uc-accountability-report-shows-high-enrollment-despite-low-state-funding/">decreased public funds for state schools</a> as <a href="http://www.dailycal.org/2014/12/22/history-uc-tuition-since-1868/">tuition rises</a>. The quality of and support for primary and secondary education <a href="https://www.theatlantic.com/business/archive/2016/08/property-taxes-and-unequal-schools/497333/">varies hugely across regions</a>. Despite our stated economic success, rising, extreme inequality is one of the most dominant economic and political issues in America today [5]. You could argue that it’s harder to be poor in America; a poor Cuban can still get free healthcare and college. On the other hand, and rich Americans get to hold on to essentially all of their income [6], while private businesses in Cuba cannot. America also has a wealth of premier universities — for those who can afford to attend.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/504/1*3k0zz4Vk0h8khC2PqNF2XA.jpeg" /></figure><figure><img alt="" src="https://cdn-images-1.medium.com/max/896/1*DKFLbIVOvLYR17BpJm__yg.jpeg" /></figure><figure><img alt="" src="https://cdn-images-1.medium.com/max/896/1*MxmbP48l-35cL1t6BCWcVw.jpeg" /><figcaption>Pieces of America.</figcaption></figure><p>After both the 2016 Brexit referendum and the 2016 US General Election: Trump triumphant, nationalism filled international news as a primarily anti-immigrant sentiment, one fueled by shifting demographics and a global refugee crisis. This nationalism is one that conflates whiteness with country, one that willfully ignores history in order to proclaim that real Americans, or real Brits, have pale skin and practice certain faiths. In studying American history, you can read about the <a href="https://en.wikipedia.org/wiki/Indigenous_peoples_of_the_Americas#United_States">pre-Columbian indigenous groups</a>, <a href="https://www.smithsonianmag.com/history/lesser-known-history-african-american-cowboys-180962144/">black cowboys</a>, <a href="https://www.uscitizenship.info/Chinese-immigration-and-the-Transcontinental-railroad/">immigrant Chinese transcontinental railroad workers,</a><strong> </strong><a href="https://www.nps.gov/articles/irish-soldiers-in-the-union-army.htm">immigrant German and Irish Catholic soldiers in the Union army</a>, <a href="http://oxfordre.com/americanhistory/view/10.1093/acrefore/9780199329175.001.0001/acrefore-9780199329175-e-217">Mexican and Filipino farmworkers</a>, and quickly realize that this brand of nationalism is based on a selective history that is perversely and profoundly <em>wrong</em>. (You c̶o̶u̶l̶d̶ should also read the book <a href="https://www.zinnedproject.org/materials/peoples-history-of-the-united-states"><em>A People’s History of the United States</em></a> by Howard Zinn.)</p><p>This nationalism grew stronger in response to an immigration crisis at our southern border. That humanitarian crisis is in turn caused by decades of our own shortsighted foreign policy in regions such as Central America. In contrast, the nationalism that gripped the anti-colonial independence movements of the 20th century is tied to the struggle against empire and, loosely, emphasizes a timely unity over division as a means to overthrow a common suppressor.</p><p>Just as it is understandably difficult for us modern Americans to truly grasp colonialism, I think it even more difficult to understand nationalism that is not <em>exclusionary</em> in our current political climate.<strong> </strong>Has there ever been a significant instance of modern Americans being united against a common oppressor, when so many groups have been disenfranchised for so long in our nation’s history? The World Wars may seem like good candidates, but there’s a stark difference between an oppressor and an enemy. It’s been a long time since our nation was truly oppressed by another, making it difficult to imagine growing a nationalism based in unity [7].</p><p>It’s also a bit scary that our relationships with countries like Cuba remain clouded by the Cold War, the striking split between capitalism and socialism leaving little room for nuance (social democracy, anyone?) in discussing issues like healthcare, education, industry, and trade. A <a href="https://www.economist.com/leaders/2019/02/14/millennial-socialism">total swing in the other direction</a> isn’t necessarily the best response to every problem of the status quo. We should at this point be able to have a nuanced discussion that acknowledges the successes and failings of both of these political and economic ideologies, going beyond political labels by recognizing both as just that: ideologies, not policies. The implementation details matter the most.</p><p>This discussion has the potential to heal some of America’s current political divisiveness. We should have it, and looking at the current House of Representatives and the Democratic candidate platforms beginning to form for 2020, it seems that we will. In doing so, we should be free to design our society around our values. Unlike many other nations, we don’t really have the excuse of a colonizer. No one is stopping us but ourselves.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/489/1*FLfeofHf0BL3QAfmsGLwFw.jpeg" /></figure><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*l6G8ztdOUnBa1FoMIuRRoA.jpeg" /></figure><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*Uu8Egk0kfS70-r2clA_H3w.jpeg" /></figure><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*8VLYISmAGUOEdUkTNM1xBA.jpeg" /></figure><h4><strong>Notes and Further Reading</strong></h4><p>[1] Factors contributing to economic decline or stagnation included the <a href="https://en.wikipedia.org/wiki/Opium_Wars">Opium Wars</a> in China, the mandated export of raw materials and import of finished British goods in India, and the <a href="http://www.bl.uk/learning/histcitizen/trading/story/trade/4tradingplaces.html">forced opium trade of the East India Company</a> in both countries.</p><p>[2] The <a href="https://www.nytimes.com/2008/03/02/books/review/Kurtz-Phelan-t.html">United Fruit Company</a> played a prominent role in 20th century US imperialism in Latin America and the Caribbean.</p><p>[3] Imperialism, Colonialism, Socialism, and Communism are all overloaded terms, as is Capitalism. I’ve typically seen and associated colonialism with the British and European empires, imperialism with the US, socialism with Latin America, the US, Modern Europe, and the USSR, and communism with China and the USSR, but this simple breakdown isn’t always accurate. Wikipedia is pretty enlightening.</p><p>[4] <a href="https://en.wikipedia.org/wiki/Hillbilly_Elegy">Hillbilly Elegy</a> was a book that helped popularize this discussion, but there’s plenty of criticism out there over its potential <a href="https://www.washingtonpost.com/opinions/i-grew-up-in-poverty-in-appalachia-jd-vances-hillbilly-elegy-doesnt-speak-for-me/2017/08/30/734abb38-891d-11e7-961d-2f373b3977ee_story.html?utm_term=.dddce24b5e6e">stereotyping</a> and the concept of a <a href="https://www.wsj.com/articles/the-false-stereotype-of-two-americas-11545869142">binary divide</a><em>.</em></p><p>[5] <a href="https://www.independent.co.uk/news/world/americas/us-developing-nation-regressing-economy-poverty-donald-trump-mit-economist-peter-temin-a7694726.html">US has regressed to developing nation status, MIT economist warns</a>. See also: <a href="https://www.inequalitymedia.org/">Inequality Media</a>.</p><p>[6] Due to US tax code, it’s <a href="https://www.washingtonpost.com/news/wonk/wp/2016/10/03/why-thousands-of-millionaires-dont-pay-federal-income-taxes/?utm_term=.1dbf237e7326">possible to be a millionaire and pay little to no federal income tax</a>.</p><p>[7] Optimism says climate change can be this unifier; pessimism says it will be a catalyst for conflict over limited resources.</p><p>| <a href="https://www.nationsonline.org/oneworld/History/Cuba-history.htm">Cuban History Reference</a> | <a href="https://www.youtube.com/watch?v=P2Obp6YS4SY">Estela Bravo’s 2001 Documentary on Fidel Castro</a> | <a href="https://www.huffingtonpost.com/topic/90-miles">Huffpost Series on US-Cuba Relations</a> | <a href="https://www.bbc.com/news/world-latin-america-12159943">BBC US-Cuba Timeline</a> |</p><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=60fbf39b996b" width="1" height="1" alt="">]]></content:encoded>
        </item>
        <item>
            <title><![CDATA[What We Talk About When We Talk About Bias (A guide for everyone)]]></title>
            <link>https://medium.com/@ayesharbajwa/what-we-talk-about-when-we-talk-about-bias-a-guide-for-everyone-3af55b85dcdc?source=rss-3902f5af4e8a------2</link>
            <guid isPermaLink="false">https://medium.com/p/3af55b85dcdc</guid>
            <category><![CDATA[bias]]></category>
            <category><![CDATA[machine-learning]]></category>
            <category><![CDATA[big-data]]></category>
            <category><![CDATA[artificial-intelligence]]></category>
            <category><![CDATA[fairness]]></category>
            <dc:creator><![CDATA[Ayesha Bajwa]]></dc:creator>
            <pubDate>Fri, 17 Aug 2018 23:07:49 GMT</pubDate>
            <atom:updated>2019-03-09T03:46:04.215Z</atom:updated>
            <content:encoded><![CDATA[<p><em>Adapted from my paper for the </em><a href="https://www.media.mit.edu/courses/the-ethics-and-governance-of-artificial-intelligence/"><em>Ethics and Governance of AI class</em></a><em> at the MIT Media Lab. Written for those with concerns about bias in artificial intelligence (AI) systems but don’t have the relevant technical background. However, it’s not too watered-down. Don’t worry, there’s a glossary and many links!</em></p><p><em>While bias is most often introduced in data and observed in predictive, real-world applications, it usually surfaces at multiple stages in the development pipeline. We attempt to cover the underlying concepts and vocabulary in order to have a meaningful discussion about various perspectives on bias. In this post, we introduce AI systems from a traditional machine learning (ML) perspective in a fair amount of detail. We provide examples and discussion of how data bias, model bias, learning algorithm bias, system bias, and human-in-the-loop bias are observed in ML systems. Hopefully, we’ll begin to bridge the technical basics of ML with the larger societal concerns of bias in the process.</em></p><h3><strong>Introduction</strong></h3><p>The field of <em>machine learning</em> (ML) — which is a subfield of <em>artificial intelligence</em> (AI) — is enormously empirical, with even the most experienced researchers often lacking a satisfactory explanation for why a particular technique does or does not work well. Approaches that work well in practice on <em>datasets</em>, as measured by error or accuracy values, can become field-wide standards without rigorous justification. An<a href="https://xkcd.com/1838/"> xkcd comic</a> provides a succinct illustration of this philosophy.