Reproducible, Reusable, and Robust Reinforcement Learning (Professor Joelle Pineau)
I was sad to miss this talk, but lots of people told me it was great so I’ve bookmarked it to watch later.
Investigations into the Human-AI Trust Phenomenon (Professor Ayanna Howard)
A few interesting points in this talk:
Children are unpredictable, and working with data generated from experiments with children can be hard. (for example, children will often try to win at games in unexpected ways)
Automated vision isn’t perfect, but in many cases it’s better than the existing baseline (having a human record certain measurements) and can be very useful.
Having robots show sadness or disappointment turns out to be much more effective than anger for changing a child’s behaviour.
Humans seem to inherently trust robots!
Two cool experiments:
To what extent would people trust a robot in a high-stakes emergency situation?
A research subject is led into a room by a robot, for a yet-to-be-defined experiment. The room fills with smoke, and the subject goes to leave the building. On the way out, the same robot is indicating a direction. Will the subject follow the robot?
It turns out that, yes, almost all of the time, the subject follows the robot.
What about if the robot, when initially leading the subject to the room, makes a mistake and has to be corrected by another human?
Again, surprisingly, the subject still follows the robot in the emergency situation that follows.
The point at which this stopped being true was when they had the robot point to obviously wrong directions (e.g. asking the subject to climb over some furniture).
This research has some interesting conclusions, but I’m not completely convinced. For one, based on the videos of the ’emergency situation’, it seems unlikely that any of the subjects would have believed the emergency situation to be genuine. The smoke is extremely localised, and the ‘exit’ just leads into another room. It seems far more likely to me that the subjects were trying to infer the researchers’ intentions for the study and decided to follow the robot since that was probably what they were meant to do.
Unfortunately this turns out to be because their ethics committee wouldn’t give them approval to run a more realistic version of the study in a separate building, which is a real shame. More on this study in the the paper, or in this longer write-up.
Does racism and sexism apply to people’s perception of robots?
Say we program a robot to perform a certain sequence of behaviours, and then ask someone to interpret the robot’s intention behind that behaviour. Will their interpretation be affected by a robot’s ‘race’ or ‘gender’?
It turns out that, yes, it is. For example, when primed to be aware of the ‘race’ of a robot, subjects are more likely to interpret the behaviour as angry when the robot is black than when the robot is white.
But when humans are not primed to pay attention to race (and just shown robots with different colours), the effect disappears. Paper here.
I just finished reading Influence: The Psychology of Persuasion, and found it fascinating. It’s like an applied version of Thinking, Fast and Slow, with lots of examples on how cognitive biases are exploited and how you can protect yourself against each of them.
I picked it up because it’s recommended by Derek Sivers as one of the top books on sales and marketing. There’s some great content in there, and as I found myself sharing my notes with a few friends I decided to put them up here for everyone to read. I don’t think the below summary is a replacement for reading the book, but it should give you a good idea of what the main ideas are and whether it’s the kind of book you’d like to read.
Derek Sivers also has his own notes on the book here. I’d highly recommend buying and reading the full book though, for £6/$8 it’s a great investment!
Some animal behaviours are completely governed by reactions to simple triggers. E.g. mother turkeys will care for anything that cheeps, but won’t care for a baby turkey that doesn’t. This can also apply to humans.
The contrast principle. When two different things are shown one after the other, the differentiating characteristics of the second will appear exaggerated. So if you’re selling a suit and a jumper to the same person, sell the suit first. Then the jumper will appear cheaper. This is related to the concept of anchoring.
Human societies have developed an evolutionary urge to reciprocate. This means that by giving someone a gift/doing them a favour, you can generate goodwill and potentially a return gesture. This can work regardless of whether the initial gesture was invited or not, and even if it was unwanted.
The reciprocation rule also applies to concessions — so if you ask someone for a big favour and they say no and you concede by asking them for a smaller favour, they are then likely to concede and say yes to that smaller favour.
Arrangements agreed with perceived concession generate increased feelings of responsibility (the opponent feels that he influenced the negotiation) for and satisfaction with the outcome.
The key to resisting these tactics, if used on you, is to recognise them for what they are. There’s no need to refuse all favours, but if you recognise a favour as a sales tactic then respond to it not as a favour, but as a sales tactic. There’s no need to reciprocate to a sales tactic.
Commitment and Consistency
The drive to be (and look) consistent is a highly potent weapon of social influence, often causing us to act in ways that are clearly contrary to our own best interests.
One way toy stores use this: run ads for amazing toys before Christmas. Undersupply stores with those toys so parents have to buy substitutes. After Christmas, increase supply again so parents who promised their kids they’d buy those toys and want to remain consistent end up buying Christmas toys twice.
Commitments are most effective in changing a person’s self-image and future behaviour when they are active, public, and effortful. But even more importantly, the person must think they have chosen to commit in the absence of strong outside pressures.
This suggests that we should never heavily bribe or threaten children to do the things we want them truly to believe in. If we want them to believe in the correctness of what they have done, then we must somehow arrange for them to accept inner responsibility for the actions we want them to take.
Lowballing: an advantage is offered that induces a favoured behaviour or decision. The subject justifies this decision to themselves by changing their views to fit the decision. Then the advantage is taken away, and the behaviour/decision is fully supported by the subject’s new views.
Two ways to fight back against opponents attempting to use your need for consistency against your best interests:
1. If you get that weird feeling in your stomach and realise what is happening, call them out on it. Say that you don’t want to continue purely for the sake of consistency.
