Researchers trained an AI model to ‘think’ like a baby, and it suddenly excelled

Shutterstock

Susan Hespos, Western Sydney University

In a world rife with opposing views, let’s draw attention to something we can all agree on: if I show you my pen, and then hide it behind my back, my pen still exists – even though you can’t see it anymore. We can all agree it still exists, and probably has the same shape and colour it did before it went behind my back. This is just common sense.

These common-sense laws of the physical world are universally understood by humans. Even two-month-old infants share this understanding. But scientists are still puzzled by some aspects of how we achieve this fundamental understanding. And we’ve yet to build a computer that can rival the common-sense abilities of a typically developing infant.

New research by Luis Piloto and colleagues at Princeton University – which I’m reviewing for an article in Nature Human Behaviour – takes a step towards filling this gap. The researchers created a deep-learning artificial intelligence (AI) system that acquired an understanding of some common-sense laws of the physical world.

The findings will help build better computer models that simulate the human mind, by approaching a task with the same assumptions as an infant.

Childish behaviour

Typically, AI models start with a blank slate and are trained on data with many different examples, from which the model constructs knowledge. But research on infants suggests this is not what babies do. Instead of building knowledge from scratch, infants start with some principled expectations about objects.

For instance, they expect if they attend to an object that is then hidden behind another object, the first object will continue to exist. This is a core assumption that starts them off in the right direction. Their knowledge then becomes more refined with time and experience.

The exciting finding by Piloto and colleagues is that a deep-learning AI system modelled on what babies do, outperforms a system that begins with a blank slate and tries to learn based on experience alone.

Cube slides and balls into walls

The researchers compared both approaches. In the blank-slate version, the AI model was given several visual animations of objects. In some examples, a cube would slide down a ramp. In others, a ball bounced into a wall.

The model detected patterns from the various animations, and was then tested on its ability to predict outcomes with new visual animations of objects. This performance was compared to a model that had “principled expectations” built in before it experienced any visual animations.

These principles were based on the expectations infants have about how objects behave and interact. For example, infants expect two objects should not pass through one another.

If you show an infant a magic trick where you violate this expectation, they can detect the magic. They reveal this knowledge by looking significantly longer at events with unexpected, or “magic” outcomes, compared to events where the outcomes are expected.

Infants also expect an object should not be able to just blink in and out of existence. They can detect when this expectation is violated as well.

Infants can detect when objects seem to defy the basic laws governing the physical world. Shutterstock

Piloto and colleagues found the deep-learning model that started with a blank slate did a good job, but the model based on object-centred coding inspired by infant cognition did significantly better.

The latter model could more accurately predict how an object would move, was more successful at applying the expectations to new animations, and learned from a smaller set of examples (for example, it managed this after the equivalent of 28 hours of video).

An innate understanding?

It’s clear learning through time and experience is important, but it isn’t the whole story. This research by Piloto and colleagues is contributing insight to the age-old question of what may be innate in humans, and what may be learned.

Beyond that, it’s defining new boundaries for what role perceptual data can play when it comes to artificial systems acquiring knowledge. And it also shows how studies on babies can contribute to building better AI systems that simulate the human mind.

Susan Hespos, Psychology Department at Northwestern University Evanston, Illinois, USA and Professor of Infant Studies at MARCS Institute, Western Sydney University

This article is republished from The Conversation under a Creative Commons license. Read the original article.

Three Eggs in Creatures Village

Three eggs waiting to be hatched

The official manual for Creatures Village, available through GOG.com, says in its walkthrough that you can choose any one of three eggs to start your game, however only two are provided when you start up a new world.  This was drawn to my attention by a recent help topic in the GOG.com forums.

The most comprehensive way of getting more eggs and norns in Creatures Village is to install Grendel Man's CV Population Update, which gives four eggs at startup and allows you to keep 8 norns, and changes the genetics of the creatures to improve their memory.

If you'd just rather recover the three eggs ability, download this file and put it in your C drive > GOG Games > Creatures Village > Bootstrap > 001 Main folder - it will replace a file in there, restoring the 3 eggs ability.

To see results, start a new world.

How to make a simple remover COB for a Creatures 2 official object

Download Slink's Creatures 2 scriptorium and unzip it.  

Then open the "c2scrp.htm" file in your web browser and you'll get a list that looks a little like this: with sprite files, names, class numbers.


The Scriptorium browser that's available with the Creatures 2 CAOS Tool is more comprehensive, as it comes directly from your copy of the game, particularly for more complex objects, like the Vibraphone, but for a quick reference to sprites and class numbers, Slink's Creatures 2 Scriptorium is where it's at.