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/371/1*UKIHA2AHtB9WPG-KrfwSZg.png" /><figcaption>(<a href="https://xkcd.com/1838/">Source</a>)</figcaption></figure><p>ML papers sometimes produce their own explanations of why particular techniques work without any other evidence from the field. An example is the phenomenon of “internal covariate shift,” a problem both hypothesized and solved by the authors of the <a href="https://arxiv.org/abs/1502.03167">original paper</a> on <em>batch normalization</em>, a now-common technique used for training <em>neural networks</em>. While this phenomenon is barely discussed beyond this paper, batch normalization is a technique that many researchers employ — because it works. The theme of using whatever works well, without having explained or justified it, portends dangerous consequences for when automated systems get it wrong.</p><p>Due to its potential for high impact and profit, <em>deep learning</em> moves at fast pace and is especially pervasive in industry labs, though it <a href="https://www.technologyreview.com/s/603984/googles-ai-explosion-in-one-chart/">impacts many areas</a> of science and technology. This competitiveness requires high operational secrecy, as evidenced by the research divisions of large companies which <a href="https://www.bloomberg.com/news/articles/2018-02-13/in-the-war-for-ai-talent-sky-high-salaries-are-the-weapons">pay their top researchers enormous salaries</a> but externally publish just some fraction of their results (and <a href="https://www.wired.com/story/ai-and-enormous-data-could-make-tech-giants-harder-to-topple/">almost never share their datasets</a>). The increasing availability of big data held by private companies underpins the current AI revolution in industry. However, these big datasets can also be a primary source of bias.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/572/1*Igl3QF6u0OLCVMm0oSgKRg.jpeg" /><figcaption>From Moritz Hardt’s Fair ML course. (<a href="https://fairmlclass.github.io/">Source</a>)</figcaption></figure><p>Once neglected by the research mainstream, discussions of bias and fairness in ML have become a new norm in both research and industry. Moritz Hardt, ML researcher and UC Berkeley professor, teaches a <a href="https://fairmlclass.github.io/">Fairness in Machine Learning course</a> and has written <a href="https://medium.com/@mrtz/how-big-data-is-unfair-9aa544d739de">clear and comprehensible blog posts</a> on the subject. Kate Crawford, a principal researcher at Microsoft Research and NYU who focuses on the social implications of AI, was a <a href="https://www.youtube.com/watch?v=fMym_BKWQzk">2017 keynote speaker at NIPS</a>, a top-tier ML conference. Researchers predict that the European Union’s <a href="https://www.eugdpr.org/">General Data Protection Regulation</a> (GDPR), passed in 2016, will raise the standards for industry by <a href="https://arxiv.org/abs/1606.08813">effectively creating a “right to explanation”</a> for individuals who have been subject to algorithmic decisions. A <a href="https://obamawhitehouse.archives.gov/sites/default/files/docs/big_data_privacy_report_may_1_2014.pdf">2014 report</a> by the White House details findings about bias in algorithmically-generated models, such as “redlining in the digital economy,” and expresses concern that most “research was not able to determine exactly why a racially biased result occurred.” The tech and AI giants Amazon, Apple, Microsoft, DeepMind, Google, Facebook, and IBM are the <a href="https://www.npr.org/sections/alltechconsidered/2016/09/28/495812849/tech-giants-team-up-to-tackle-the-ethics-of-artificial-intelligence">founding partners</a> of the Partnership on AI to Benefit People and Society, a consortium dedicated to establishing best practices for AI and increasing transparency. Google and DeepMind have established <a href="https://www.theguardian.com/technology/2017/jan/26/google-deepmind-ai-ethics-board">internal ethics committees</a> with an AI focus, though little information on them is publicly available.</p><p>The increasing interest in bias and fairness in machine learning by the research community, companies, public interests, and governments suggests that bias is more than a technical problem. However, some level of technical understanding is helpful in understanding how bias arises.</p><p>Before delving into a discussion of bias, we first establish basic ML concepts. There are two common types of machine learning, <em>supervised</em> and <em>unsupervised learning</em>. We focus on supervised learning, which requires as input a <em>training dataset</em> of input data examples with associated <em>labels</em>; a typical example is a dataset of animal photos, each labeled with a class such as “cat”. We want to create a <em>model</em> that maps inputs to output predictions. A supervised learning algorithm learns a model from data by starting with some generic model type and optimizing the model <em>parameters</em> in order to minimize the <em>training error</em> on the training dataset — this can also be viewed as maximizing the model’s accuracy. It does so by minimizing an <em>objective function</em>, which is generally based on the error between the actual labels and the predicted values.</p><p>We can visualize this optimization by imagining the model parameters incrementally changing in the appropriate directions, based on each example the current model sees as input, in order for the model to generate the appropriate output (the model sees the labels in training). Optimization algorithms such as <em>gradient descent</em> simply push the parameters in the right directions, decreasing error. Intuitively, we should minimize the training error because we want to achieve the correct predictions on the training data as an indication that the model is working, producing expected results for data it has seen. For example, when training a <em>deep neural network (DNN)</em> to distinguish images with cats from images without cats, we understand that our model will probably not work well until it is able to correctly distinguish cat and non-cat images in the <em>training dataset</em>. Once the model achieves some reasonable accuracy, we can try it out on a new, <em>test dataset </em>— which is simply a dataset of the same type, with the same categories of associated labels as the training set. However, the trained model is now trying to predict the labels of the test dataset.</p><p>In this post, we break down the issue of bias by considering the stages at which it may be introduced and observed. These stages fall along a pipeline of how many ML systems are constructed. The premise is to learn patterns from data, so we begin with <strong>data bias </strong>(Section 2), which we argue is essential to describe precisely, as it is usually the stage at which bias is introduced. We move upwards to <strong>model bias</strong> (Section 3) and <strong>learning algorithm bias</strong> (Section 4), which are very closely tied. This concludes the core of any ML development pipeline. We further consider <strong>system bias</strong> (Section 5) in the malicious intent and emergent behavior cases. Finally, we consider how human behaviors can bias <strong>human-in-the-loop systems</strong> (Section 6); this discussion is crucial due to the increasing deployment of ML systems in real-world situations. We put forth the opinion that, while bias is most often introduced in data and observed in predictive, real-world applications, it affects multiple stages in the pipeline.</p><p>We don’t provide a formal definition of bias or claim to cover all cases. We simply lay out a framework of stages at which it’s useful to consider bias. Loosely, we think of bias as the ways in which a system that appears to function correctly may fall short of designer, legal, or social expectations and definitions of fairness. These expectations and definitions are subjective and need to be formally defined by societies and governments in the long term.</p><h3><strong>Section 1: Case Scenarios</strong></h3><p>We define a pipeline to illustrate various stages at which bias may occur in the following sections. Consider three case scenarios that we will trace through the first part of the pipeline (data bias through learner bias):</p><ul><li><a href="https://www.propublica.org/article/machine-bias-risk-assessments-in-criminal-sentencing">ProPublica’s findings</a> that Northpointe’s COMPAS, used for predicting criminal risk scores, appears to be “biased against blacks.”</li><li>A toy example about mortality in the ICU.</li><li>An <a href="https://intelligence.org/files/AIPosNegFactor.pdf">ML parable</a> popularized by Eliezer Yudkowsky about the US Army’s failed attempt to train a neural network to distinguish camouflaged enemy tanks in forest settings.</li></ul><h3><strong>Section 2: Data Bias</strong></h3><p>Since machine learning problems are usually formulated as learning a predictive model from real data, it’s natural to find real-world biases reflected in automated systems. We place a huge amount of blind faith in the quality of datasets when training data-hungry models.</p><p>One can train a good algorithm on skewed data and obtain a biased model — in fact, this outcome would be expected. The opaque nature of many <em>state-of-the-art</em> ML systems means that detecting, explaining, or correcting this bias is difficult.</p><p>A high profile societal example of bias in automated systems is <a href="https://www.propublica.org/article/machine-bias-risk-assessments-in-criminal-sentencing">ProPublica’s work on criminal justice</a>. Julia Anguin and a team of ProPublica reporters found that, when predicting risk assessment scores, the false positive and false negative rates are biased against black individuals and favorable to white individuals. Though this phenomenon is often termed “machine bias” (as in the article title) or “algorithmic bias,” there is a clear aspect of data bias surrounding the issue; critics point out that <a href="https://fivethirtyeight.com/features/technology-is-biased-too-how-do-we-fix-it/">blacks are arrested more than whites, even when committing crimes at the same rate</a>. Law professor Ifeoma Ajunwa, who gave a <a href="https://www.eeoc.gov/eeoc/meetings/10-13-16/ajunwa.cfm">testimony</a> on big data implications before the Equal Employment Opportunity Commission, argues that the number of convictions a person has is “not a neutral variable.”</p><p>The implications of data bias are enormous. Data bias means that, regardless of the particular model used, <em>any</em> statistical prediction model relying on pattern recognition will exhibit the same kind of biases present in the training data.