The ambiguity problem illustrated:
User: ”Siri, call me an ambulance!”
Siri: ”Okay, I will call you ’an ambulance’.”
You’ll never reach the hospital, and end up bleeding to death.
Two potential solutions come to mind:
A. machine builds general knowledge (”common sense”)
B. machine identifies ambiguity & asks for clarification from humans (reinforcement learning)
The whole ”common sense” problem can be solved by introducing human feedback into the system. We really need to tell the machine what is what, just like a child. This is iterative learning, in which trials and errors take place. However, it is better than trying to adapt an unescapably finite dataset into a close to finite space of meanings.
But, in fact, A and B converge by doing so. Which is fine, and ultimately needed.
To determine which solution to an ambiguous situation is proper, the machine needs contextual awareness; this can be achieved by storing contextual information from each ambiguous situation, and being explained ”why” a particular piece of information results in disambiguity. It’s not enough to say ”you’re wrong”, but there needs to be an explicit association to a reason (concept, variable). Equally, it’s not enough to say ”you’re right”, but again the same association is needed.
1) try something
2) get told it’s not right, and why (linking to contextual information)
3) try something else, corresponding to why
4) get rewarded, if it’s right.
The problem is, currently machines are being trained by data, not by human feedback.
New thinking: Training AI pets
So we would need to build machine-training systems which enable training by direct human feedback, i.e. a new way to teach and communicate with the machine. It’s not a trivial thing, since the whole machine-learning paradigm is based on data, not meanings. From data and probabilities, we would need to move into associations and concepts that capture social reality. A new methodology is needed. Potentially, individuals could train their own AIs like pets (think having your own ”AI pet” like Tamagotchi), or we could use large numbers of crowd workers who would explain the machine why things are how they are (i.e., create associations). A specific type of markup (=communication with the machine) would probably also be needed, although conversational UIs would most likely be the best solution.
Through mimicking human learning we can teach the machine common sense. This is probably the only way; since common sense does not exist beyond human cognition, it can only be learnt from humans. An argument can be made that this is like going back in time, to era where machines followed rule-based programming (as opposed to being data-driven). However, I would argue rule-based learning is much closer to human learning than the current probability-based one, and if we want to teach common sense, we therefore need to adopt the human way.
Machine learning may be at par, but machine training certainly is not. The current machine learning paradigm is data-driven, whereas we could look into ways for concept-driven AI training approaches. Essentially, this is something like reinforcement learning for concept maps.
Introduction. Hal Daumé wrote an interesting blog post about language bias and the black sheep problem. In the post, he defines the problem as follows:
The ”black sheep problem” is that if you were to try to guess what color most sheep were by looking and language data, it would be very difficult for you to conclude that they weren’t almost all black. In English, ”black sheep” outnumbers ”white sheep” about 25:1 (many ”black sheep”s are movie references); in French it’s 3:1; in German it’s 12:1. Some languages get it right; in Korean it’s 1:1.5 in favor of white sheep. This happens with other pairs, too; for example ”white cloud” versus ”red cloud.” In English, red cloud wins 1.1:1 (there’s a famous Sioux named ”Red Cloud”); in Korean, white cloud wins 1.2:1, but four-leaf clover wins 2:1 over three-leaf clover.
Thereafter, Hal accurately points out:
”co-occurance frequencies of words definitely do not reflect co-occurance frequencies of things in the real world.”
But the mistake made by Hal is to assume language describes objective reality (”the real world”). Instead, I would argue that it describes social reality (”the social world”).
Black sheep in social reality. The higher occurence of ’black sheep’ tells us that in social reality, there is a concept called ’black sheep’ which is more common than the concept of white (or any color) sheep. People are using that concept, not to describe sheep, but as an abstract concept in fact describing other people (”she is the black sheep of the family”). Then, we can ask: Why is that? In what contexts is the concept used? And try to teach the machine its proper use through associations of that concept to other contexts (much like we teach kids when saying something is appropriate and when not). As a result, the machine may create a semantic web of abstract concepts which, if not leading to it understanding them, at least helps in guiding its usage of them.
We, the human. That’s assuming we want it to get closer to the meaning of the word in social reality. But we don’t necessarily want to focus on that, at least as a short-term goal. In the short-term, it might be more purposeful to understand that language is a reflection of social reality. This means we, the humans, can understand human societies better through its analysis. Rather than trying to teach machines to imputate data to avoid what we label an undesired state of social reality, we should use the outputs provided by the machine to understand where and why those biases take place. And then we should focus on fixing them. Most likely, technology plays only a minor role in that, although it could be used to encourage balanced view through a recommendation system, for example.
Conclusion. The ”correction of biases” is equivalent to burying your head in the sand: even if they magically disappeared from our models, they would still remain in the social reality, and through the connection of social reality and objective reality, echo in the everyday lives of people.
