David Zen's Blog

A growing area of focus of my AI-related activities is how machines can learn directly from humans (rather than just from data), and conversely how we can better understand human behavior from AI behavior. I call this "human-machine learning". Such interactions will be a key part of the evolution from current (impressive but narrow) achievements of machine learning to artificial general intelligence.

This blog post is a call for collaboration on a demonstration project in this area.

OpenAI "Debate" Proposal

OpenAI has recently advanced the thinking in this area in a fascinating blog post AI Safety via Debate, which proposes using human evaluation of debates between AI agents to better understand the AIs' functions. After proposing this general model, the blog states that "today’s natural language modeling is insufficient for this task".

So to provide a concrete example, the blog describes a preliminary experiment involving guessing (by a machine classifier "judge", not a human) of which handwritten numeric digit (0 through 9, from the widely-used MNIST database) is depicted, based upon only six non-black pixels (out of 784 pixels in each image). The pixels are presented by two other machine agents -- one trying to convey the real digit to the human, and the other trying to convey a different digit to the human (but using only pixels from the real digit).

The result of this experiment was that the judge was able to guess the right digit 88.9% of the time (between the two choices), significantly better than the 59.4% success rate of the classifier from six pixels without the "debate" between the agents (although it's not entirely clear from the blog post whether the judge in the latter case was choosing among 10 or only 2 digits, or how the pixels were selected).

BayesMaze Idea

Reading about the OpenAI experiment prompted me to seek crowdsourced help for a different demonstration of human-machine learning that I have been developing for the past couple of months. I call the idea BayesMaze -- "Bayes" comes from my interest in Bayesian probabilistic methods for machine learning (including because human interactions inherently involve multiple dimensions of uncertainty); "Maze" will be self-evident.

The goal of the experiment is for humans to teach a machine agent (a) to understand natural language instructions on how to navigate a maze and (b) to be able to give such instructions to a human. The key constraints are:

1. The human and machine are working together in a defined environment with a defined goal.

2. At the start of the experiment, the machine agent (like a human baby) has no knowledge of human language, but would be able to react to sentiment in human language. (Without this constraint, the problem is trivial -- it would be easy to build a chatbot with knowledge of language that can explain or understand how to navigate a maze.)

This experiment avoids two of the constraints of the OpenAI "debate" experiment by (a) using natural language and (b) involving humans in the learning process.

I do not claim that this idea has any obvious immediate practical application. And I know there is plenty of ongoing work on how chatbots can learn from humans (for example, here and here). But I believe that substituting a machine agent into a (simple) model of the basic human language learning process may produce interesting insights about machine and human learning.

If you are interested in getting involved (or have any other comments), please reach out! The current version of the specification for the project is below.

BayesMaze Specification

BayesMaze will operate via a publicly-accessible website, on which users can choose either (1) to instruct a machine agent how to navigate a maze or (2) let the machine agent instruct them how to navigate a maze. The logical flow for each of these options is set out below (pseudocode is available upon request, and I am moving slowly towards implementation in Python.)

A. Human Instructs Machine
1. Generate random maze including start point and goal (free resources include Mazify and Maze Generator)
2. Set position to start
3. Human gives command C in natural language
4. Update training of word embeddings for C (likely using word2vec)
5. Generate computer move based upon maze position and structure (portion seen by computer), word2vec embeddings and human command (likely RNN with 4 element output layer for 4 possible directions)
6. Human evaluates move in natural language
7. Evaluate sentiment of human evaluation (likely using NLTK)
8. Update training of RNN move generator based upon command, move, sentiment and whether goal has been reached
9. If goal not reached, return to step 3
10. Output closing message (e.g. number of steps taken)

B. Machine Instructs Human
1. Generate random maze including start point and goal (show start point but not rest of maze)
2. Set position to start
3. Machine calculates optimum move toward goal, out of 4 directions (use free resources)
4. Machine chooses command from library of commands (gathered from human instructing machine above) based upon best sentiment association with chosen move (plus Gaussian noise)
5. (optional) Chosen command evaluated for abusiveness (to avoid human manipulation), likely using sentiment analysis or similar tool
6. Machine gives command in natural language
7. Human moves
8. Machine evaluates human move based upon responsiveness to machine command and progress towards goal
9. Output machine evaluation of move
10. If goal not reached, return to step 3
11. Output closing message

I am worried by the increasing inability of Americans (and people in other Western countries -- I live in London) to engage in calm, productive dialogue with those of different views. Bret Stephens of the New York Times took a good look at this problem this week in an op-ed "The Dying Art of Disagreement".