</p><p>We encounter similar issues in domains where data is uneven or sparse, and we must pay attention to baselines. Consider a toy example in which we want to predict mortality in a particular ICU. Assume the baseline mortality rate is 7%. If we train a model which gives us 93% accuracy on a test set, we should immediately be suspicious. Any trained model which always predicts survival in this ICU achieves 93% accuracy, but is nonetheless useless. Keep in mind that 100% training accuracy is not necessarily desirable either, due to <em>overfitting</em> (see Section 3: Model Bias). If the data is uneven, as it is in this case where 93% of the samples in a sufficiently large dataset will be associated with the survival label and 7% will be associated with the death label, it’s common to observe this form of bias. When there are large imbalances in <em>labels</em>, baselines are key.</p><p><a href="https://medium.com/@mrtz/how-big-data-is-unfair-9aa544d739de">Moritz Hardt writes</a>, “the lesson is that statistical patterns that apply to the majority might be invalid within a minority group,” and when standard approaches are used at the population-level, the learned outcomes are those that best represent the majority. This disparity easily translates into unfairness for minority groups, since predictions for a majority class will be more accurate than predictions for any minority class. The below example from <a href="https://medium.com/@mrtz/how-big-data-is-unfair-9aa544d739de">Hardt’s blog post</a> illustrates how a model may be worse for a minority group. Imbalances in the data result in worse predictive power for examples with a non-majority label, as in the ICU example.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/661/1*p0W2T3_iYG5_1YiDbPjI3A.png" /><figcaption>This trained Gaussian mixture model, shown with solid and dashed ellipses, achieves a poor fit on the minority data (fewer points) in the bottom left corner compared to the majority data (more points) at the top right. From Moritz Hardt’s blog post (<a href="https://medium.com/@mrtz/how-big-data-is-unfair-9aa544d739de">Source</a>).</figcaption></figure><p>While we may assign particular labels to a training dataset, the labeled attributes might not be what is actually learned by a model. A <a href="https://www.jefftk.com/p/detecting-tanks">classic ML parable</a> about army tanks illustrates a common pitfall of the requirement for similar training and test datasets, reminding us that we often lack insight into <em>how</em> a model is learning to associate input data with particular labels. According to the story, the US army developed neural network models for detecting camouflaged enemy tanks. The trained model correctly distinguished the 50 photos of camouflaged enemy tanks from the 50 photos of the forest, and the performance on the held-out test data — also 50 photos of camouflaged enemy tanks and 50 of the forest — was also highly accurate. However, after the model was shipped to the Pentagon, it was rejected as no better than humans at detecting enemy tanks. The catch: in both the training and test datasets (which were split from the same original dataset), photos of tanks were taken on cloudy days, while forest photos were from sunny days. The model had learned to distinguish sunny from cloudy days, unsurprisingly giving poor performance when employed to detect enemy tanks.</p><p>Large technology companies have come under scrutiny for the bias discovered in their products. For instance, <a href="https://www.bloomberg.com/graphics/2016-amazon-same-day/">Amazon was accused of discrimination and redlining</a> when rolling out its same-day service, despite the fact that it does not use racial or ethnic data in its maps. While it is often illegal to use race as an algorithmic consideration, companies such as Amazon find themselves in a bind just by using seemingly reasonable metrics, such as ZIP codes with a high concentration of Prime members. Apparently benign metrics can function as proxies for race or socioeconomic status, reinforcing inequality of access to products or services in disadvantaged areas.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/400/1*Eb3mQzvVahba4mCO3CeHQA.png" /><figcaption>A breakdown of current data interrogation standards from the Dataset Nutrition Label team. (<a href="https://arxiv.org/pdf/1805.03677.pdf">Source</a>)</figcaption></figure><p>A recent <a href="https://arxiv.org/pdf/1805.03677.pdf">survey</a> conducted by an interdisciplinary team of the Assembly program at the MIT Media Lab and the Berkman Klein Center for Internet &amp; Society at Harvard found that in practice, data scientists as a whole lack consistent standards, with existing standards being highly volatile and domain dependent. Out of the 34 individuals surveyed, just under half (44%) reported that their organizations lacked best practice standards for data analysis, while approximately a quarter reported that their organizations have best practice standards. This team developed the concept of a <a href="http://datanutrition.media.mit.edu/">Dataset Nutrition Label</a>, a tool which gives data scientists some standard information — both qualitative and quantitative — about the fitness and viability of the data, what sorts of biases it may contain, and what sorts of application domains might be reasonable. This project highlights a lack of rigor with respect to choosing the right dataset for a particular task. Standards such as the Dataset Nutrition Label aim to drive accountability and mitigate harm, but widespread adoption is required and even then they don’t completely prevent bias. A common problem with data bias is that, while introduced at the data level, its effects are often unclear until further down the ML development pipeline.</p><h3><strong>Section 3: Model Bias</strong></h3><p>While data bias usually leads to problems further down the pipeline, the model is another stage at which bias may become apparent or amplified. Even with unbiased data, models may be wrong or unfair in a variety of ways. Since real-world data is rarely unbiased, the particular type of model evaluation can surface varying symptoms of bias.</p><p>One common type of model bias is <em>overfitting</em>, the bane of Intro ML students. Overfitting occurs when a model is fit so specifically to a training dataset that it gives poor performance on the test dataset. Even with the important assumption that the training and test datasets are drawn from the same distribution or generated in the same way, overfitting is dangerous.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/522/1*tLp-EwEvkT31KlU2GRu4Rw.jpeg" /><figcaption>A simple 2D example of polynomial overfitting.</figcaption></figure><p>Overfitting is usually avoided by constraining model complexity using <em>regularization</em>. In the language of mathematical optimization, regularization introduces a new constraint. When the learning algorithm tweaks the parameters to achieve minimization of the error, it must do so while satisfying the regularization constraint. This usually increases the <em>training error</em> slightly, but decreases the <em>test error</em> due to an improved ability to generalize. We show an overfitted model (orange) contrasted with a regularized one (green) as a visual examples.</p><p>Addressing overfitting with regularization is largely an art, not a science, and certain types of regularization approaches (especially ones that induce <em>sparsity</em>) may not be suitable for unevenly labeled data. Overfitting and regularization are examples of how standard ML issues and techniques, when used in societal contexts, may provide varying fairness and accuracy for different groups.</p><p>In our toy ICU and enemy tank scenarios, the models are wrong in simple ways. In the ICU example, the fact that the model always predicts one class is suspicious. Looking at the baselines, it becomes obvious that a 93% accuracy is not predictive. In the camouflaged enemy tanks example, the 100% training and testing error is fishy, and the datasets provided should be more varied for a system expected to function in the real-world scenario of camouflaged enemy tank detection. Both the ICU and enemy tank model flaws may be caught by researchers with minimal probing and evaluation of the models.</p><p>Symptoms of bias vary depending on how we evaluate a model. Consider the varying evaluations of <a href="http://www.northpointeinc.com/files/downloads/FAQ_Document.pdf">Northpointe’s COMPAS</a>. Cathy O’Neil, data scientist and author of <a href="https://weaponsofmathdestructionbook.com/"><em>Weapons of Math Destruction</em></a>, <a href="https://fivethirtyeight.com/features/technology-is-biased-too-how-do-we-fix-it/">says</a> “Northpointe answers the question of how accurate it is for white people and black people, but it does not ask or care about the question of how inaccurate it is for white people and black people: How many times are you mislabeling somebody as high-risk?”— and how does this affect how we judge fairness? Matthias Spielkamp, a founder of Berlin-based nonprofit <a href="https://algorithmwatch.org/en/">AlgorithmWatch</a>, points out in an MIT Technology Review <a href="https://www.technologyreview.com/s/607955/inspecting-algorithms-for-bias/">article</a> that the ProPublica reporters and the rebuttal by Northpointe don’t actually contradict each other, given varying definitions of fairness. While the ProPublica study compared false positive and false negative rates for blacks and whites respectively, finding them unfavorable to blacks and favorable to whites, the Northpointe study compared the positive predictive value across races and found the model to be consistent. While both types of evaluation may be technically valid, the “correct” one to use depends on societal values and definitions of fairness. Until these definitions are made explicitly legal, regulated, and enforced, there is no incentive to perform the appropriate type of model evaluation.</p><p>Unsupervised ML approaches, though free of labels, are not necessarily free of bias. Researchers at Boston University and Microsoft Research have demonstrated that <em>word embeddings</em>, a popular technique for representing text as numerical vectors, often <a href="https://arxiv.org/abs/1607.06520">exhibit high gender bias</a> when trained on large, real-world datasets such as Google News. Since word embeddings are commonly used by ML researchers, there is a risk that gender bias will not only be reflected but also amplified in applications where word embeddings are used. After quantifying these biases, the researchers devised a sequence of steps to neutralize and “de-bias” word embedding algorithms, with techniques such as ensuring gender-neutrality of known gender-neutral words and enforcing distance constraints between particular embeddings. Their approach successfully retains the usefulness of word embeddings while decreasing the gender bias they exhibit. Practical research endeavors such as this are critical to ensure prolonged traction and interest in the issues of bias and fairness within research communities.