Remembering and Forgetting, Saturation in Neural Networks

This study by Rosenzweig, Barnes, and McNaughton highlights the importance of forgetting in order to make the best use of the brain cells we have,

http://frank.itlab.us/forgetting/making_room.html

If we fail to forget, our neural networks will saturate and become useless.  Saturation in a neural network does not merely mean that a network cannot learn more, it can mean that a network could fail to respond to input in an appropriate manner.

Consider a very simple network consisting of two input neurons I₁ and I₂, and two output neurons O₁ and O₂.  A neural network learns by increasing the strength of connections between associated inputs and outputs.  For instance should an input signal be present at I₁ while an output signal is also present at O₂, then the connection I₁↔O₂ would be strengthened.  Consider Ivan Pavlov, his dog, a dog treat, and Pavlov’s bell.

A trained neural network acts by pro-actively triggering appropriate output neurons when specific input signals are present.  Should a signal trigger the first input neuron in our example, I₁, the second output neuron would be triggered pro-actively.

• Given the “learned” synaptic connection: I₁↔O₂
• Assuming: I₁
• Triggered: O₂ (I₁→O₂)

Consider Pavlov’s dogs salivating when the bells rang regardless if treats were provided.

If a second training exercise triggered input I₁ but instead the first output neuron was triggered in lieu of the second, the connection I₁↔O₁ would also be strengthened.  We now have,

• I₁↔O₁
• I₁↔O₂

After the second training session, should I₁ be triggered once again, which output neuron would trigger?  Without any further weighting functions to apply to our connections, a I₁ signal would trigger both outputs,

I₁→O₁∪O₂

Consider a situation where Pavlov’s dogs were sometimes offered treats when the bells rang, or sometimes were given electric shocks.  What would the dogs have expected the next time bells rang?  Would they have expected treats, electric shocks, or both?

Perhaps this state of affairs is desirable, perhaps it is not.  Now that this cross-association is saturated however, there exists no way to trigger only O₂ given I₁.  Even if all future training sessions reinforce the I₁↔O₂ connection, the system will remain ambiguous forever.

It is likely that nature’s first, simple neural networks exhibited this kind easy saturation.  Perhaps early critters could only adapt to very limited environmental conditions during their very short lives.  Perhaps these critters simply died from indecision if they encountered natural oddities they weren’t prepared for.  In the competitive evolutionary race however, those critters who occasionally reset their saturated networks would have an evolutionary advantage over those who did not.  To reset an easily saturated neural network would have been to allow the forgetting of anomalies.  These critters would have had a better chance of survival in the real, random natural world.  They would relearn their most common and important lessons and forget the oddities which simply did not pertain to most circumstances of their lives.

In the context of the article, 4-(3-phosphonopropyl) piperazine-2-carboxylic acid (CPP) provides an occasional “reset” function to spatial memory that allows de-saturation and re-learning.  CPP is one of nature’s “dirty tricks” that helps to alleviate the downsides of easily saturated neural networks.  Nature has converged upon many such dirty tricks over the eons, including:

• Chemical washes (CPP)
• Inhibition, “pulsing” and other mild periodic reset mechanisms
• Network segmentation (slows saturation)
• Physical growth and degeneration
• Specialty circuits (e.g., “instinct”)
• Preferential learning such as that which provides increased weight to electric shocks versus pleasurable food treats
• Consciousness (self-awareness)
• Concept formation and other information compression mechanisms
• Emotion, heuristic, magical thinking, social deference and economic behavior in humans

The basic lesson is that, short of ameliorating effects, all neural networks easily saturate.  For any cognitive function, researchers should ask two questions:

• How does the associated network saturate?  What are the effects?
• What solutions has evolution converged upon to de-saturate the network?

Data Compression in the Brain

This article, “How the brain assigns objects to categories“, describes a great example of how multiple, segmented neural networks accomplish the fine art of “data compression”. Evolutionarily speaking, this is a “cheap and dirty trick”, but it works!

http://medicalxpress.com/news/2011-07-brain-assigns-categories.html

Once the generic category is formed in the prefrontal cortex, the straitia are free to focus on other details.  The first few exemplars of a category may linger in memory, somewhere, but new exemplars are likely ignored.

