What’s the difference between a “core loop” and a “compulsion loop”?

Joseph Kim on Gamasutra has one of the better articles about “compulsion loops” in games.

Edit: see also my article on motivation loops.

I’m inclined to think that there’s something useful in the concept of a compulsion loop. But whether and how this concept differs from “core loops” I’m not sure.

Here’s Kim’s definition:

Compulsion loop: A habitual, designed chain of activities that will be repeated to gain a neurochemical reward: a feeling of pleasure and/or a relief from pain.

That sounds pretty good. I suspect it can be tightened up if someone were to put a more robust theory behind it, but it serves.

Here’s Kim’s definition of a “core loop”, which he says is different from a compulsion loop even though the two are often confused:

Core loop: The chain of activities associated with the primary user flow. What is the user primarily doing over and over again?

Hmm. So the differences are…

  • Compulsion loops are “habitual” and “designed” but core loops…aren’t?
  • Core loops are associated with the primary user flow and compulsion loops aren’t?
  • Users complete the activities in a compulsion loop to gain a neurochemical reward, but they complete the activities in a core loop to gain… something else?

I’m open to the possibility that there’s a useful distinction to be made here, but I don’t think Kim has yet satisfactorily made it.

Two key tools to know your state of mind

Being able to know your state of mind is a prerequisite skill for success and happiness. It’s foundational. The state of mind that you’re in determines how you’ll see and evaluate the world, which determines your decisions and actions. Additionally, your state of mind determines the amount of enjoyment or dissatisfaction you’ll get out of the experiences you’re having.

If you’re not good at knowing your state of mind, you can learn. If you’re good at it, you can get better. It’s one of those skills you never stop practicing. But if you haven’t started already, the best time is now.

Here are two of the best tools you can use to level up your skills.

Vipassana meditation

“Vipassana” means something like “clear seeing”. In this type of meditation, you’re practicing the skills of attentional control and awareness of somatic sensations. In the most common form of practice (as taught by a teacher named Goenka), you’ll sit on a pillow, close your eyes, and systematically move your attention through your physical body, part by part, noticing what sensations arise at each part. Your mind will wander furiously, and each time you notice that you’ve wandered off into some daydream or imaginary conversation, you’ll bring your attention back to the part of your body you’re supposed to be focusing on, and you’ll resume noticing whatever sensations arise there.

By doing this, you’ll gradually improve your ability to hold your attention where you want it. In the process, you’ll be confronted again and again with the meanderings of your mind — meanderings that, normally, you would be simply swept away with, but that now you see as alternate paths that you could travel down but that you don’t necessarily have to engage with. By spending many minutes in this state, you’ll get to know your mind in a way you haven’t before. Next, as you start to be able to sustain your attention where you want it to be, you’ll learn to recognize, but not react to, the physical sensations in your body. You’ll see first-hand the somatic component of your emotions. But rather than reacting like the impulse monster you normally are, you’ll be able to simply observe and grow wiser.

In my experience, the best way to learn meditation is to do a 10 day meditation retreat. The Goenka centers (www.dhamma.org) are fantastic. You’ll be silent for the entire 10 days, meditating for about 10 hours a day. You’ll get recorded instruction from Goenka, a master vipassana teacher who died a few years ago. The cost is free, but at the end you’ll have the chance to make a donation so that another student can have the experience you just had.

This is the best way to learn to meditate. You can certainly do it on your own, or using guidance like Headspace, but in my experience, you’ll only develop shallow skills in this way. A 10-day retreat is the way to quickly become proficient and learn some deep, lasting skills.

Psychotherapy, especially IFS

In vipassana your goal is to avoid getting involved with the stories and explanations floating in your mind. In psychotherapy, you dive headlong into that content. The tension between these two approaches is irreconcilable — they cannot both be used at the same moment. But like a hammer and a screwdriver, they’re valuable in different moments.

Hopefully you have insurance that will cover psychotherapy. If not, find a way to get it. Then find a therapist you like. Don’t be afraid to try a few different people out. If after a few visits you feel like you don’t connect — like you’re not understood, or like this person isn’t going to be able to help you — then try a different person. When you find a great therapist it’ll be worth all the effort.

A psychotherapist will help you see the stuff you’re not seeing. It’s like having a good friend, who’s wise and in fact professionally trained, there to just listen and help you work through your stuff. It’s incredible.