</p><p><a href="https://www.poetofcode.com/">Joy Buolamwini</a> of the MIT Media Lab <a href="https://medium.com/mit-media-lab/incoding-in-the-beginning-4e2a5c51a45d">discusses her personal experiences</a> with bias as a black woman whose face was often unrecognizable by facial recognition software that had no trouble with white faces. Buolamwini argues that code libraries — which improve efficiency by allowing developers to build on previous work instead of re-implementing it — can help contribute to a “coded gaze,” reflecting a lack of inclusion in the technology sector as well as in training datasets. The solution is largely to solve the data bias; by promoting the utilization and widespread adoption of more diverse training datasets, Buolamwini thinks the problem of the “coded gaze” can be solved. In her examples, addressing data bias fixes the model bias as well.</p><iframe src="https://cdn.embedly.com/widgets/media.html?src=https%3A%2F%2Fwww.youtube.com%2Fembed%2F162VzSzzoPs%3Ffeature%3Doembed&amp;display_name=YouTube&amp;url=https%3A%2F%2Fwww.youtube.com%2Fwatch%3Fv%3D162VzSzzoPs&amp;image=https%3A%2F%2Fi.ytimg.com%2Fvi%2F162VzSzzoPs%2Fhqdefault.jpg&amp;key=a19fcc184b9711e1b4764040d3dc5c07&amp;type=text%2Fhtml&amp;schema=youtube" width="854" height="480" frameborder="0" scrolling="no"><a href="https://medium.com/media/d1ed3623daa41100850340bb26be544e/href">https://medium.com/media/d1ed3623daa41100850340bb26be544e/href</a></iframe><p>Another key point from ML basics is that we generally learn <em>parameters</em>, not entire <em>models</em>. When we train a <em>deep neural network</em> for a particular task, we must somewhat arbitrarily choose a network architecture upfront. Examples of chosen, as opposed to learned, architectural parameters for a neural network may include the number of layers, the types of each layer (e.g. <em>fully-connected</em>, <em>convolutional</em>), the number of <em>nodes</em> in each layer, and the <em>activation functions</em> (e.g. <em>ReLU</em>, <em>Sigmoid</em>, <em>Softmax</em>) to use. Fixed parameters that are not modified during training are called <em>hyperparameters</em>. Other parameters which are not architectural but are also not learned from data include the <em>learning rate</em> or <em>step size</em>, model constants, and initialization values. Sometimes, these are chosen with the help of a <em>validation dataset</em>, which functions as a pre-test dataset that a model produced from a training dataset might be immediately tested on for tweaking of hyperparameters. The<em> objective function</em> must also be chosen well. The metric of success is nearly always accuracy; bias and fairness are not usually considered because of the enormous trust placed in the veracity of the data.</p><p>Those who study machine learning distinguish <em>estimation error </em>and <em>structural error</em>. Formally, estimation error occurs when the parameters of a hypotheses cannot be estimated well based on the training data, while structural error occurs when the <em>hypothesis class</em> corresponding to a particular model (e.g. a neural network architecture) cannot produce a specific hypothesis (e.g. the network architecture with specific learned parameters) that performs well on test data. Estimation error decreases with larger training datasets, as more training data allows the model to refine the parameters and perform better on test data. However, even when training data is increased (and overfitting is prevented), a model may still suffer from structural error — it may simply be the wrong sort of model for the problem, no matter how we tweak the parameters.</p><p>Model evaluation in the context of societal applications can be difficult when models have varying levels of interpretability. While simple models (e.g. linear regression) have fairly high interpretability and explainability, the large and complex <em>state-of-the-art</em> neural networks that power many modern AI systems are opaque in comparison — they are <em>black-box</em> models. There has been a prioritization of accuracy over interpretability and explainability within the research community that makes black-box models<em> </em>acceptable. However, many agree that only explainable models should be used in certain applications; complex deep neural networks, for example, are not commonly used in criminal justice.</p><p>While complexity and inexplicability are clear hazards to fairness, simplicity does not guarantee a bias-free model. The purportedly race-neutral model <a href="http://www.predpol.com/how-predictive-policing-works/">PredPol</a>, for forecasting crime as an aid to predictive policing, is based on just three data characteristics: the type, location, and date and time of previously committed crimes. It is still found to be <a href="https://medium.com/in-justice-today/setting-the-record-straight-on-predictive-policing-and-race-fe588b457ca2">biased under certain evaluations</a>. The criminal justice researchers Kristian Lum and William Isaac respond to criticism that their analysis is based on drug crime data, which is not the type of crime the software is typically used to forecast. They write that PredPol “has been used by some jurisdictions to allocate resources to police other crime types that suffer from similar statistical bias induced by the highly discretionary nature of enforcement.” As Cathy O’Neil and others have pointed out, the method of evaluation matters as just much as the model, so evaluating models on data sufficiently similar to that used in practice is critical to any analysis. And, as in the case of PredPol, not all models are used as the designers intended.</p><h3><strong>Section 4: Learner Bias</strong></h3><p>We use <em>learning algorithms</em>, or <em>learners</em>, to obtain trained models from data. The types of models we consider for an ML problem generally have associated learning algorithms to optimize the model parameters, so it is somewhat artificial to separate this section from the one on model bias. The dataset, model, and learning algorithm form the core of an ML development pipeline.</p><p>Learning algorithms take in training data and produce the tuned parameters of the given model. To learn the weight vector and offset for a linear classifier, we may use <em>Perceptron</em>. For a complex <em>neural network</em> architecture, we use <em>gradient descent</em> and <em>backpropagation</em> to learn the weight and bias parameters of the network from the training data. The pairings between models and their associated learning algorithms implies that learning algorithm bias is closely tied to the issue of model bias. The common phrase of “algorithmic bias” can be a misnomer, since it usually refers not to the learning algorithm but to the test application of a model with learned parameters. People who talk about “algorithmic bias” are often really talking about data bias emerging in a model after training.</p><p>Problems of model bias are relevant to learning algorithms as well, including the blind faith in data as well as the potentially arbitrary nature of the hypothesis class, or type of model (and associated learning algorithm), chosen for a particular problem. A learning algorithm that works well is expected to tweak a model’s parameters to replicate statistical patterns present in the data — bias included. As Moritz Hardt <a href="https://medium.com/@mrtz/how-big-data-is-unfair-9aa544d739de">writes</a>, protected classes such as race and gender can easily be “latent in the observed attributes,” so “nothing prevents the learning algorithm from discovering these encodings.” It is possible for a learning algorithm that sees no data about race to — picking up on statistical correlations — produce a model that, when evaluated, is found to produce worse outcomes for those of minority races.</p><p><a href="https://ieeexplore.ieee.org/document/8025197/">A study on emotion recognition</a> in images attempts to correct for some bias in data imbalances by “(1) identifying the set(s) of minority classes, (2) developing specialized learners that address the minority class via special focus on the class, and (3) developing a specialized learner that combines signals from both the minority and majority class models,” focusing on the learning algorithms used. The researchers combine a generalized learning algorithm with specialized learners for the particular minority classes they identify in a conscious effort to combat bias.</p><p>General considerations and limitations of models are true of learning algorithms as well. We consider the issue of <em>robustness</em>, which is a question of perturbation and stability. A robust model will give similar results for similar inputs, and a robust learning algorithm will output similar models when fed similar input data. Models and learning algorithms that are not robust in this manner may be more susceptible to bias.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/662/1*FVdVbjnuKaeXQQ2PJlyq2w.png" /><figcaption>An input image and visualization of features detected at locations throughout the image. From “The Building Blocks of Interpretability” (<a href="https://distill.pub/2018/building-blocks/">Source</a>)</figcaption></figure><p>The issues of interpretability and explainability are relevant to learning algorithms, as they are for models; they currently enjoy a huge focus in the ML research community. The European Union’s GDPR is <a href="https://arxiv.org/abs/1606.08813">expected to become a strong policy incentive</a> for this work, since it effectively creates a “right to explanation” when algorithmic decisions are used. <a href="https://distill.pub/">Distill</a>, an online publication venue that aims to make advances in ML more transparent and accessible, is further evidence that the field is pivoting in this direction. In one <a href="https://distill.pub/2018/building-blocks/">article</a>, researchers break down learned <em>features</em> of a neural network along the <em>hidden layers </em>for a variety of input images in an interactive environment. We show above one of their examples of an input image alongside a visualization of features detected at a layer early in the network, with individual component images scaled by intensity; this representation makes it clear that edges in the images are strongly detected. Distill focuses on interactive interfaces that make the learning process easier to follow and interpret. With increased interpretability and explainability in learning algorithms and models, bias can become clearer at the time of training — since we have a perspective on the implicit features being learned — instead of after model deployment.