Open up Bobcob and choose the autoscript function:

Choose preferences

Tick Source type: Creatures 2, and output of Remover.  Then use the C2 Scriptorium list from before to choose the correct class number, then click Build:

It will then create a new script for you.

inst

enum 2 13 12

kill targ

next

scrx 2 13 12 1

scrx 2 13 12 4

scrx 2 13 12 5

scrx 2 13 12 6

endm

Note that this script will remove all the purple balls from the world:

inst

enum 2 13 12

kill targ

next

and all the associated scripts:

scrx 2 13 12 1

scrx 2 13 12 4

scrx 2 13 12 5

scrx 2 13 12 6

before quitting:

endm

To have a script that simply removes the balls while leaving their scripts intact, you would use the code:

inst

enum 2 13 12

kill targ

next

endm

Move the script from the "Remover Script" section to the "Install Scripts" section, and remove the script from the remover section.

This will allow you to inject it as if it was any other COB.

Final touches

In the Picture tab, use Slink's C2 Scriptorium to find the correct sprite file name from your game and import the image - this can help prevent crashes.  The sprite for the purple ball is dbal.s16.

Now, the little * in the name of the COB means that the file needs to be saved, so I'll go ahead and do that.  The amount being -1 means there are infinite amounts of purple ball remover.  

Download

If you just want to get rid of those purple balls right this second, download the fruits of this tutorial.

Moving on

It's possible to stack remover scripts together in one script so that you can remove more things at once, for example to write a script removing the purple ball (2 13 12) and the green tennis ball (2 13 5) AND all their scripts, you would use the following:

inst

enum 2 13 12

kill targ

next

enum 2 13 5

kill targ

next

scrx 2 13 12 1

scrx 2 13 12 4

scrx 2 13 12 5

scrx 2 13 12 6

scrx 2 13 5 1

scrx 2 13 5 2

scrx 2 13 5 3

scrx 2 13 5 4

scrx 2 13 5 5

scrx 2 13 5 6

endm

Note that the green tennis ball has more scripts associated with it than the purple one does.  I would not recommend creating a 'Remove All' remover COB like this because the inst command makes things happen in 1/10th of a second - so putting too many things in one remove script might make the game crash.

Another use for a simple remover script like this:

inst

enum 2 13 12

kill targ

next

scrx 2 13 12 1

scrx 2 13 12 4

scrx 2 13 12 5

scrx 2 13 12 6

endm

would be to put it in an installation script for a new object using the same class number - so that the purple ball is removed before a new object is created.  

Free Art Source - Pixabay.com!

Pixabay.com is a website with the purpose of providing art, photographs, and videos for free.

The images can be searched for by whether they are photographs, vector graphics, illustrations (including CGI), or videos.  Furthermore, you can search by if the image is in portrait or landscape orientation, and by several broad categories, like Nature/Landscapes.  

When you find a picture, you can download a smaller resolution version for free, under a Creative Commons 0 license.  A larger version is available if you sign up for free.  Pixabay is partnered with Shutterstock, so make sure you are careful not to fall in love with the first image you see - there is a line of Shutterstock content for sale up the top of all search results, and to the right.  

There are some oddities with spriting for the Creatures series that Pixabay images  may need to be altered to fit - in particular, pure black  appears transparent in-game.  However, Pixabay has a wide variety of free art which could easily be adapted into background images for digital collage metarooms.

Updated Metaroom Map

Just a quick note to say that Pilla of Pilla's DS Agents has created an updated metaroom map, with 99 known metarooms on it!  This is based on a collaboration at the Creatures Wiki, the Creatures 3 and Docking Station metaroom coordinates.  Don't forget to open this page to see what the different metarooms in the map are!

Fixed Turkey and Fixed Pudding

I hope everyone enjoyed the Creatures Community Spirit Festival 2016 - I know I did.  Congratulations to Allekha and Doringo for pulling it all together!

I thought I should post a CAOS walkthrough of the Fixed Christmas Pudding and Fixed Turkey (they are identical, bar the sprites and class numbers used) to demonstrate how they work.


First, the original script:
scrp 2 6 5 1
snde chwp 
stim writ from 10 255 0 0 35 250 34 10 57 150 0 0 
pose 1 
endm

Note how the pudding doesn't vanish from the world (using KILL) in the eat script here - this is accomplished by the drop script in the original, leading to a food item that only vanishes once it is dropped - not so good for learning!

The fixed editions have an additional feature that turns these from one-shot treats to easy-to-find emergency rations.