Don’t be one of those fools who thinks that therapy is just for people who have some problem or are “damaged”. That’s like imagining that physical exercise is only for the physically unwell. Your mental health is incredibly important, and you’re making a dumb choice if you wait until your problems are large before investing in it.

A type of psychotherapy that I’ve had great success with is called Internal Family Systems therapy, or IFS. This type of therapy came out of family therapy, in which a therapist would see all the members of a family and focus improving the relations between them. To do so, the therapist would have to understand and empathize with each member, understand the dynamics between them, and help the members understand and empathize with one another. Eventually, therapists began to realize that the metaphor of a dynamic system composed of a number of separate agents each with their own interests, desires, fears, beliefs, superstitions, and idiosyncracies applies surprisingly well to the individual and the “parts” of a person that want, believe, and fear different things. By focusing on these parts as individuals, and getting to know them and empathize with them as one would a member of a family, IFS therapists and their clients were able to achieve huge and rapid progress in bringing harmony to the clients’ inner worlds.

If it sounds weird for you to think of people in terms of separable “parts”, know that in IFS you’re not necessarily taking it all literally — as if these “parts” are little homunculi living in your head. It’s a metaphor. And it’s one that you can work with to achieve a level of clarity and connection that would otherwise be difficult. But also know that the notion that a human personality is not a unitary thing is well supported by psychology and neuroscience. Check out “Incognito” by David Eagleman for an interesting tour of neuroscience research that’ll give you a feel for how stitched-together our consciousness really is.

When you find a good psychotherapist, dig in. Don’t hold back. Don’t waste your time trying to make yourself look good. You have too much to gain.


These two tools are the best ones I know. They provide foundational skills and vocabulary that you’ll use to unlock may other tools and techniques. Start here.

Look upstream, change downstream: the importance of knowing and managing your state of mind

Do you ever find yourself faced with a project that you must work on right now, but you just don’t feel like it? Maybe it’s a report you have to produce or an email you have to write. Or have you ever found yourself hanging out with friends on a trip that’s supposed to be fun, but you’re feeling grumpy and you kinda don’t want to be there?

Sometimes, the right course of action isn’t to change what’s going on outside, but to change what’s going on inside. You have to change your state of mind.

But that’s hard to do.

First of all, some part of you believes that your current state is the Right state. It’s what you Should be feeling. Change my state of mind? This is me. This is what I’m feeling. To want to be feeling something else would be to want to not be me.

But that’s largely bullshit. You’re lots of things. You’re the observer as well as the actor. The rider as well as the elephant. Most certainly, “you” are not coextensive with the current contents of your consciousness.

So the question is really one of practical rationality. Is the state of mind that I’m in going to serve me well in the pursuit of my goals? Or would a different state suit me better?

This is the real hard part. You have to know what state you’re in, you have to know what state you Want to be in, you have to know how to successfully move yourself from this state to that state, and then you have to actually do it. That’s a lot. And it’s hard.

But that skill is hugely empowering. Your state of mind is the upstream source of not only all your actions, but also of the quality of your experience. It is the source of your success and your happiness. If you can learn this skill you can benefit enormously.

This skill consists of three steps. Each of these steps is an art that takes years to master. I haven’t mastered any of them. But I have practiced a lot and learned some things.

Step 1: Understand your current state.

What are you feeling? How did it come to be? What’s the significance of those feelings?

Tools for developing this skill:

  • Mindfulness meditation
  • Psychotherapy
  • Affect naming
  • Talking about feelings, emotions, states of mind with friends

Step 2: Identify the state you want to be in.

Imagine a parallel universe in which you’re in the same situation, but you’re in a great mood, and whatever that thing is that you have to do, you actually want to do it, and you knock it out of the park.

What state of mind were you in in that universe? Cool. That’s the state you want to be in.

But it’s a little more complicated than that. The you in that other universe must have gone through a chain of causes that was different from yours in this universe in order to have wound up in a different state of mind. It’s not certain that you’ll be able to get to that exact state of mind.

So really, you’re looking for the best accessible state of mind. And knowing what’s accessible isn’t easy. Tools again:

  • Mindfulness meditation
  • Psychotherapy
  • Remembering past experiences of success
  • Watching and learning from others who are doing this well
  • Practice

Step 3: Make a plan for how you’re going to move from this state to that state.

How do you move from this state to that state? Here you want a step-by-step plan; a formula. And it matters that the plan is correct, in the sense that if you do the steps, then you will get the intended outcome.

This again takes a lifetime to master. There aren’t a lot of generalizable skills that I can name. It’s mostly about learning the mappings one by one: I’m in this state, and I can get to that state by taking these steps.