</p><h3><strong>Section 5: System Bias</strong></h3><p>Context and evaluation are important in identifying bias. While “system” is an overloaded term, we consider two important types of bias which may be observed when ML systems are deployed at a large scale in the real world. Firstly, there are the malicious cases in which an existing model — its inner workings may or may not be opaque to users — is manipulated by someone who has spent a long time studying its inputs and outputs in order to produce the malicious actor’s desired output. Secondly, there is the emergent case in which we observe network effects from simple human behaviors — an obvious example is clickbait.</p><p><a href="https://arxiv.org/abs/1602.02697">Targeted adversarial attacks</a> on large-scale ML systems are a common form of malicious behavior. A team of researchers <a href="https://arxiv.org/abs/1606.06565">write</a> that many such general vulnerabilities come from “unintended and harmful behavior that may emerge from machine learning systems when we specify the wrong objective function, are not careful about the learning process, or commit other machine learning-related implementation errors” —that is, from accidents that increase system susceptibility to adversaries. However, even <em>state-of-the-art</em> open-source models (which presumably don’t suffer from simple implementation errors or thoughtlessness) show a remarkably high adversarial susceptibility to targeted attacks. In one case, <a href="https://arxiv.org/abs/1804.08598">an MIT study</a> demonstrated that Google’s InceptionV3 classifier was fooled into thinking that a 3D printed turtle, covered with a carefully engineered adversarial pattern, was actually a rifle.</p><iframe src="https://cdn.embedly.com/widgets/media.html?src=https%3A%2F%2Fwww.youtube.com%2Fembed%2FpiYnd_wYlT8%3Ffeature%3Doembed&amp;display_name=YouTube&amp;url=https%3A%2F%2Fwww.youtube.com%2Fwatch%3Fv%3DpiYnd_wYlT8&amp;image=https%3A%2F%2Fi.ytimg.com%2Fvi%2FpiYnd_wYlT8%2Fhqdefault.jpg&amp;key=a19fcc184b9711e1b4764040d3dc5c07&amp;type=text%2Fhtml&amp;schema=youtube" width="854" height="480" frameborder="0" scrolling="no"><a href="https://medium.com/media/580d586496417703f973730ca610f4f9/href">https://medium.com/media/580d586496417703f973730ca610f4f9/href</a></iframe><p>On social platforms, human users click or share clickbait more often, and content ranking algorithms usually favor content that is more clicked and shared. A recent <a href="http://science.sciencemag.org/content/359/6380/1146">study</a> on the spread of fake news on Twitter found that false rumors spread faster than true news; the top 1% of false rumors spread to between 1000 and 100,000 people, while the truth reached only 1000 users. Platforms such as Facebook have responded to the fake news problem by <a href="https://www.forbes.com/sites/kathleenchaykowski/2018/03/06/facebooks-latest-algorithm-change-here-are-the-news-sites-that-stand-to-lose-the-most/#204bd04f34ec">modifying their ranking algorithms</a> to prioritize content posted by friends and family over content originally shared by news platforms on social media. The findings that high-engagement, often highly polarizing content takes priority in social networks may have as much to do with human behavior as it does with the ranking algorithms of the social media platforms themselves.</p><p>Other systems can surface bias in the results of algorithms that find patterns in large amounts of data. <em>Collaborative filtering</em> is a technique commonly used in recommender systems, such as suggested purchases on Amazon based on your purchase history and what “people like you” bought. The simplest version of recommender systems could surface problems of bias, especially since recommendations can change behavior, but <a href="https://arxiv.org/abs/1506.04135">evaluation of a recommendation algorithm on a historical dataset</a> would not reflect these changes. Once we acknowledge that bias may be in issue in recommender systems, the immediate question of how to fairly censor bias arises. It is unclear what legal or regulatory implications there might be for a biased recommendation system, since suggestions of products to purchase is not an obvious harm — contrast recommendations to, for example, Harvard Professor Latanya Sweeny’s findings that <a href="https://arxiv.org/abs/1301.6822">online advertisement delivery could show emergent racism</a> when fed black-sounding names.</p><p>While the phrase “system bias” is overloaded, we should recognize two prevalent cases of adversarial susceptibility of large systems and the emergent behavioral trends, both of which can lead to biased outcomes.</p><h3><strong>Section 6: Human-In-The-Loop Bias</strong></h3><p>A simple case of human-in-the-loop bias can occur when a human, such as a judge, a doctor, or a human safety driver in a self-driving vehicle, acts in conjunction with an automated system. While it it is clear that people may hold their own unconscious biases or overt prejudices, it is far more complicated to consider the interaction between human actors and automated systems acting in unison.</p><p>A common human-in-the-loop system is the safety driver that companies such as Waymo and Uber use for their autonomous vehicle (AV) testing. In the case of <a href="https://www.nytimes.com/interactive/2018/03/20/us/self-driving-uber-pedestrian-killed.html">the first AV fatality</a>, a self-driving Uber vehicle hit a pedestrian crossing a section of road at night, while the safety driver was distracted and looking away from the road. According to some experts, the safety driver <a href="https://www.bloomberg.com/news/articles/2018-03-22/video-said-to-show-failure-of-uber-s-tech-in-fatal-arizona-crash">would have been able to intervene</a> and potentially save the pedestrian had he not been looking away from the road seconds before. The safety driver’s own confidence that the AV would behave as expected was a form of bias that subverted the safety driver’s purpose in the human-in-the-loop system, leading to a fatal outcome. This example also suggests that in human-in-the-loop systems, it is difficult to ensure that the human actor behaves as intended. Human actors are likely to adapt their behavior over time after interacting with a system.</p><p>Clinical contexts for human-in-the-loop systems are growing, though many say AI is <a href="https://www.technologyreview.com/s/610552/google-x-ray-project-shows-ai-wont-replace-doctors-any-time-soon/">not likely to replace doctors in the near future</a>. Clinicians have been found to have <a href="https://www.ncbi.nlm.nih.gov/pubmed/12414468">cognitive biases that affect how they provide care</a>. <a href="https://www.ncbi.nlm.nih.gov/pubmed/28472533">Implicit biases have been found along racial lines</a> in clinical settings, with white people favored, but the exact clinical implications of this finding have not been characterized. While decision support tools (DSTs) are being adopted in order to mitigate human error and bias, they <a href="https://www.ncbi.nlm.nih.gov/pubmed/21335682">don’t entirely prevent bias</a> in clinical outcomes, especially when clinicians defer to their judgment in close cases. This deference to DSTs is termed <a href="https://www.ncbi.nlm.nih.gov/pubmed/21335682">automation bias</a>, occurring in situations in which “users tend to accept computer output without sufficient thought.”</p><p>We have already considered criminal justice as an important area for detecting and preventing automated bias. While risk scores “<a href="https://www.technologyreview.com/s/608011/secret-algorithms-threaten-the-rule-of-law/">seem scientific, an injection of computational rationality into a criminal justice system riddled with discrimination and inefficiency</a>,” various analyses have found them to be flawed or biased against particular populations. However, <a href="https://www.jstor.org/stable/42705348">human judges are susceptible to various forms of bias</a> as well. Given the imperfectness of current risk scoring systems, it’s easy to imagine that human bias coupled with bias in a risk score, through confirmation bias, could lead to even more biased outcomes — especially when trends in enforcement, which lead to biased datasets, are reflected in judicial biases.</p><p>We may also consider the interactions of ML systems with entire communities and societies when considering human-in-the-loop bias. For example, when Google Photos in 2015 began <a href="https://www.wired.com/story/when-it-comes-to-gorillas-google-photos-remains-blind/">tagging black people as “gorillas”</a> in photos, the public outcry on social media was an indication of a cultural bias — in this case, a strong reaction to an offensive historical context. We can be almost certain that Google’s engineers did not use training data in which black people were labeled as gorillas or develop a model of racist historical contexts to inject into their systems. Due to the lack of interpretability and explainability in how image classification models learn, there was no easy fix for Google, so they simply removed “gorillas” (as well as “chimp” and “monkey”) as potential tags. This mislabeling anecdote is a fluke, a specific instance which raised an outcry, among a much wider variety of one-off misclassification errors. If we consider users of Google Photos to be human-in-the-loop moderators, most errors don’t elicit strong responses. Therefore, most errors are simply tolerable and not worth fixing. At the time of this writing, Google Photos remains “blind” to gorillas — the original problem was never properly solved, and <a href="https://www.wired.com/story/when-it-comes-to-gorillas-google-photos-remains-blind/">the patch fix remains</a>.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/681/1*sm2FcZHQtB6i50Hn-CuBXw.png" /><figcaption>MIT Media Lab professor Iyad Rahwan’s visualization of a human-in-the-loop system. (<a href="https://medium.com/mit-media-lab/society-in-the-loop-54ffd71cd802">Source</a>)</figcaption></figure><p>Human-in-the-loop bias is an important consideration as AI systems continue to pervade everyday life. A key takeaway is that, in human-in-the-loop systems, the weak point in terms of predicting overall system behavior is often the human. Human-in-the-loop systems (which can extend to <a href="https://medium.com/mit-media-lab/society-in-the-loop-54ffd71cd802">society-in-the-loop systems</a>) make us consider the tradeoff between removing potential human biases and leaving decisions to automated systems that may lack explanatory capabilities. A standing assumption is that a good tradeoff between the two will enable humans to make better decisions in fields such as criminal justice and medicine.</p><h3><strong>Consequences and Future Directions</strong></h3><p><a href="http://www.katecrawford.net/">Kate Crawford</a> defines two types of “harms” in her discussions on AI bias. The first is “allocative harm,” which occurs when a system is biased against particular populations; racial bias is one example. The second is “representational harm,” occurring when biased systems model the world in a manner that preferences certain viewpoints; an example is to search the Internet for images of “CEO” and to see only images of men as a result. While both these examples likely stem from data bias and are detectable at multiple pipeline stages, allocative harm has historically been the focus of the research community and regulatory groups. We have discussed both types of harms in previous sections. In many anecdotes, allocative and representational harms could have been been prevented either by developers or by regulators, in multiple stages along the development pipeline.