2. If you’re not sure what you really believe, ask yourself and pay special attention to your immediate instinctive/emotional response. You can lie to yourself and rationalise things when thinking intellectually, but not as easily in these basic responses.
People use others’ opinion as another shortcut to figuring out the truth. This sometimes even works when something mimics others’ opinion even though we know it’s fake — e.g. laughter tracks make things seem funnier.
The pluralistic ignorance effect or bystander effect happens under uncertainty, when the social cues that everything is fine overcome the concern that there might be an emergency.
In an emergency, ask for help clearly. Direct your request at one person at a time. Your best strategy when in need of emergency help is to reduce the uncertainties of those around you concerning your condition and their responsibilities.
1978 Jonestown, Guyana. Jim Jones, leader of The People’s Temple, instructs his followers to commit mass suicide. 910 people did so. This was after, a year earlier, the entire community moved from San Francisco to the jungle in Guyana.
Surrounded by uncertainty, people look to the actions of others to guide their own actions.
Be wary of situations where social feedback is faked (laughter tracks, ads,…) and consciously disengage your social autopilot and examine the evidence independently. Look up and around periodically whenever locked onto the evidence of the crowd.
People are more likely to buy from or follow others they know or like, as exploited by Tupperware parties. Separately, the halo effect induces us to imbue attractive others with unrelated positive attributes by assumption.
Physical attractiveness, similarity, compliments, contact and cooperation (familiarity), conditioning and association can all influence how much we like someone.
Kids away at camp. To increase hostility, separate them physically in different areas and give the separate groups names. Then put them in competition with one another. To increase harmony, construct situations where they have to cooperate, and where competition would be harmful to both sides.
Shakespeare: “The nature of bad news infects the teller.” A lot of strange behaviour can be explained by the fact that people understand the association principle well enough to strive to link themselves to positive events and separate themselves from negative events — even when they have not caused the events.
When talking about wins, sports fans often use “we” to associate themselves with success. When talking about losses they’ll often use “they” to distance themselves — even though it’s clear that that there is no causation in either case. People with a weakened self-image will feel a stronger need to do this.
Countering the liking strategy: it is too difficult to prevent ourselves from liking someone, but we must be aware of it when making decisions. In making a compliance decision, it is always a good idea to keep separate our feelings about the requester and the request.
Milgram experiment, 1965. ‘Teacher’ subjects continued to administer increasingly painful shocks under instruction of the researcher, even when the subject was clearly in agony and begged for them to stop.
Obedience to authority is another mental shortcut we use because in general it pays off — authority figures (teachers when we’re kids, doctors/lawyers/professionals when we’re adults) often have more information than we do. But this reflex can be abused by others taking advantage of their authority status or masquerading with fake status.
Again the pretence works as well as the real thing — even actors playing doctors are seen as having authority.
Titles, clothes, trappings — not only do these have great effect on our actions, we consistently underestimate the effect they have on our actions.
Countering abuse of authority — ask two questions:
Is this authority truly an expert?
How truthful can we expect the expert to be here?
Take care — sometimes the expert might make small concessions to appear more truthful. Seeming to argue against their own financial interests can sometimes serve those interests well, by establishing them as credible, helpful authority figures.
When customers show casual interest in an item on sale, a salesperson might pretend he thinks it has sold out — but that he could check the store room if they would be interested in buying it if it’s available. The apparent scarcity increases the perceived value of the item, and triggers a commitment that can then be used to the salesperson’s advantage when they return with the item and a sale contract to sign.
Related to the above limited-number technique is the deadline tactic. An offer is listed with an expiration date, triggering increased interest.
Primarily, our penchant for scarcity is again a shortcut based on the (generally correct) assumption that scarce items are more valuable.
A second explanation is that of psychological reactance: as opportunities become less available, we lose freedoms; and we hate to lose freedoms we already have. According to this theory, whenever choice is limited or threatened, the need to retain our freedoms makes us desire them (as well as the goods and services associated with them) significantly more than previously.
The power of psychological reactance is most clearly manifested in two age groups: twos and teens.
Under conditions of scarcity (e.g. an item being made illegal) we imbue the scarce item with more positive attributes than we otherwise would.
Often banning or censoring information can make that information appear more valuable or truthful.
When is scarcity most effective as a weapon of influence?
Newly experienced scarcity is more powerful than constant scarcity. Scarcity can be manifested as a lack of rights, and groups who have rights taken away become much more incensed than those who never had them in the first place. Children who lose established privileges are much harder to control than those lacking never-posessed ones.
Even more powerful is scarcity that is generated by demand. This is seen in black Friday sales, and with a lover becoming more keen with the appearance of a rival. It can also be seen in open-bid auctions, where participants and up bidding much more than they intended to. In a way, the winner of these auctions is often the loser.
Countering scarcity: as soon as we feel the tide of emotional arousal that flows from scarcity influences, we should use that rise in arousal as a signal to stop short. We need to calm ourselves and gain a radical perspective.
Once that is done we can calm ourselves why we want the item under consideration. If it is primarily for the purpose of owning it (as a rare commodity) then we should use its availability to help gauge how much we want to spend on it. However, if it is something we want primarily for its function, then we must remember that the item under consideration will function equally well whether scarce or plentiful. Quite simply, we need to recall that the scarce cookies didn’t taste any better.
All these weapons of influence can be used for good or bad ends. They are based on shortcuts we use which, when the evidence they use is legitimate, serve us well. Practitioners who use these ‘weapons’ ethically serve our best interests and allow us to make effective decisions quickly.
Practitioners who fabricate or distort evidence, or construct scenarios which affect our decision making do not serve our best interests and should be treated with caution.