To learn those mappings, you mainly have to experiment. Tools:

  • Try things. Try lots of things, and pay attention to the results. Record it, maybe, so that you’ll have a good chance of remembering it. If it worked, great. Don’t forget it. If it didn’t, why not?
  • Psychotherapy can be a great way to get ideas for how to move from one state to another.
  • Mindfulness meditation often is the tool you need to move from the state you’re in to the target state.
  • Talk to friends about what works for them.

This is barely scratching the surface. But it’s appalling to me that we as humans don’t have this better figured out by now. So much rides on it. And we hardly even talk about it explicitly. And to my knowledge, there isn’t any coordinated, systemized effort. Maybe it’s time to start one.

Should you stay or should you leave your job?

So you’re thinking about leaving your job. Should you stay or should you go?Here are 5 things to consider, in no particular order.

People — fun. What percent of your moments are on the positive side, and what percent are on the negative side? Your coworkers probably play a huge part in this. Do you like them? Do you want to be more like them?

People — learning. Are there people around that know way more than you do about things you want to know more about? People with hard-won wisdom that you can’t get on Coursera? Are you learning from them?

Projects — learning. The stuff you’re working on — are you learning a lot from it? Or has it become rote? If you’re learning, how much is that learning going to serve you in the future?

Projects — alignment. The stuff you’re working on — how much do you care about it? Is it aligned with your goals and values? Does it matter? 10 years from now, will you look back and be glad that you were doing this?

Compensation . How much are they paying you? How much would you make elsewhere? How much room is there to move at your current place on base — would they pay you more? Suggestion: gain information by doing an interview somewhere else. 

Instrumental and non-instrumental desires

When I desire something instrumentally, I desire it because it’ll help me get something else. For example, I might desire to know my password to a website, because I want to log in to the website and knowing my password will enable to me to do so.

When I desire something non-instrumentally, I desire it for no instrumental reason whatsoever. If it led to no outcomes beyond itself, I would still desire it.

What, then, would be an example of a thing desired non-instrumentally? Pleasure? Discomfort alleviation? Cf “incentive salience” – a kind of desire distinguishable from the explicitly declarative, instrumental kind of desire.  http://www.ncbi.nlm.nih.gov/pmc/articles/PMC2756052/

The one thing you can’t fake in rapid prototyping: motivation

Rapid prototyping on real users is an incredibly effective way to search solution space for product configurations that will actually resonate with users. But there’s one critical mistake that I see rookie user testers making on the regular. If you do this, your test is bunk, and if you believe the bunk results of your bunk test, your product might be bunk too.

In a rapid prototyping session, you make a very rough approximation of the product use scenario. You can ask a user to fake lots of things. Pretend this piece of paper is an app on your phone. Pretend you’re on the bus. Pretend I’m not here. Pretend you’re 16 years old. Pretend you’re a mother. It’s actually fine to ask a user to pretend all of those things, as long as you help the user get into the state of mind that you’re requesting.

But there’s one thing you cannot ask the user to fake: motivation. The moment you ask a user to pretend that he’s interested, or pretend that he wants some outcome that your product is trying to deliver, the test is a charade. The user is now faking the very thing you’re here to test.

Usually it goes like this:

Researcher: What would you do on this screen?

User: I’d probably quit because I’m bored.

Researcher: …okay. Pretend you clicked “Next” and you saw this next screen. What would you do here?

User: Um…. I guess I’d click Sign Up?

No! Bullshit! The key question is whether the user cares. As soon as you ask him to fake it, you might as well just tell him to go home, because you can sit in your armchair and speculate about what he would do just as well as he can. You are no longer learning about the critical question: what do people care about? What gets them excited?

So, never ask users to pretend about motivation.

Names matter more than most people think

A rose by any other name would release the same chemical odorants, but if a rose was called “pukeweed” you probably wouldn’t buy one for your sweetheart.

Names are important. Product people often underestimate how important they are. And since naming a product or brand can be really hard, it’s easy to say we’re spending too much time on this. It’s not that important. Let’s just pick one and move on.

But names matter. A lot. If you’re trying to name your product and you haven’t found a name that feels right, you probably shouldn’t put the search to rest until you do.

Here’s why names matter.

Nearly every time a person encounters your product they encounter your name. That name creates a snap emotional judgment. This judgment occurs before declarative thoughts and conscious intentions. It’s a feeling that precedes conscious awareness (in the Blink sense), and it creates the context in which the brain goes about forming declarative opinions about the object. After that first instant, the thing already has established its emotional valence. All that’s left now is for the brain to find reasons to explain why the thing is so good or so bad.