</p><p>There are a number of initiatives to address data bias, including the previously mentioned <a href="http://datanutrition.media.mit.edu/">Dataset Nutrition Label project</a>, <a href="https://www.ajlunited.org/">Joy Buolamwini’s advocacy work</a>, and a growing awareness within industry and academia that data, though powerful, should not be blindly trusted. Research on <a href="https://arxiv.org/abs/1607.06520">de-biasing word embeddings</a> and on <a href="https://ieeexplore.ieee.org/document/8025197/">combining demographic-specific learning algorithms</a> are key examples of how the values of fairness and representation can be instilled within research communities to mitigate both allocative and representational harms, while providing usable tools and standards to do so.</p><p>Model and learning algorithm bias become easier to detect as we are able to better inspect and explain the learning process. Work on explainability and interpretability in the ML community within both industry and academia (often with collaborations between the two) is driven by both the promise of increasing the accuracy of models and the incentive to explain system behavior in applied settings. Explaining system effects and human-in-the-loop behaviors can help build more robust systems and decrease bias in the output. The <a href="https://www.forbes.com/sites/kathleenchaykowski/2018/03/06/facebooks-latest-algorithm-change-here-are-the-news-sites-that-stand-to-lose-the-most/#2f7ef4e334ec">evolution of Facebook’s content ranking algorithm</a> is one example of addressing bias in a real system interacting with user populations. The idea of <a href="http://which can extend to society-in-the-loop systems">society-in-the-loop </a>systems is also compelling, but for the most part, we’re still working towards the infrastructure and standards that would make that possible.</p><p>Ultimately, the goal of researchers, developers, and regulators should be to detect and combat bias along multiple stages of the development pipeline. While testing the output of a system or product for <a href="https://techcrunch.com/2017/04/30/algorithmic-accountability/">“algorithmic accountability”</a> is a good start, this ideal implies a need for industry to test for bias at various points in the development process instead of at the output alone. It implies questioning the canonical blind faith in data, particularly in large datasets. It will be important to be vigilant about bias from all angles, with a special focus on building inclusive and representative datasets, thorough and appropriate model evaluation, an informative legal and regulatory framework that incentivizes fairness, and a comprehensive perspective on the real-world contexts in which AI systems are applied.</p><h3>Acknowledgements</h3><p>Thanks to Amy Zhang and Natalie Saltiel for their feedback, and to Kasia Chmielinski for providing a summary of survey data.</p><h3><strong>Glossary</strong></h3><p><a href="https://en.wikipedia.org/wiki/Activation_function"><strong><em>Activation Function</em></strong></a> — The function associated with a <em>neural network node </em>that specifies the node output given the input(s). Generally used to introduce non-linearity into <em>models</em>.<br> <br> <a href="https://en.wikipedia.org/wiki/Artificial_intelligence"><strong><em>Artificial Intelligence (AI)</em></strong></a> — Intelligence demonstrated by automated systems. A very broad field; <em>machine learning</em> (ML) is one part of it.</p><p><a href="https://en.wikipedia.org/wiki/Backpropagation"><strong><em>Backpropagation</em></strong></a> — The calculus-based technique used in <em>gradient descent</em> algorithms to optimize the <em>parameters</em> of <em>neural network</em> models.</p><p><a href="https://en.wikipedia.org/wiki/Batch_normalization"><strong><em>Batch Normalization</em></strong></a> — An empirical technique that, in practice, improves the performance and stability of <em>neural networks</em> by normalizing the inputs at each layer of the network.</p><p><a href="https://en.wikipedia.org/wiki/Black_box"><strong><em>Black-Box</em></strong></a><strong><em> </em></strong>— In general, systems for which we know or understand only the inputs and outputs. In ML, black-box <em>models</em> are models for which we lack explainability. <em>Deep neural networks </em>are the typical example.</p><p><a href="https://en.wikipedia.org/wiki/Statistical_classification"><strong><em>Classification</em></strong></a> — The ML problem of identifying what category a particular input belongs to by learning function mappings. Binary classification involves distinguishing between two classes, while multiclass classification involves distinguishing between a set of classes.</p><p><a href="https://en.wikipedia.org/wiki/Collaborative_filtering"><strong><em>Collaborative Filtering</em></strong></a><strong><em> </em></strong>— a technique used by recommender systems such as online vendors (Amazon, Netflix). Can be item-based or user-based.<br> <br> <a href="https://en.wikipedia.org/wiki/Convolutional_neural_network"><strong><em>Convolutional Neural Network (CNN)</em></strong> </a>— A subset of n<em>eural networks</em> that use convolution (or cross-correlation) operations. They are commonly used in image <em>classification</em> because they preserve the 2D structure of images.<br> <br> <strong><em>Data</em></strong> — The data used in ML problems may be labeled, where each data point comprises of an input and the known corresponding output, or unlabeled, where each data point is just the input because there is no specific corresponding output or it is not known.<br> <br> <a href="https://en.wikipedia.org/wiki/Deep_learning"><strong><em>Deep Learning</em></strong> </a>— A highly current subfield of <em>machine learning </em>that includes models such as <em>CNNs</em>, <em>RNNs</em>, and other <em>deep neural network (DNN)</em> architectures.<br> <br> <a href="https://en.wikipedia.org/wiki/Deep_learning#Deep_neural_networks"><strong><em>Deep Neural Network (DNN)</em></strong> </a>— <em>Neural networks</em> with multiple <em>hidden layers</em>, or layers between the input and output layers.<br> <br> <strong><em>Estimation Error</em></strong> — Occurs when the <em>parameters</em> of a <em>hypothesis class</em> cannot be estimated well from the <em>training data</em>.<br> <br><strong><em> </em></strong><a href="https://en.wikipedia.org/wiki/Feature_(machine_learning)"><strong><em>Features</em></strong></a> — Specific attributes or measurable properties of of input <em>data</em> that are used to train <em>models</em> via <em>learning algorithms</em>. The process of choosing features must be done carefully in order to achieve high accuracy.<br> <br> <a href="https://en.wikipedia.org/wiki/Feedforward_neural_network"><strong><em>Feedforward</em></strong></a> — A <em>neural network </em>architecture in which there are no cycles and every <em>node</em> in a layer is connected to every node in the previous and next layers. A <a href="https://en.wikipedia.org/wiki/Multilayer_perceptron">Multilayer Perceptron</a> is one type of feedforward network. Contrast with <em>convolutional</em> architectures.<br> <br> <a href="https://en.wikipedia.org/wiki/Gradient_descent"><strong><em>Gradient Descent</em></strong></a> — A calculus-based approach to optimizing <em>model parameters</em> by finding error minima. Many <em>learning algorithms</em> are based on gradient descent.<br> <br> <a href="https://www.quora.com/What-is-the-role-of-a-hidden-layer"><strong><em>Hidden Layers</em></strong></a> — Layers between input and output layers in a <em>neural network</em>. Neural networks with multiple hidden layers are by definition <em>DNNs</em>.<br> <br> <a href="https://en.wikipedia.org/wiki/Hyperparameter_(machine_learning)"><strong><em>Hyperparameters</em></strong></a> — Hyperparameters of a <em>model</em> are not learned, but are set beforehand and do not change. They are common for similar models.<br> <br> <a href="https://www.quora.com/What-is-hypothesis-in-machine-learning"><strong><em>Hypothesis Class</em></strong></a> — The set of <em>models</em> (learned functions) that can be realized by a <em>learning algorithm</em>, which tunes the <em>model parameters</em>. <br> <br> <a href="https://en.wikipedia.org/wiki/Labeled_data"><strong><em>Labels</em></strong></a> — Correct outputs or output classes corresponding to particular inputs. Labeled <em>data</em> is used in <em>supervised learning</em>.<br> <br><strong><em> </em></strong><a href="https://en.wikipedia.org/wiki/Outline_of_machine_learning#Types_of_machine_learning_algorithms"><strong><em>Learning Algorithm</em></strong></a> — The algorithm that can be applied to data to produce a <em>model</em> belong to the <em>hypothesis class(es)</em> associated with that learning algorithm.<br> <br> <strong><em>Learning Rate </em></strong>— The <em>hyperparameter</em> controlling the speed of the updates in <em>learning algorithms</em> such as <em>gradient descent</em>.<br> <br> <a href="https://en.wikipedia.org/wiki/Machine_learning"><strong><em>Machine Learning (ML)</em></strong></a> — A subfield of <em>artificial intelligence</em> (AI) that is premised on learning patterns from <em>data</em>. Closely tied to data science and statistics. (Note: can be synonymous with <em>deep learning</em> in modern contexts).<br> <br> <a href="https://www.quora.com/What-are-different-models-in-machine-learning"><strong><em>Model</em></strong></a> — A function that is a particular instance of a <em>hypothesis class</em> that is learned from <em>data</em> by a <em>learning algorithm</em> and used to predict test outputs given the corresponding inputs.<br> <br><strong><em> </em></strong><a href="https://en.wikipedia.org/wiki/Neural_network"><strong><em>Neural Network</em></strong></a> — A generally complex <em>model</em> represented by architecture of <em>nodes</em> or neurons, usually with multiple layers. The learned <em>parameters</em> of such a model are weights and biases. While <em>neural networks </em>generally have input and output layers, <em>deep neural networks</em> also contain <em>hidden layers</em>.<br> <br> <strong><em>Node</em></strong> — A unit in a <em>neural network</em>. Also called a neuron. Each node has an associated <em>activation function</em>.<br> <br> <a href="https://en.wikipedia.org/wiki/Loss_function"><strong><em>Objective Function</em></strong></a> — The objective of the optimization problem in any ML formulation. It is usually an error to minimize, subject to some <em>regularization</em> constraints.