There is enormous concern about machine learning and AI replacing human workers. However, according to several economists, and also according to past experience ranging back all the way to the industrial revolution of the 18th century (which caused major distress at the time), the replacement of human workers is not permanent but there will be new jobs to replace the replaced jobs (as postulated by the Schumpeterian hypothesis). In this post, I will briefly share some ideas on what jobs are relatively safe from AI, and how should an individual member of the workforce increase his or her chances of being competitive in the job market of the future.
“Insofar as they are economic problems at all, the world’s problems in this generation and the next are problems of scarcity, not of intolerable abundance. The bogeyman of automation consumes worrying capacity that should be saved for real problems . . .” -Herbert Simon, 1966
What jobs are safe from AI?
The ones involving:
- creativity – machine can ”draw” and ”compose” but it can’t develop a business plan.
- interpretation – even in law which is codified in most countries, lawyers use judgment and interpretation. Cannot be replaced as it currently stands.
- transaction costs – robots could conduct a surgery and even evaluate before that if a surgery is needed, but in between you need people to explain things, to prepare the patients, etc. Most service chains require a lot of mobility and communication, i.e. transaction costs, that are to be handled by people.
How to avoid losing your job to AI?
Make sure your skills are complementary to automation, not substitute of it. For example, if you have great copywriting skills, there was actually never a better time to be a marketer, as digital platforms enable you to reach all the audiences with a few clicks. The machine cannot write compelling ads, so your skills are complementary. The increased automation does not reduce the need for creativity; it amplifies it.
If the machine would learn to be creative in a meaningful way (which is far far away, realistically speaking), then you’d do some other complementary task.
The point is: there is always some part of the process you can complement.
Fear not. Machines will not take all human jobs because not all human jobs exist yet. Machines and software will take care of some parts of service chains, even to a great extent but in fact that will enhance the functioning of the whole chain, and also that of human labor (consider the amplification example of online copywriting). New jobs that we still cannot vision will be created, as needs and human imagination keep evolving.
The answer is in creative destruction: People won’t stop coming up with things to offer because of machines. And other people won’t stop wanting those things because of machines. Jobs will remain also in the era of AI. The key is not to complain about someone taking your job, but to think of other things to offer, and develop your personal competences accordingly. Even if you won’t, the next guy will. There’s no stopping creativity.
- Scherer, F. M. (1986). Innovation and Growth: Schumpeterian Perspectives (MIT Press Books). The MIT Press.
- David, H. (2015). Why are there still so many jobs? The history and future of workplace automation. The Journal of Economic Perspectives, 29(3), 3–30.
Ethics of machine learning algorithms has recently been raised as a major research concern. Earlier this year (2017), a fund of $27M USD was started to support research on the societal challenges of AI. The group responsible for the fund includes e.g. the Knight Foundation, Omidyar Network and the startup founder and investor Reid Hoffman.
As stated on the fund’s website, the fund will support a cross-section of AI ethics and governance projects and activities, both in the United States and internationally. They advocate cross-disciplinary research between e.g. computer scientists, social scientists, ethicists, philosophers, economists, lawyers and policymakers.
The fund lays out a list of areas they’re interested in funding. The list can be seen as a sort of a research agenda. The items are:
- Communicating complexity: How do we best communicate, through words and processes, the nuances of a complex field like AI?
- Ethical design: How do we build and design technologies that consider ethical frameworks and moral values as central features of technological innovation?
- Advancing accountable and fair AI: What kinds of controls do we need to minimize AI’s potential harm to society and maximize its benefits?
- Innovation in the public interest: How do we maintain the ability of engineers and entrepreneurs to innovate, create and profit, while ensuring that society is informed and that the work integrates public interest perspectives?
- Expanding the table: How do we grow the field to ensure that a range of constituencies are involved with building the tools and analyzing social impact?
As can be seen, the agenda emphasizes the big question: How can we maintain the benefits of the new technologies while making sure that their potential harm is minimized? To answer this question, a host of studies and perspectives is definitely needed. Read here a list of other initiatives working on the societal issues of AI and machine learning.
Feature analysis could be employed for bias detection when evaluating the procedural fairness of algorithms. (This is an alternative to the ”Google approach” which emphasis evaluation of outcome fairness.)
In brief, feature analysis reveals how well each feature (=variable) influenced the model’s decision. For example, see the following quote from Huang et al. (2014, p. 240):
”All features do not contribute equally to the classification model. In many cases, the majority of the features contribute little to the classifier and only a small set of discriminative features end up being used. (…) The relative depth of a feature used as a decision node in a tree can be used to assess the importance of the feature. Here, we use the expected fraction of samples each feature contributes to as an estimate of the importance of the feature. By averaging all expected fraction rates over all trees in our trained model, we could estimate the importance for each feature. It is important to note that feature spaces among our selected features are very diverse. The impact of the individual features from a small feature space might not beat the impact of all the aggregate features from a large feature space. So apart from simply summing up all feature spaces within a feature (i.e. sum of all 7, 057 importance scores in hashtag feature), which is referred to as un-normalized in Figure 4, we also plot the normalized relative importance of each features, where each feature’s importance score is normalized by the size of the feature space.”