That initial movement will start a feedback loop. You hear the name. You feel good. You look for reasons the thing is good. You find them. You declare them. Now, noticing those, and getting your ego wrapped up in it, you feel even more good feelings about the thing. And so on. Once that process starts, the ball will tend to keep rolling down the same side of the hill.

If the initial feelings are negative, the relationship between the person and product is likely to end right there. If the initial feelings are positive, the product has a foot in the door. The person is interested. Their brain is already at work generating reasons why this thing is cool or valuable. There can be further interaction. A relationship might blossom.

By this process the seemingly tiny factor of the first emotional, aesthetic response to the product’s name has an enormous effect on the final outcome of the relationship between the person and the product. I’m reminded of the cheesily awesome explanation of chaos theory delivered by Ian Malcolm to Ellie Sattler in the Jeep on Jurassic Park.

Consider virality. If I have a negative feeling about a product, it’s unlikely that I’m going to tell my friends about it. If when I say the name I feel a negative shadow, a twinge of embarrassment, an urge to defend or convince you that no, this thing is not what you think, it’s actually cool, hear me out — then I’m probably not going to talk about it. And the product won’t spread. But if I feel good about it, and am excited about it, I might be happy to talk about it.

That’s virality at the micro level. Now consider the macro level. For your product to be successful, it’ll need millions of people to hear about it, check it out, and develop a relationship with it.

How to test if a name is good

For me, there are two steps to a name. When you have a candidate for a name, here’s how to evaluate it.

First: Say it to someone. Describe your product, and use the name as if it’s already been decided upon. Choose a person whose judgment you trust, and who isn’t predisposed to like everything you say. Not your mother. Coworkers and critical friends are good. And when you say the name, pay very close attention to how you feel. Did it feel good? Did you enjoy saying the name? Did you feel embarrassed? Was it a bit of a strain? Your answer there is all you need to know, because as Don Draper says, “You are the product. You, feeling something. That’s what sells.”

Second: If your name passed step one, great. But it now it has to pass another step: the sanity check. Someone recently pointed out to me that “chlamydia” is actually a pretty word, if you separate it from its meaning. But yeah. That meaning. I won’t be naming my daughter Chlamydia. If the name is trademarked by your competitor, or if it refers to a venereal disease, keep looking.

Anonymous daily writings

Julie Zhou has a great post about why she writes, how she got herself to do it, and how you can do it too.

In particular, this:

In all the times before that I have failed to get something on paper, it was because I had thoughts like the following: geez, what if I hit publish and nobody reads this? That’d be embarrassing and pointless. Or I only want to publish something if it’s really good and makes me seem smart, witty, and knowledgeable. Or What if I say this and somebody disagrees and tells me I’m wrong? Or Hmm, I should only write when inspiration hits me, and right now I don’t feel inspired.

In every creative endeavor — not just writing — this train of thought paralyzes. I have experienced it enough times to know that holding yourself to some lofty standard when you are just starting out is like blowing a deathkiss to your chances of success.

That resonates with me. Those thoughts occur, and suddenly I’ve found some other activity that’s more urgent and important than continuing to try to write.

And then her advice:

Instead, if you’d like to write, I offer the following tips:

1. Set a writing goal that is purely about the mechanical act of doing.Maybe, like me, it’s Hit the publish button every third Tuesday, Maybe it’sWrite 3 journal entries a week. Or maybe it’s Write 500 words a day. (In case you wonder how all your favorite authors complete their novels, I have it on good information that pretty much all of them do it via daily word-count/time-spent-writing goals.)

2. Tell yourself that nothing else matters besides #1. The thing you publish every third Tuesday does not have to fit any particular theme (in my case, not having any better ideas at the time, I’ve published poetry,listicles, and essays about my dog.) Your journal entries can be one sentence long. Your 500 daily words can be crap words. Don’t obsess over your audience. Don’t try to write what you think other people will want to read. Write about what you are excited about, because the best writing tends to reveal a piece of yourself anyway. The point is to bust down any possible barrier that might get in the way of you being able to achieve #1.

3. Commit to doing #1 for long enough that you will have built a habit out of it. A week or a month isn’t sufficient. Try 6 months or a year. By then, the act of writing will have molded to your life like a favorite sweatshirt, and you will begin to feel its effects on the way you think, reflect, and process the world.