<br> <br> <a href="https://en.wikipedia.org/wiki/Overfitting"><strong><em>Overfitting</em></strong></a> — The phenomenon in which a particular dataset is fit too closely by a <em>model</em>, resulting in poor generalizability for that model. Mitigated with <em>regularization</em>.<br> <br> <a href="https://machinelearningmastery.com/difference-between-a-parameter-and-a-hyperparameter/"><strong><em>Parameters</em></strong></a> — The parameters of a model are learned during training as they are modified by the <em>learning algorithm</em>. They are usually initialized to something strategic that will allow training to proceed smoothly.<br> <br> <a href="https://en.wikipedia.org/wiki/Perceptron"><strong><em>Perceptron</em></strong></a> — A simple, linear classifier and associated <em>learning algorithm</em>. Analogous to a <em>neural network</em> with <em>linear activation functions</em> and just one <em>node</em> or neuron.<br> <br><a href="https://en.wikipedia.org/wiki/Recurrent_neural_network"> <strong><em>Recurrent Neural Network (RNN)</em></strong></a>— A type of <em>neural network</em> commonly used to model and generate sequential data.<br> <br> <a href="https://en.wikipedia.org/wiki/Regularization_(mathematics)"><strong><em>Regularization</em></strong></a> — The usual method to address <em>overfitting</em> by introducing additional optimization constraints. Mathematically, a common form of regularization is to penalize the norm of the <em>parameter</em> (weight) vector.<br> <br><strong><em> </em></strong><a href="https://en.wikipedia.org/wiki/Rectifier_(neural_networks)"><strong><em>Rectified Linear Unit (ReLU)</em></strong></a>— A simple, common <em>activation function</em> (including its corresponding <em>node</em>) that simply outputs zero for negative inputs and outputs positive inputs without modification. <br> <br> <strong><em>Robustness</em></strong> — Informally, the stability of the <em>learning algorithm</em> or <em>model</em> output when the input is perturbed. Robustness is important for consistency and reliability.<br> <br><a href="https://en.wikipedia.org/wiki/Semi-supervised_learning"> <strong><em>Semi-supervised Learning</em></strong></a> — A form of supervised ML in which unlabeled examples are also leveraged. Contrast to <em>supervised learning</em> and <em>unsupervised learning</em>.<br> <br> <a href="https://en.wikipedia.org/wiki/Sigmoid_function"><strong><em>Sigmoid</em></strong></a> — A common <em>activation function</em>. Useful because it is differentiable (one can take the derivative) at every point, allowing for optimization through <em>backpropagation</em>.<br> <br> <a href="https://en.wikipedia.org/wiki/Softmax_function"><strong><em>Softmax</em></strong></a> — A function used to represent categorical distributions in <em>classification</em>, usually applied at the output layer of <em>neural network </em>classifiers. Also called the normalized exponential function.<br> <br> <a href="https://www.quora.com/What-do-engineers-mean-by-sparse-features-in-machine-learning-in-laymen-terms"><strong><em>Sparsity</em></strong></a> — The statistical concept of <em>data</em> that is mostly empty or zero. Intuitively, the opposite of dense. Useful because it can be used to reduce the complexity of very high-dimensional data.<br> <br> <strong><em>State-Of-The-Art</em></strong> — In general, the cutting edge technology. In ML, state-of-the-art usually refers to the <em>models</em> and <em>learning algorithms</em> which achieve the highest accuracy.<br> <br> <strong><em>Step Size</em></strong>— Same as <em>learning rate</em>.<br> <br><a href="https://en.wikipedia.org/wiki/Stochastic_gradient_descent"> <strong><em>Stochastic Gradient Descent (SGD)</em></strong> </a>— The form of <em>gradient descent</em> commonly used in practice, which leverages randomization for faster learning.<br> <br> <strong><em>Structural Error</em></strong> — Occurs when the <em>hypothesis class</em> instance corresponding to a particular <em>model</em> cannot produce a specific hypothesis that performs well on <em>test data</em>, regardless of how the <em>parameters</em> are tuned.<br> <br><strong> </strong><a href="https://en.wikipedia.org/wiki/Supervised_learning"><strong><em>Supervised Learning</em></strong></a> — The form of ML in which known input-output pairs are used to learn the <em>model</em>, or function mapping between input-output pairs, for a particular problem formulation. Contrast to <em>unsupervised learning</em>.<br> <br> <a href="https://en.wikipedia.org/wiki/Training,_test,_and_validation_sets"><strong><em>Training Data</em></strong><em> </em></a>— The data fed to a <em>learning algorithm</em> in order to determine a <em>model</em> by tuning <em>parameters</em>. May be labeled or unlabeled <em>data</em>. Contrast to <em>test data</em>.<br> <br><strong> <em>Training Error</em></strong> — The error rate or proportion of the <em>training data</em> when run through the trained <em>model</em>. Contrast to <em>test error</em>.<br> <br><strong> </strong><a href="https://en.wikipedia.org/wiki/Training,_test,_and_validation_sets"><strong><em>Test Data</em></strong></a>— The <em>data</em> input to a trained <em>model</em> that the model has not yet seen. Used to evaluate the model’s performance and generalizability to new data. Contrast to <em>training data</em>.<br> <br><em> </em><strong><em>Test Error</em></strong>— The error rate or proportion of the <em>test data </em>when run through the trained <em>model</em>. Contrast to <em>training error</em>.<br> <br> <a href="https://en.wikipedia.org/wiki/Unsupervised_learning"><strong><em>Unsupervised Learning</em></strong></a>— The form of ML in which a function is inferred to describe hidden structure from unlabeled <em>data</em>, or data inputs for which the corresponding output is not known upfront. Contrast to <em>supervised learning</em>.<br> <br><strong> </strong><a href="https://en.wikipedia.org/wiki/Training,_test,_and_validation_sets"><strong><em>Validation Data</em></strong> </a>— The <em>data</em> input to a trained <em>model</em> is neither the training nor test data. This dataset is often used to manually tweak <em>hyperparameters</em>. Contrast to <em>training data </em>and <em>test data</em>.<br> <br><strong> <em>Validation Error </em>— </strong>The error rate or proportion of the <em>validation data</em> when run through the trained <em>model</em>. Contrast to <em>training error</em> and <em>test error</em>.<br> <br><strong> </strong><a href="https://en.wikipedia.org/wiki/Word_embedding"><strong><em>Word Embeddings</em></strong></a><strong> — </strong>A category of feature-learning techniques in the ML subfield of natural language processing (NLP). Used to map text into high-dimensional vectors.</p><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=3af55b85dcdc" width="1" height="1" alt="">]]></content:encoded>
        </item>
        <item>
            <title><![CDATA[On Street Harassment]]></title>
            <link>https://medium.com/@ayesharbajwa/on-street-harassment-1f783da4f4c0?source=rss-3902f5af4e8a------2</link>
            <guid isPermaLink="false">https://medium.com/p/1f783da4f4c0</guid>
            <category><![CDATA[fear]]></category>
            <category><![CDATA[gender-roles]]></category>
            <category><![CDATA[feminism]]></category>
            <category><![CDATA[gender-equality]]></category>
            <category><![CDATA[street-harassment]]></category>
            <dc:creator><![CDATA[Ayesha Bajwa]]></dc:creator>
            <pubDate>Sun, 23 Oct 2016 20:20:11 GMT</pubDate>
            <atom:updated>2016-10-23T20:21:06.001Z</atom:updated>
            <content:encoded><![CDATA[<p><em>All experiences discussed below are my own or those of my immediate friends, occurring in NYC between June and August 2016. </em><strong><em>Note: contains strong language and blunt description.</em></strong></p><p>New York City is a wonderful place. I would move back, though there is one aspect that gives me pause. Street harassment — such as catcalling, leering, being followed, being inappropriately crowded in a fairly empty subway car, and the like — happens often.</p><p>Here are a few of my theories. I think you’re safer when wearing sunglasses, safer when walking briskly. The length of your hemline or the amount of bared shoulder may or may not matter. Heels make it worse. Do not stop on certain streets of Midtown, or men will come up to you. On the subway, be prepared to move. If someone stopping you to ask for directions does not begin with “Excuse me…” then he is probably not asking for directions, but rather your age or evening plans (sample size 3). Your height may make you more of a target in either case, as can eye contact or lack thereof. If you are not obviously Caucasian, your ethnicity will undoubtedly be fetishized. “Where you from? Damn, what are you?” Time of day can matter less than vicinity. The closer you look to a stereotypical young working professional or girl going out, the more unwanted attention you will draw.</p><p>Walking with a male friend usually helps; people assume. Walking with other females does not, and people think they can get away with more. If you and a female friend are accosted — basically, approached by strangers and told how pretty you are — and after three minutes you are the one to say flatly that you are not interested, you will be called ugly. A default assumption is that, for girls, being told you’re beautiful is the best possible compliment, being called ugly the worst possible insult.</p><p>If a creepy old man says your dress is “unfair” because you have “great legs,” your stony silence will draw further aggression. Angry yells follow as you walk away from these interactions. “B*tch can’t take a compliment! Why’d you stop, then? F*cking weirdo.”</p><p>It can’t be a compliment when it’s intended to objectify you.</p><p>At the very least, these incidents are an annoyance. It only takes one jerk to ruin your day. At their worst, they reinforce senseless societal narratives of what defines a woman’s worth and make you fear for your physical safety.</p><p>They will also make you paranoid. If you’re not safe from unwanted attention while browsing in a <em>bookstore</em>, then are you safe at a networking event? Or should you think twice about that guy telling you he’s single, though it was in the perfectly natural context of a discussion on work-life balance? You feel bad for feeling uncomfortable, for feeling suspicious. If I hadn’t felt a hundred percent safe in my own workplace, I might have lost my mind.</p><p>After three months in the city, I found no obvious way to address it other than to speak up about it (particularly so when certain types of people are running for President!) and to cope. You don’t want to minimize the larger issue, though you have no choice but to minimize its effect on your daily life.</p><p>After all, reports of such behavior will not be taken seriously until after you have been physically harmed. Does that seem extreme?</p><p>I mean, I’ve had friends who were catcalled by men in NYPD uniforms.</p><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=1f783da4f4c0" width="1" height="1" alt="">]]></content:encoded>