After quite a bit of thinking about what to do about this difficult problem, my main conclusion is that we all need to take small steps back to civility. And I recently developed a personal manifesto for my own steps in that direction ("Otter" is my nickname at home).

The Otter Manifesto

1. All people deserve respect and kindness by default. Incivility and bad behaviour should be discouraged. These two principles can conflict, and when they do my bias is towards respect and kindness.

2. Reasoned debate is better than divisiveness and anger. I commit always to talk politely with those of different views — not with the goal of convincing them of my point of view, but to promote mutual understanding and trust.

Taking this manifesto seriously has already had some significant positive effects in my life.

Almost 3 1/2 years ago, I began building a portfolio of investments in UK start-ups that use artificial intelligence (AI) -- or, if you prefer, machine intelligence (MI) -- to solve business problems (see portfolio companies below). This investment approach is getting a lot of attention these days, including from UK venture capital funds like Octopus Partners and Forward Partners. I like to think I was early to the game.

These investments have led me to pay a lot of attention to the business and science of AI (on the latter, check out my pinboard on AI for text analysis, on the platform of my investee Sparrho). A recent insight from this effort -- and the reason for the quotes around "intelligence" in the title of this blog -- is that it is misplaced (at least from a business perspective) to think about this opportunity in terms of creating human-like intelligence. Much is written about design for human-like AIs that can meet the Turing test or lead to the "singularity". But the real advantage to business from AI/MI is that it provides tools to do things faster or better than humans (and not necessarily by taking our jobs), rather than exhibiting "intelligence".

The start-ups in which I have invested use AI/MI ranging from quite sophisticated AI techniques to now run-of-the-mill processing of big data:

Qlearsite. AI and data visualization for human resources.

Sparrho. Science search and recommendation using AI.

Lexoo. Finding lawyers more efficiently and cost-effectively.

Contego. Customized management of big data for regulatory compliance.

Insurance. An insurance technology start-up still in stealth mode.

A common thread is that all of these companies are focused on delivering a business solution rather than building the smartest AI. And the smart AI/MI building blocks that they need to do this are readily available for free from big players like Google / Deep Mind, Facebook, Amazon, IBM Watson and OpenAI, and the many university faculty and students who are researching and writing about AI. Imitating human intelligence is a good field for researchers, futurists and science fiction writers, but it has very little to do with successful investment.

It is early days for this portfolio (no exits yet), but (touch wood) things are looking promising!

I recently came across The Happiest Kids in the World: Bringing up Children the Dutch Way, which provides persuasive evidence that parenting is done very well indeed in the Netherlands. In the US, family life seems to be breaking down.

And then I read today that Dutch crime rates are so low (and prison sentences so short) that prisons are being closed. The US has by far the largest prison population in the world, with China, Russia and Brazil some distance behind.

Could there be a connection?

I love this "America First, Netherlands Second" video. But maybe it should be the other way around?

As a US/UK dual citizen, I have been immersed in the news of the two surprise votes of 2016 -- for Brexit and for Donald Trump. What’s genuinely disturbing is the increasing polarization of rhetoric and positions in both US and UK society. There is an urgent need to reverse this trend of polarization.

People should speak out for what they believe in. The issues that divide the UK and the US are real issues. Broad agreement on these issues will not be easy to find. But discussions of the issues must not routinely descend into acrimony.

I often have productive discussions with people with whom I do not agree. I learn things. We find common ground, and identify the reasons for our differences, which are often based in differences of experience. Even where our values turn out to be fundamentally different, I still usually gain understanding and, with understanding, greater skill in making my views heard.

For those now engaged in discussion of the issues that challenge our society:

1. Listen to those with different points of view, and respond with conviction but never anger.

2. Seek common ground, remembering that we share a future and that win-win outcomes are often possible.

In my lifetime, I have seen some apparently intractable political conflicts improve markedly -- notably in South Africa (where apartheid was ended by the mutual courage of Nelson Mandela and F. W. de Klerk) and in Northern Ireland (where once bitter enemies Ian Paisley and Martin McGuiness cooperated to build a shared peace).

This evening, I walked home with my wife from a London pub dinner, and we played a game of discussing divisive issues of the day, taking turns with one of us right-wing and one of us left-wing on each issue. It was instructive to feel that when one is arguing a disliked point of view, the views one expresses may seem wrong, but not evil or deserving of scorn. I highly recommend it.

My own desire for finding common ground is based in my own study of Eastern philosophy and religion -- Buddhism, Taoism and Confucianism. This approach is not for everyone. Many tell me that they feel a need to be (and some say that I should be) more adamant in taking on those who are "in the wrong". Perhaps. But hard times demand tough choices. I choose a determined search for common ground to build our shared future.