But the first thing she did was to write an anonymous blog. Because her fear was keeping her from writing.

Yes, I was afraid.

To write publicly is to put yourself out there. To take a stance on something, propose an idea, have a point of view. It is to give someone else — someone you may not know and may never even meet — a piece of evidence with which to form an opinion of you. I cared deeply what others thought of me. (When I was little, I refused to ask grocery store clerks simple questions likeWhere are the oreos? for fear of seeming incompetent. As you can guess, this sacrifice cost me dearly in terms of snack-time utility.) I worried about what it would mean to admit weaknesses publicly, to write about touchy topics like gender and bad behavior and all the things that I’m learning. I worried what friends and coworkers would think.

And apparently, using the cover of anonymity, she was able to lower her fear enough to start writing.

In 2012, I sat down in January and scrawled a New Year’s Resolution on a sticky note: Write a blog. I did it the only way I knew how at the time: facelessly and anonymously. And that helped to get the words flowing. I wrote and published twice a week. The anonymity helped me share stories like what I learned from negotiating my first salary, tactics for interruptions, and what it felt like to harbor a Jekyll-Hyde impostor syndrome most days of the week.

My bet is that that step was crucial. Writing is hard because it’s so many things at once. Anonymity allowed her to condition her fear down while building the mechanical habits of writing. Only after she did that for a while — it sounds like she started in Jan 2012, and petered out in April 2012 — did she go nonymous.

Talking Machines podcast Ep 5 transcript: distributed representations vs symbolic representations

The output from a neural network asked to over-interpret an image clouds. Google Research

This is a transcript of an interesting segment from episode 5 of the Talking Machines podcast, which contains part 1 of a conversation between three leaders in machine learning: Geoffrey Hinton, Yoshua Bengio, and Yann LeCun.

The full episode is available here: http://www.thetalkingmachines.com/blog/2015/2/26/the-history-of-machine-learning-from-the-inside-out

21:36

Ryan (host):

It seems like thinking hard about distributed representations is some of the most exiting stuff that’s come out of this resurgence. It’s a very different way of — I would say it kind of challenges a long history of knowledge representation. It feels very biological, right? Geoff can you talk a little bit more about distributed representations and maybe explain that to our audience.

Geoff:

Ok the idea is that you have a large number of neurons and they’re conspiring together to represent something and they each represent some tiny aspect of it, and between them they represent the whole thing and all its wonderful properties.

And it’s very different from a symbol. Where a symbol is just something that is either identical or not identical to another symbol. Whereas these big patterns, these distributed representations, have all sorts of intrinsic properties that make them relate in particular ways to other distributed representations. And so you don’t need explicit rules, you just need a whole bunch of connection strengths, and one distributed representation will cause another one in just the right way.

For example you could read an English sentence and get a distributed representation of what it means, and that could cause a distributed representation that creates a French sentence that means the same thing. And all of that can be done with no symbols.

Yann:

So the power of that concept can be seen in the fact that all of us, in all of our labs, are essentially working on embedding the world — you can think of it this way. So how do we find vector representations for words, for text in various languages, for images, for video, for everything in the world. For people, actually, so you can match people’s interests with content, for example, which is something that Facebook is very interested in.

So finding embedding is a very interesting thing. And there’s a lot of methods for doing this. For text there’s the very famous method called word2vec invented by Thomas Mikolov[?].

And following the neural language model that Yoshua had worked on before that, Geoff and I also had worked separately on different methods to do high-level embeddings rather than low-level embeddings. So things that could be applied to images for example. So I guess this could be called metric learning. So this is situations where you have a collection of objects and you know that two different objects are actually the same object with different views or the same category. So two images of the same person, or two views of the same object, or two different instances of the same category.

And so you have two copies of the same network, you show those two images and you tell the two networks ‘produce the same output. I don’t care what output you produce, but your output should be nearby.’ And then you show two objects that are known to be different, and then you can push the output of the two networks away from each other.

Geoff had a technique called NCA to do this neighborhood component analysis. […] And then Jason Weston and Sonny Bengio came up with a technique called Wasabi which they used to do image search on Google. Google used that as a method to build vector representations for images and text so you could match them in search. At Facebook we’re using techniques like this for face recognition. So we find embedding spaces for faces, which allows us to search very quickly through hundreds of millions of faces to find you in pictures, essentially.

So those are very powerful methods that I think we’re gonna use increasingly over the next few years.