        </item>
        <item>
            <title><![CDATA[Four Weeks]]></title>
            <link>https://medium.com/@ayesharbajwa/four-weeks-1237fa22d4ee?source=rss-3902f5af4e8a------2</link>
            <guid isPermaLink="false">https://medium.com/p/1237fa22d4ee</guid>
            <category><![CDATA[education]]></category>
            <category><![CDATA[teach-for-america]]></category>
            <category><![CDATA[teaching]]></category>
            <dc:creator><![CDATA[Ayesha Bajwa]]></dc:creator>
            <pubDate>Fri, 05 Feb 2016 05:07:16 GMT</pubDate>
            <atom:updated>2016-02-05T05:08:57.387Z</atom:updated>
            <content:encoded><![CDATA[<p>In between college semesters, I was a teacher for a month. In reality I was a teacher for just a few days, or hours, or perhaps mere moments. I felt like a real teacher when the voice in my head piped up to say, <em>what you’re doing right now matters to people — it’ll be missed when you go away</em>. As a sophomore at an elite technical school, I never feel this way on campus. If I weren’t enrolled at MIT, someone equally qualified and committed would be in my place. Not so for teaching.</p><p>My host teacher understands this, since she is one of a mere handful of STEM mentors at her school. She’s earning her teaching credential through Teach For America, one of the largest national organizations in the field of education. TFA trains mostly freshly minted college graduates to teach in the nation’s most underserved schools. By that I mean schools in poor communities that serve mostly students of color and tend to be located in inner cities or rural areas.</p><p>Despite an intense summer training program, most TFA teachers learn on the job. Some are excellent; others crack under the stress of a new environment and give up on their students. Some enter the program with savior mentalities reminiscent of the White Man’s Burden, others participate to gain “real-world experience” after college, others simply want to serve, and yet others want a fast track to a teaching credential. I’m happy to say that my host was in fact overqualified to teach her subject, that she cared about teaching well, and that she invested time and effort in understanding each of her students. She certainly didn’t deem herself their savior.</p><p>I believe that TFA is overall a good organization with noble goals. However, that does not make each corps member an automatically good teacher, much less a saint. If, for example, you majored in a humanities field, haven’t even taken math yourself since high school, and have no passion whatsoever for that subject, then you probably shouldn’t be teaching it. In many cases, TFA will still let you. Since decent STEM teachers are hard to come by when more lucrative industry jobs exist, TFA lets virtually <em>anyone</em> in their corps teach math and science.</p><p>I had relatively little preparation, certainly less than a TFA participant, for the drastic change in environment I experienced when I began teaching in the Rio Grande Valley in south Texas. As a Bay Area born and Silicon Valley raised MIT student majoring in a technical field, my view of America was limited to tech-focused places full of high schoolers bound to elite colleges. I was vaguely aware that most of America was different, but the gap between knowing and comprehending was, in my case, profound. Further narrowing my view was the fact that both my parents, and in fact all four of my grandparents, attended college. My parents are engineering PhDs. There had never been a doubt, financial or otherwise, that I’d attend college.</p><p>My strikingly different socioeconomic and educational background made me initially wary. The Valley (not the CA but the TX one) was racially, religiously, culturally, socioeconomically, culinarily, linguistically, and educationally different from anywhere I’d lived. While I took in the sights of flat plains, water towers, grackle-covered power lines, and drive-thrus enthusiastically enough during our daily commute, being on campus made me nervous.</p><p>I’m naturally introverted. I was somewhat shy during my first week at the school. In the classroom, I didn’t initiate many conversations other than homework help. Mostly, I sat in the back as my host taught her lessons as usual. During the engineering project sessions I led, I went over technical concepts with groups of three or four students. I didn’t know how my name and ethnicity would be perceived, so I went by a generic name (one with no hint of my ethnicity) for simplicity.</p><p>While most of the school’s teachers were welcoming and exchanged more than a few words with me, others treated me coldly in the staffroom and even said not-so-nice things about me to my host. I sometimes had to interact quite gingerly, in a manner that made me even more acutely aware of my differences. There were teachers who referred to me as “MIT.” I was even handed a poorly-written science fair grading rubric with MIT written on top in place of my name. The residue of these odd interactions seeped into the classroom, where I perpetually feared being misunderstood or viewed as a privileged snob. I worried that I was inadvertently and genuinely elitist when I noticed English teachers making grammar mistakes and Math teachers doing sample problems incorrectly — was it my duty to speak up? Many of the juniors I taught were below grade level, and I had to cover the basics of fraction multiplication and addition before I could explain how to simplify trig expressions. I resented every math teacher who had encouraged using a calculator as a crutch, leaving the students stumbling blindly without one. In the back of my mind, a constant reminder throbbed: <em>I’m just three years older than these kids.</em></p><p>It took me more than two weeks to realize that a focus on differences is counterproductive.</p><p>Once I started mentoring after-school activities more regularly, I learned how to interact with the students on a personal level that rendered all other circumstances irrelevant. Nobody thinks about socioeconomic status in the heat of a soccer scrimmage. Nobody cares about ethnicity when trying to drive a robot. Nobody cares about your native language when playing with an LED circuit. Teaching comes into its own outside the classroom, since the best conversations do not usually contain tested content.</p><p>The first lesson I taught went fairly well; I believe I fell short of no standards but my own — I thought I could have done better. A mere three days later, I taught another lesson. I ran class as usual, taking more liberties to explain concepts in my own way. Start with the basics. Build up. Start <em>again</em> with the basics, and this time ask the students for input. Let them finish your sentences, ask “are you sure?” no matter what answers they offer, and when they start something the wrong way, gently nudge them to realize why what they’re doing isn’t helpful. Most of all, do not ever dismiss them as wrong without explanation. They internalize it as evidence of their own stupidity — which is untrue and completely uncalled for — in a stubborn, heartbreaking way.</p><p>I resent the focus on testing. The goal of achieving the maximum passing percentage of students doesn’t incentivize teachers to encourage those already performing well, who are sadly the ones who might do anything in the relevant field after high school. The truly interested students aren’t being encouraged any further. A focus on test scores means that a teacher who instructs the students to use their calculators like life-support will receive a better rating than a teacher who attempts to actually teach math. The students themselves believe they are intelligent or stupid based on what scores they tend to receive, leading to a vicious cycle beginning in elementary school (and I believe the initial input to such a cycle is likely something as simple as having some books in the house or eating proper meals).</p><p>I truly loved teaching when I knew it was going well. You feel immense gratification when your normally worst-behaved student asks you after class, “Miss, can you always teach? I like how you explain things.” You feel it when a student realizes, unprompted, that he can find the magnitude of a vector using the Pythagorean Theorem and gives an audible exclamation of excitement. You feel it when a student writes you a letter saying that because of a conversation he had with you, he knows where he wants his life to go after high school. You feel it when, two weeks after you explained over and over how distribution was the opposite of factoring, you see one of your students meticulously explaining it to another. You feel it when a student who usually sleeps in class says that your positive attitude makes him feel like working.</p><p>I had more credibility after teaching the lesson on my own. I think I came across as genuine more successfully than ever before. The intense, insightful, hilarious, and profound conversations I had with students in my final week of teaching far outweighed any negativity I’d ever felt in the staffroom. I discovered that one student liked to read Jung in his free time, while another was strongly pro-choice because of a family member’s unwanted pregnancy, and that another possessed a secret poetic streak.</p><p>They wrote me letters for my last day. I got wind of it beforehand, tipped off by a few overenthusiastic juniors, and I resolved to write each student a farewell note. They were short but specific, individualized with anecdotes and inside jokes, since I had finally learned every single face and name. The letters I received were sweet, hilarious, artistic, and genuine. I spent nearly all five hours of my flights back to Boston rereading them.</p><p>Now that my month of teaching has ended, I won’t claim to have saved anyone, done anything particularly noble, or inherently changed anyone’s life. If anything, I’m the one who has learned.</p><p>Some enthusiastic young teachers enter their first classroom only to be disillusioned for life by the impossibility of singlehandedly mending the current, broken education system. While I can understand this sentiment after my month in the classroom, I see it slightly differently. Your job is not to fix the entire education system. Your job is to help your students realize their own brilliant capacity to learn, to care, and to succeed.</p><p>It is not impossible.</p><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=1237fa22d4ee" width="1" height="1" alt="">]]></content:encoded>
        </item>
    </channel>
</rss>