Much has recently been written and said about the filter bubble -- i.e. the phenomenon that online algorithms designed to show us items of interest have the perverse effect of insulating us from alternate points of view.

I recently had a simple idea: machine learning can be used to insulate us from opposing ideas, and it could also be used to introduce us to opposing ideas. My nascent concept for implementing this idea is a website called The Opposite View, which learns the preferences of a subscriber (e.g. by being authorized to follow their social media feeds), and then displays content favored by people who have "opposite" views (what "opposite" means would require some thinking, and could be adjustable by the user).

Such a website would of course only be of interest to those who really are interesting in learning views that differ from their own. But it could make an interesting contribution to challenging the filter bubble. Anyone interesting in taking this idea further?

It is a commonplace that life is uncertain. Blaise Pascal said it well more than 350 years ago: "We sail within a vast sphere, ever drifting in uncertainty, driven from end to end."

In a world where nothing is certain, one might ask, how can we act with conviction? But this question misses the point, by putting uncertainty and conviction in opposition. By conviction I mean confidently taking a course of action under the risk of being wrong. As I recently realized with some astonishment, this kind of conviction marks the path to meaning and success in an uncertain world.

This realization came to me after an uncertain situation that produced two opposing outcomes, and got me thinking about the technical, emotional and creative dimensions of the interaction between uncertainty and conviction.

An Opportunity To See Both Outcomes. The situation was a personal matter in which someone close to me wanted to take an action with which I disagreed. I stated my objections calmly and forcefully, but declined to take an available opportunity to enforce my preferred approach, because I accepted that my friend's point of view was not unreasonable in an uncertain situation. And indeed my friend went ahead and took her preferred approach, with which I disagreed.

But then she changed her mind, basically for the reasons that I had pressed. This was a great outcome, much better than it would have been to enforce my approach. And it was fundamentally because I accepted that the right course of action was uncertain, while defending my point of view with conviction.

The Technical Dimension: Risk Management. One simple reason that conviction helps under conditions of uncertainty is that a firm decision can produce an improved outcome. I am an avid sailor, and I often say that when sailing towards a rock, it may be difficult to know whether turning right or left is better, but either is preferable to no decision at all. More generally, this is about the distinction between uncertainty (which is the unknown itself) and risk (which depends upon how we manage the unknown). There is an excellent explanation of this distinction by Kelvin Stott here.

The Emotional Dimension: Confidence Is Self-Fulfilling. Conviction under uncertainty becomes more interesting and complex when we realize that honest and confident conviction is often self-fulfilling, because it builds conviction in others. Conviction is not a zero sum game. This was a primary effect in the incident with my friend described above. She was ultimately convinced (at least in part) by my conviction. Another example, about which I have written before, is the difference between anger and wrath, where the former is negative and damaging, and the latter involves constructive conviction.

The Creative Dimension: Meaning From Zero. Most fundamentally, conviction is about creating meaning where there is none. As British philosopher A C Grayling memorably said at a lecture I attended at St George's House in Windsor Castle, the meaning of life is obvious -- it is whatever meaning we choose to give to it. And therein lies the challenge: creating meaning where all is uncertain. As Peter Thiel has written in Zero to One: Notes on Startups, or How to Build the Future, starting with nothing (i.e. zero) and creating something new (i.e. one) is the fundamental challenge in entrepreneurship -- and I would say more broadly in life. Doing so usually requires great conviction.

I believe in democracy. I grew up in the United States, where democracy is practically a religion. And I believe in the tradition of democracy, as it developed from roots in the ancient world (especially Athens, according to many western historians) and helped form modern republican traditions like those of my native United States and current home the United Kingdom.

But I have started to wonder about whether democratic government is always the best route in the modern world - or more precisely, whether governments that we in the West call undemocratic deserve to be criticized on that grounds.

Enlightened Absolutism. Might there be a role for the "enlightened absolutism" that was championed by European monarchs of the 18th century and French philosopher Voltaire? As Encyclopaedia Britannica defines it, monarchs advocating this form of government "typically instituted administrative reform, religious toleration, and economic development but did not propose reforms that would undermine their sovereignty or disrupt the social order."

Belarus. I started asking this question because, in the past couple of years, I have spent a lot of time in Belarus, whose President Alexander Lukashenko has been called "Europe's last dictator", resulting in EU and US sanctions against the country. What I have learned, though, is that Belarus is a peaceful, stable country with very little visible poverty (the poverty rate is one of the lowest in Europe). The country as a whole remains of modest means, including because of Lukashenko's failure to embrace western capitalism, but it seems a rather contented lack of means. I am not an apologist for Lukashenko's human rights violations (including locking up political opponents - most of whom he recently freed), but Belarus does not seem to be suffering from lack of democracy. Compare the situation in Ukraine, which has been literally torn about by elections and shifting power. Maybe that's the right path for Ukraine, but I don't think one can say it's obviously a better path for the country's people than the one taken in Belarus.