25:21

Ryan:

Is there a point where you need to have discrete grammars on top, or can it be distributed the whole way down?

Geoff:

My belief — if you’d asked me a few years ago, I’d have said well maybe in the end we need something like a discrete grammar on top. Right now I don’t think we do. My belief is we can get a recurrent neural network — that is something with an internal state that has connections to itself so it sort of keeps going over time. We can get that kind of network to translate from one language to another — this has been done at Google, and it’s been done in Yoshua Bengio’s group — we can do that with nothing that looks like symbols with symbolic rules operating on them. It’s just vectors inside.

It works very well. It’s at about the state of the art now, both at Google and at Yoshua’s lab. And it’s developing very fast.

And I think the writing’s on the wall for people who think the way you get implications from one sentence to the next is by turning the sentence into some kind of mentalese that looks a bit like logic and then applying rules of inference. it seems that you can do a better job by using these big distributed vectors, and that’s much more likely to be what people are up to.

Yann:

There’s a very interesting white paper or position paper by Leon Bottou the title is “From Machine Learning to Machine Reasoning” which basically advocates the idea that we can use those vector representations as the basic components of an algebra for reasoning, if you want. Some of those ideas have been tried out but not to the extent that we can exploit the full power of it.

And you start seeing work now, so for example my colleague [?] Fergus […] and someone from Google, worked on a system that uses distributed representation that identifies mathematical identities. And it’s one of those problems that is very very sort of classical AI — like solving intervals and stuff like that — that involves reasoning and search and stuff like that. And we can do that recurrent nets now to some extent.

Then there are people working on how do you augment recurrent networks with sort of a memory structure. So there’s been ideas going back to the early 2000’s or late 90’s, like LSDM which is pretty widely used at Google and other places. So it’s a recurrent net that has a sort of a separate structure for memory. You can think of it as sort of a processor part and a memory part, where the processor can write and read from the memory.

So neural Turing Machine is one example, there’s another example. Jason Weston [and others] have proposed something called a Memory Network which is kind of a similar idea. It’s somewhat simpler than SDM in many ways. […]

And there’s a sense that we can use those types of methods for things like producing long chains of reasoning, maintaining a kind of state of the world if you want. So there’s a cute example in the memory network where you can tell a story to the network, like say Lord of the Rings, so “Bilbo takes the ring and goes to Mt Doom, and then drops the ring, and blah blah blah.” You tell all the events in the story, and at the end you can ask a question to the system, so, “Where’s the ring?”and it tells you, “Well, it’s in Mt Doom.” Because it maintains sort of an idea of the state of the world, and it can respond to questions about it.

So that’s pretty cool cool because that starts to get into the stuff that a lot of symbolic AI people said neural networks will never be able to do.

Yoshua:

I’d like to add something about the question you asked regarding distributed representations and why they are so powerful and behind a lot of what we do.

So one way to think about these vectors of numbers what they really are are attributes that are learned by the machine, or by a brain if we think that’s how brains work. So a word or an image or any concept is going to be associated with these attributes that are learned.

Now associating attributes to concepts is not a new idea. Linguists will define things like the gender or plural or this is an animal or this is alive or not. And people trying to build semantic descriptions of the world do that all the time. But here the difference is that these attributes are learned. And the learning system discovers all of the attributes that it needs to do a good job of predicting the kind of data that we observe.

The important notion here is the notion of composition — something which is is very central in computer science and also in many of the older ideas of AI. Cognitive scientists thought that neural nets cannot do composition.

Actually composition is at the heart of why deep learning works. In the case of the attributes and distributed representation I was talking about it’s because there are so many configurations of these attributes that can be composed in exponentially many ways that these representations are so powerful.

And when you consider multiple levels of representations, which is what deep learning is about,

then you get an extra level of composition that comes in, and that allows you to represent even more abstract things.

Geoff:

A nice example of distributed representations where you can see them at work in people, is if you just have symbols, you might have a symbol for a dog and a symbol for a cat, and a symbol for a man and a symbol for a woman. But that wouldn’t explain why you can ask anybody the following question, and young kids can do this. If you say “You’ve gotta chose: either dogs are male and cats are female, or dogs are female and cats are male.” People have no doubt whatsoever. It’s clear that dogs are male and cats are female.

And that doesn’t make any sense at all. And the reason it’s clear is because the vector for dogs is more like the vector for man, and the vector for cats is more like the vector for woman. And that’s just obvious to everybody. And if you believe in symbols and rules, it doesn’t make any sense.