*Eat script for the pudding
scrp 2 6 5 1
*Make a noise
snde chwp
*stimulate the thing that made the script fire, jiggle its brain a little, and give it some nutrients
stim writ from 10 255 0 0 35 150 34 30 57 150 0 0
*strike a pose
pose 1
*WAIT for a second
wait 10
*If there's only one pudding in the world (the eaten pudding in the creatures' paws)...
doif totl 2 6 5 eq 1
*instantaneously
inst
*create a new pudding
new: simp holi 3 0 3500 0
*identical in class to the one before it
setv clas 33948928
*put it in the kitchen
mvto 2934 900
*same attributes
setv attr 67
*same behaviours
bhvr 0 1
*it says hello to the world, and is ready to eat
mesg writ targ 8
*end the if-there's-only-one-pudding check
endi
*the eaten pudding removes itself in its own eat script
kill ownr
*and the script ends
endm

Playing a science-based video game? It might be all wrong

Alex Leith, Michigan State University
You look down from the sky, manipulating the world and seeing how it responds to your changes. You are able to alter vegetation and climate while watching their effects on the surrounding organisms. In this way, and many others, digital games provide excellent opportunities for players to learn about complicated subjects, including the concept of evolution through natural selection. Even games designed for fun and not specifically for education can provide rich, concise, dynamic representations of complex science, technology, engineering and math topics.
Since I was young, digital games have successfully supplemented the educational process in a range of topics, including math, science and biology. Research shows that if these games are going to actually teach those concepts, they must represent them accurately. Games that include incorrect depictions teach the wrong lessons.
Since Charles Darwin and Alfred Russel Wallace, evolution has been understood as a process based on genetic differences between individual organisms of the same species. There are three key principles:

1. Organisms with genetic advantages for their environment are more likely to survive and reproduce, while organisms whose genes make their bodies less suited will die more quickly and reproduce less effectively.
2. Because these characteristics are genetic, they are passed on to offspring.
3. Organisms with genes that improve their survival will have more successful offspring, which will in turn pass on their advantages to the next generation.

Some colleagues and I looked into how well current games could serve as educational tools, specifically about evolution. We examined how Darwinian evolution was represented in 22 games, which we located either through game databases like GameSpot or IGN, or through Google searches. Most games got evolution at least partly wrong. Only five accurately represented all three key principles of evolution.

The five that got it right

A screen-capture of a Norn interacting with its environment in ‘Creatures 2,’ developed by Creatures Labs and published by Mindscape. Wikipedia

“Creatures” provides a rare example of the three principles. In that game, players create cartoon-like creatures called “norns,” through a process that allows norns to be altered not just in terms of appearance, but at the genetic level. For the most accurate representation of evolution, the game offers a play mode called “wolfling run.” In that mode, players cannot directly affect their norns, but can observe their relative fitness for a particular in-game scenario. The potential variations in both norn creation and the environment they must survive in provide for an astonishing number of evolutionary possibilities.
Maxis, best known for creating the “SimCity” game series, and its spinoff “The Sims” collection, also made a set of games called “SimEarth” and “SimLife.” Like “SimCity,” both give players top-down control of a world. “SimEarth” was designed for players to make major changes to the weather, landscape and animals to create an environment. Players were then able to see how the animals would fare in this created environment. “SimLife” was more specific: it has players engage with the animals (rather than merely creating them) to learn about the biology surrounding their survival.

A screen-capture of ‘Who Wants to Live a Million Years,’ playable on the Science Channel website. from sciencechannel.com

We also found two academically oriented games that loosely presented the three mechanics of evolution: “Selection Game” and “Who Wants to Live a Million Years” (which was later renamed “Charles Darwin’s Game of Survival”). The two games were designed to be simple tools that could be played quickly in places like museums. Despite the limited mechanics present in such games, they still clearly show each element of the evolution process.

Market success doesn’t mean accuracy

The most commercially popular game we found didn’t quite get evolution right. “Spore” left out something many other games did, too: Organisms' genetic differences didn’t affect their survival rates. Instead, organisms whose genes were unfit for the environment would not necessarily die more often, in keeping with evolutionary principles. Rather, players could intervene and increase an organism’s likelihood for success by, say, helping it move more intelligently and strategically, beyond the scope of its genetically predisposed movements.
Nevertheless, “Spore” does a reasonable job presenting the broader concept of evolution to players, and is the best such game made this century. (“Creatures,” “SimEarth,” and “SimLife” are all from the 1990s.) “Spore” is also still available for purchase, so it is the only game readily usable by the average educator or student.
But other findings were disappointing. Most games inaccurately portrayed evolution, usually in the same way Spore did – allowing player intervention to save organisms that were unfit for survival.
For these other games, evolution becomes more akin to mutation during a single organism’s life than a process that occurs through generations. In “E.V.O.: Search for Eden” and “L.O.L.: Lack of Love,” players earn points they can spend to modify their organisms. In “Eco,” at the end of each level, the player arbitrarily changes an attribute, though not necessarily one that affects an organism’s survival prospects. In each of these cases, what the game calls “evolution” is actually external genetic manipulation, rather than inheriting particular traits.
These inaccuracies may confuse those unsure of what evolution actually is. If other scientific subjects are similarly poorly depicted in video games, the potential educational benefits of these games could be lost. However, as game designers become more adept at modeling scientific themes, it could herald an educational revolution.

Alex Leith, Doctoral Candidate in Media and Information Studies, Michigan State University
This article was originally published on The Conversation. Read the original article.