China. Moving to a bigger stage, China (where much of my investment work is focused) also has an undemocratic government. Watching the impressive ongoing progress of the Chinese economy (which continues notwithstanding the recent slowdown of growth), I wonder whether these achievements would have been possible without the ability of the Communist Party to make difficult decisions about economic change without the threat of real political opposition (although some commentators have suggested that central control has been a poor approach to the recent turmoil in Chinese financial markets).

The Chinese Communist Party claims that it is moving the country towards "socialist democracy", but this means something very different from Western democracy. At the end of a recent event I attended at the Chinese embassy in London, I was handed a copy of a book What Kind of Democracy Do Chines People Want by Zhang Minshu, which articulates a less individualistic concept of democracy than the Western one.

Free Speech and Human Rights. The biggest problem in my view with absolutist regimes like those in China and Belarus are their weakness in permitting free speech, and often other human rights. The systems in these countries are far more totalitarian than the enlightened absolutist regimes of the 18th century, and have powerful and aggressive state mechanisms to suppress dissent. In the era of the Internet, these regimes perceive (with some justification) that such control is necessary to maintain stability. I recently attended the Ai Wei Wei exhibit at London's Royal Academy of Art, and Ai makes powerful arguments that the resolute and sometimes brutal Chinese suppression of unpopular views is bad for the future of the country.

United States. So how do these modern absolutist regimes stack up against that bastion of democracy, the United States? Not too badly, in my view. US "democracy" has disintegrated into a pitched battle between Republicans and Democrats, with little room for reasonable middle ground, and the final choice of the president being made in a small number of states that have the possibility of going to either party. The country is unable to control guns, and seems to be becoming increasingly xenophobic. A surprising number of people are buying the angry, selfish messages of Donald Trump (who suggested excluding Muslims from the country entirely) and Ted Cruz. This is not the kind of democracy that I want.

In sum, I remain a believer in democracy. But I am far from sure that systems of government meeting the western definition of "democracy" are the only reasonable choice in the modern world. There appear to be other approaches that can be considered to promote human opportunity and human rights.

I was delighted this week to sign up with Cassian Hall at DHH Literary Agency to represent my writing globally. DHH are a well-regarded smaller agency, and seem like an excellent choice for a new author like me.

I also attended this week at Harvard Club of the UK event on "Earning a Living in the Writing Business". Along with two successful authors, my friend Mildred Yuan of United Agents was on the panel. Mildred emphasised how difficult it is in the current publishing environment for publishers and agents to make money off new authors, leading to a focus on writers who are either (a) established or (b) already famous for something else. I feel fortunate to have found an agency that is willing to take a risk on my work -- hopefully we will prove together that this was an excellent decision.

As a US expatriate living in London for nearly 15 years (and I love living here, and most things about this country), I have never quite gotten over the frequent unwillingness of some customer-facing personnel at UK businesses to respond to reasonable service requests. In the past year, though, I have discovered a remarkably effective solution. Just make a polite and truthful complaint, through a formal process or on social media.

I have gotten these results in the past year with this approach:
- A large UK bank on two different problems (one on a personal account and one on a business account) had been frustratingly unresponsive. After I insisted on registering formal complaints (a simple process), each problem was addressed reasonably quickly, and I was offered compensation of £100 for one problem and £50 for the other.
- I bought printer cartridges in a 2-for-1 deal from a high-street stationery retailer, and the second one did not work at all (after the first one ran out). Even though I had kept a receipt, the shop would not exchange the useless cartridge. I tweeted about the problem, and was promptly offered a free cartridge by post, and an apology.
- Our neighbours have an ongoing construction project, and the builders have been very noisy, missed deadlines and been generally difficult, despite repeated requests. Again, one tweet mentioning the builder and the problems was sufficient to prompt a quick call from the company's commercial director, an apology, and a promise of better conduct and communication.

I don't know why complaints seem to work so well in the UK when polite requests for action do not. Maybe it's about fear -- people here don't perceive an imperative to act until there are apparent consequences of inaction. From my recollection , the situation is somewhat different in the US, where a customer with a reasonable grievance usually does fairly well by contacting customer service (and complaints are much more rarely necessary). An interesting cultural difference.

So for all you Britons who have just gotten used to the sometimes appalling service in the UK (of course there are many UK service people who do a good job), my new year advice to you is to consider a polite complaint when other approaches have failed.