They go on to visualize the impact of each feature (see Figure 1).
Figure 1 Feature analysis example (Huang et al., 2014)
As you can see, this approach seems excellent for probing the impact of each feature on the model’s decision making. The impact of sensitive features, such as ethnicity, can be detected. Although this approach may be useful for supervised machine learning, where the data is clearly labelled, the applicability to unsupervised learning might be a different story.
In one of the workshops of the first conference day, ”Studying User Perceptions and Experiences with Algorithms”, the participants recommended papers to each other. Here are, if not all, then most of them, along with their abstracts.
Exposure to news, opinion, and civic information increasingly occurs through social media. How do these online networks influence exposure to perspectives that cut across ideological lines? Using deidentified data, we examined how 10.1 million U.S. Facebook users interact with socially shared news. We directly measured ideological homophily in friend networks and examined the extent to which heterogeneous friends could potentially expose individuals to cross-cutting content. We then quantified the extent to which individuals encounter comparatively more or less diverse content while interacting via Facebook’s algorithmically ranked News Feed and further studied users’ choices to click through to ideologically discordant content. Compared with algorithmic ranking, individuals’ choices played a stronger role in limiting exposure to cross-cutting content.
This article reflects the kinds of situations and spaces where people and algorithms meet. In what situations do people become aware of algorithms? How do they experience and make sense of these algorithms, given their often hidden and invisible nature? To what extent does an awareness of algorithms affect people’s use of these platforms, if at all? To help answer these questions, this article examines people’s personal stories about the Facebook algorithm through tweets and interviews with 25 ordinary users. To understand the spaces where people and algorithms meet, this article develops the notion of the algorithmic imaginary. It is argued that the algorithmic imaginary – ways of thinking about what algorithms are, what they should be and how they function – is not just productive of different moods and sensations but plays a generative role in moulding the Facebook algorithm itself. Examining how algorithms make people feel, then, seems crucial if we want to understand their social power.
Eslami, M., Rickman, A., Vaccaro, K., Aleyasen, A., Vuong, A., Karahalios, K., … Sandvig, C. (2015). “I Always Assumed That I Wasn’T Really That Close to [Her]”: Reasoning About Invisible Algorithms in News Feeds. In Proceedings of the 33rd Annual ACM Conference on Human Factors in Computing Systems (pp. 153–162). New York, NY, USA: ACM.
Our daily digital life is full of algorithmically selected content such as social media feeds, recommendations and personalized search results. These algorithms have great power to shape users’ experiences, yet users are often unaware of their presence. Whether it is useful to give users insight into these algorithms’ existence or functionality and how such insight might affect their experience are open questions. To address them, we conducted a user study with 40 Facebook users to examine their perceptions of the Facebook News Feed curation algorithm. Surprisingly, more than half of the participants (62.5%) were not aware of the News Feed curation algorithm’s existence at all. Initial reactions for these previously unaware participants were surprise and anger. We developed a system, FeedVis, to reveal the difference between the algorithmically curated and an unadulterated News Feed to users, and used it to study how users perceive this difference. Participants were most upset when close friends and family were not shown in their feeds. We also found participants often attributed missing stories to their friends’ decisions to exclude them rather than to Facebook News Feed algorithm. By the end of the study, however, participants were mostly satisfied with the content on their feeds. Following up with participants two to six months after the study, we found that for most, satisfaction levels remained similar before and after becoming aware of the algorithm’s presence, however, algorithmic awareness led to more active engagement with Facebook and bolstered overall feelings of control on the site.
Epstein, R., & Robertson, R. E. (2015). The search engine manipulation effect (SEME) and its possible impact on the outcomes of elections. Proceedings of the National Academy of Sciences, 112(33), E4512–E4521.
Internet search rankings have a significant impact on consumer choices, mainly because users trust and choose higher-ranked results more than lower-ranked results. Given the apparent power of search rankings, we asked whether they could be manipulated to alter the preferences of undecided voters in democratic elections. Here we report the results of five relevant double-blind, randomized controlled experiments, using a total of 4,556 undecided voters representing diverse demographic characteristics of the voting populations of the United States and India. The fifth experiment is especially notable in that it was conducted with eligible voters throughout India in the midst of India’s 2014 Lok Sabha elections just before the final votes were cast. The results of these experiments demonstrate that (i) biased search rankings can shift the voting preferences of undecided voters by 20% or more, (ii) the shift can be much higher in some demographic groups, and (iii) search ranking bias can be masked so that people show no awareness of the manipulation. We call this type of influence, which might be applicable to a variety of attitudes and beliefs, the search engine manipulation effect. Given that many elections are won by small margins, our results suggest that a search engine company has the power to influence the results of a substantial number of elections with impunity. The impact of such manipulations would be especially large in countries dominated by a single search engine company.
