How our AI called a scenario for Trump being elected and the ‘why’ behind it
As the world was collectively watching results of the 2016 election , we got a barrage of emails from people congratulating our AI for predicting a Trump win.
Now before I dive into self praise, some context: After the first debate, we did a small experiment based on a random sampling of social media posts using labelled data feeds, a CNN (convolutional neural network; a feed-forward artificial neural network), a Bayesian methods-based network and a variant of word2vec-like algo.
To be very clear, our AI did not call a Trump win, per se, but rather conditions that would support a Trump or a Hillary win, i.e. the following — and I am quoting content from an email sent to CNN’s New Day and Bloomberg Gadfly on October 3, 2016:
a) If macro conditions stay the same, Hillary will win the election 54% to Trump’s 41%. We were planning on updating this for every debate.
b) If there is a terror attack within 2 weeks of the election or if the markets (S&P 500) drops more than 10% between now and November 8, Trump will narrowly win 52% to Hillary’s 46%.
c) Trump’s engagement shows the half life of Trump’s passive supporter base (those that engage with only 1/4 tweets or posts with affirmation to what Trump tweeted or said) is roughly 4.3 days after an engagement. What that means is that these passive supporters are likely to moderate (i.e. support a neutral engagement) after 4.3 days. In cases of a terror attack this picture changes dramatically with moderation occurring only after 11.2 days.
The scenario laid forth above is pretty clear. The “why,” however, is far more critical than the output, and I cannot stress how important the “why” is, because it reduces the overall blackbox-like effect of ANNs (artificial neural networks) — especially important in decision making. If you are an investor in AI, your ultimate bet is that the process of knowledge and reason will itself be commoditized. Too many people who pursue investments in these sectors do so with utter disregard for the process of how reasoning and knowledge work , a faculty otherwise known as epistemology.
Within the very broad, boring — yet intellectually controversial — subject of epistemology, there are two widely held beliefs — the empirical view that experience is the source of all knowledge and the rationalist view that reason is the source of all knowledge. There is no definitive answer, and philosophically both arguments are valid; the former more than the latter, as the latter implies a priori knowledge about an object.
On a more practical level, and in context of Bayesian epistemology, the relationship between knowledge, reasoning and experience can be illustrated using the sunrise problem. If we take an example of a person on a beach seeing the sunrise every day , the assumption is that after seeing X number of sunrises, said person would have enough prior experience to predict a sunrise in the future.
Mathematically, it is impossible to know whether a sunrise will occur until it actually does — somewhat similar to the Schrödinger’s cat thought experiment. However, for an entity that has seen the sunrise X number of times , the probability becomes albeit a number that can be ignored and the sunrise becomes a “universal truth” of sorts until there is anomaly. An anomaly in this context could mean any change in the surrounding — for example, what if the sunrise did occur but the person was unable to see it because of grey skies? Broadly, what all of this means is that surroundings matter in context of an output — e.g. type of weather, season, time of sunrise, etc., and these surroundings need to be weighed in real time, in the present, against prior experiences in order to achieve a qualitatively better output as close as possible to a “universal truth.”
To blame an ANN or an AI or an algorithm is to blame humanity itself.
Going back to our email — Point (a) is fairly straightforward and it depicts a win for Hillary if an election were to occur at that point in time, i.e. after the first debate. It is no secret that Hillary’s strong performance in the debate was a key driver for this output, so I want to extrapolate more on this. Point (b), however, takes into account an anomaly event. Prior experiences (from April 2015 to date) showed that support for Trump surged, following the horrific attacks in Paris and Brussels and after the stock market meltdown in January-February 2016. Ergo, the weighting for domestic security changed in a materially significant way after the attacks. Similarly economic concerns were weighed in a materially significant way after the stock market sell-off in January and February. This gives you a sense of how passive supporters and neutral voters reacted to a change in the surroundings, causing them to reweigh their support for either candidate at that point in time.
Another aspect is time – which in this case determines time taken for a weight of a particular surrounding (e.g. security, economy) to materially change and impact an action (e.g. a pro Trump or Hillary stance). This is important because the election is held on a specific date and is not a “never-ending event.” From prior experiences this pool of people tended to revert to a neural stance by engaging (e.g. liking, RT, sharing, etc.) a moderate position than Trump on the economy and security after 4.3 and 11.2 days respectively.
As we all know, neither a terrorist event nor a stock market correction of 10 percent occurred. However, two things did occur. The S&P had a major correction, falling nine consecutive days — its worst in 36 years — on the back of October’s Comey surprise — itself a security incident given the FBI’s involvement. Bottom line — the surroundings had materially changed to the point where they made an impact in how passive/neutral voters weighed them, ultimately affecting the output, i.e. the vote.
It is therefore necessary to design an AI or an ANN that weighs prior experiences of surroundings (in this case text or words) using vector relationships between different words or numbers in context of raw data (this is materially different than sentiment analysis, which has propensity to output false positives), and then apply that to present surroundings.
Additionally, from a product-design perspective, simplicity lies in explanation of the value brought forth to an end user. Doing so ultimately determines the logic and the rationale for an ANN’s output. On that level, the human race is exceptionally lucky to be good at guessing and finding everyday “universal truths” — such as the location of our car keys.
Those who have good instincts tend to be more apt in decision-making than others. The core point, though, is one can’t actually predict (e.g. weather) or even output (e.g. a summarized report) anything with absolute certainty — you are merely weighing surrounding information based on prior experiences, in the present and in relation to an entity, hoping that the sun has risen because a ray of light can be seen coming through the clouds.
On a closing note, it is important to understand the role of biases. To be very clear, each one of us has a bias. Man is, after all, a social animal. As social animals, we tend to seek out and move in groups that we are comfortable with. Nowhere is this bias reflected more than in Silicon Valley itself — where for every successful startup founded by an immigrant, like WhatsApp, there is a Theranos or a Clinkle.
The issue here isn’t necessarily success or failure, but rather the bias that extends into decision making processes. To blame an ANN or an AI or an algorithm (I am not referencing the Facebook News Feed algorithm, because that deserves a much deeper explanation as it relates to advertising and engagement) is to blame humanity itself. Biases exist and will continue to exist because the human brain is horizontally scalable, meaning that we can do at once a bunch of simple activities (eating and thinking about the election result for example) that require significant processing power, but we cannot analyze and weigh data points about every single object and reach an objective conclusion about that entity.
A good example of bias is grit. While there is a lot of harping about grit being one of the most important qualities in life, for the average human, and indeed even for an algorithm, grit could also be misinterpreted as an individual having a long streak of non-accomplishment instead of looking at it in a coherent way and analyzing each data point within that streak in an objective manner.
For most humans, this is not only not possible because of pre-existing biases but also because the human brain simply does not have enough time on hand to compute an output. It therefore becomes materially more easy to write about grit than to accept it on face value, as we tend to have a bias toward negative news (i.e. we assume non-accomplishment as the outcome of grit rather than the “grit part” of grit).
The hope with an AI or ANN therefore isn’t to remove biases, for that is mathematically not possible, but rather to have each data point weighed as efficiently as possible so as to find an anomaly that hopefully turns out to be unbiased and impacts an overall decision with or without human instinct. The sooner we come out of our collective utopias, the better it will be for humanity in building solutions that positively impact people’s lives.
Featured Image: Sara D. Davis/Getty Images