Algorithms (particularly those embedded in search engines, social media platforms, recommendation systems, and information databases) play an increasingly important role in selecting what information is considered most relevant to us, a crucial feature of our participation in public life. As we have embraced computational tools as our primary media of expression, we are subjecting human discourse and knowledge to the procedural logics that undergird computation. What we need is an interrogation of algorithms as a key feature of our information ecosystem, and of the cultural forms emerging in their shadows, with a close attention to where and in what ways the introduction of algorithms into human knowledge practices may have political ramifications. This essay is a conceptual map to do just that. It proposes a sociological analysis that does not conceive of algorithms as abstract, technical achievements, but suggests how to unpack the warm human and institutional choices that lie behind them, to see how algorithms are called into being by, enlisted as part of, and negotiated around collective efforts to know and be known.
In information societies, operations, decisions and choices previously left to humans are increasingly delegated to algorithms, which may advise, if not decide, about how data should be interpreted and what actions should be taken as a result. More and more often, algorithms mediate social processes, business transactions, governmental decisions, and how we perceive, understand, and interact among ourselves and with the environment. Gaps between the design and operation of algorithms and our understanding of their ethical implications can have severe consequences affecting individuals as well as groups and whole societies. This paper makes three contributions to clarify the ethical importance of algorithmic mediation. It provides a prescriptive map to organise the debate. It reviews the current discussion of ethical aspects of algorithms. And it assesses the available literature in order to identify areas requiring further work to develop the ethics of algorithms.
Every day, corporations are connecting the dots about our personal behaviorsilently scrutinizing clues left behind by our work habits and Internet use. The data compiled and portraits created are incredibly detailed, to the point of being invasive. But who connects the dots about what firms are doing with this information? The Black Box Society argues that we all need to be able to do soand to set limits on how big data affects our lives. Hidden algorithms can make (or ruin) reputations, decide the destiny of entrepreneurs, or even devastate an entire economy. Shrouded in secrecy and complexity, decisions at major Silicon Valley and Wall Street firms were long assumed to be neutral and technical. But leaks, whistleblowers, and legal disputes have shed new light on automated judgment. Self-serving and reckless behavior is surprisingly common, and easy to hide in code protected by legal and real secrecy. Even after billions of dollars of fines have been levied, underfunded regulators may have only scratched the surface of this troubling behavior. Frank Pasquale exposes how powerful interests abuse secrecy for profit and explains ways to rein them in. Demanding transparency is only the first step. An intelligible society would assure that key decisions of its most important firms are fair, nondiscriminatory, and open to criticism. Silicon Valley and Wall Street need to accept as much accountability as they impose on others.
There is a dearth of research on the public’s beliefs about how social media technologies work. To help address this gap, this article presents the results of an exploratory survey that probes user and non-user beliefs about the techno-cultural and socioeconomic facets of Twitter. While many users are well-versed in producing and consuming information on Twitter, and understand Twitter makes money through advertising, the analysis reveals gaps in users’ understandings of the following: what other Twitter users can see or send, the kinds of user data Twitter collects through third parties, Twitter and Twitter partners’ commodification of user-generated content, and what happens to Tweets in the long term. This article suggests the concept of “information flow solipsism” as a way of describing the resulting subjective belief structure. The article discusses implications information flow solipsism has for users’ abilities to make purposeful and meaningful choices about the use and governance of social media spaces, to evaluate the information contained in these spaces, to understand how content users create is utilized by others in the short and long term, and to conceptualize what information other users experience.
Woolley, S. C., & Howard, P. N. (2016). Automation, Algorithms, and Politics| Political Communication, Computational Propaganda, and Autonomous Agents — Introduction. International Journal of Communication, 10(0), 9.
The Internet certainly disrupted our understanding of what communication can be, who does it, how, and to what effect. What constitutes the Internet has always been an evolving suite of technologies and a dynamic set of social norms, rules, and patterns of use. But the shape and character of digital communications are shifting again—the browser is no longer the primary means by which most people encounter information infrastructure. The bulk of digital communications are no longer between people but between devices, about people, over the Internet of things. Political actors make use of technological proxies in the form of proprietary algorithms and semiautomated social actors—political bots—in subtle attempts to manipulate public opinion. These tools are scaffolding for human control, but the way they work to afford such control over interaction and organization can be unpredictable, even to those who build them. So to understand contemporary political communication—and modern communication broadly—we must now investigate the politics of algorithms and automation.