What if the straita could not forget?  What if the prefrontal cortex was not present to absorb recurrent exemplars?

Memory, Adaptation and Entropy

I will write more in the coming weeks and months about the various types of memory a life form may leverage in order to adapt to its environment.  An interesting article from ScienceDaily illustrates how epigenetics, those chemical changes which alter the way DNA is processed (or not processed) in our cells, provide an organism with an adaptation subsystem that helps it better fit its environment,

http://www.sciencedaily.com/releases/2011/07/110724135553.htm

Adaptation cannot occur without memory.  Organisms, including plants, leverage many forms of memory.  Other than chemical and physical construction, perhaps the most important characteristic which differentiates kinds of memories is the informational entropy capacities of those memories.  Memory systems with higher entropy capacities can assimilate larger informational variety.  As the informational variety (entropy) capacity of a memory system increases, so will rise the organisms potential to adapt to a greater number of environmental conditions.  That is, the higher the entropy capacity, the higher the potential utility of the adaptive system.

From the article,

Epigenetic memory comes in various guises, but one important form involves histones — the proteins around which DNA is wrapped. Particular chemical modifications can be attached to histones and these modifications can then affect the expression of nearby genes, turning them on or off. These modifications can be inherited by daughter cells, when the cells divide, and if they occur in the cells that form gametes (e.g. sperm in mammals or pollen in plants) then they can also pass on to offspring.

I will also illustrate in the coming weeks and months that adaptive system utility is not merely a function of higher information entropy capacity.  Adaptive system utility can also be extended by the system’s ability to “clean house”, “collect the garbage” and reduce information variety when the system has become saturated.

Delegating Our Precious Memories to the Cloud

An interesting set of results published in MIT Technology Review,

http://www.technologyreview.com/web/38032/?nlid=nlweb&nld=2011-07-20&a=f

In short, if we think some information will be on the internet later, we are likely to not bother remembering that information today.

One way to interpret the results of the study is that, where we have a chance to jettison information, we do.  Hopefully, “jettison” means to “compress” information into a handy rule of thumb, word, concept or theory which we reuse with efficiency, rather than to reuse the original data set over and over again in all its unwieldy bulk. Sometimes however, we can be conned into simply erasing or otherwise strongly de-prioritizing memories without bothering to create a useful summary of what we’ve lost.  This supports that idea that our memory erasure mechanisms (or de-prioritization) are separate from our symbol creation mechanisms.  This would make sense from an incremental, evolutionary perspective.  It would also suggest that I should not assume my “transaction model” of cognition represents the working of a single mechanism.

It is interesting that one of the researchers (Wegner) refers to this as “transactive memory” (this might be related to transactional analysis in psychology).

Heuristics Over Logic

We have evolved to favor heuristics over logic precisely because we have evolved to make decisions,

http://www.troyhunt.com/2011/07/science-of-password-selection.html

To make a decision is to reduce data in a problem domain to some compressed form, call it a concept or a word, then reuse that form in the future.  If we had to use all data we had ever learned to adapt to every environmental change, we would keep slowing down as we learned more, eventually to be crippled by the ambiguity of the data we had collected over time.  While this behavior helps us to maximize the utility of the brain system we have, it also leaves us with a glaring hole: we may fail to realize that the use of a specific word, heuristic or concept is invalid or otherwise fraught with some kind of risk.

Remember, decision-making and concept formation are one-way trips.  Information is discarded.  This means we had better make good decisions the first time because we might not be able to easily give ourselves second chances to reconsider.

---

Update 20 July 2011: See also,

http://www.michaelshermer.com/2002/09/smart-people-believe-weird-things/

It is Always a People Problem. Always.

There are no such things as technology problems, only people problems.

No technology can build itself, nor use itself, nor correct its own problems.  Even self-replicating machines, built using any technology in use (or even in conception) today, would merely execute the delayed choice of their builders.  Consider the case of a man, eager to protect his home against theft, who installs an anti-theft device which would kill any unwanted intruder, perhaps with a bullet to the head.  The homeowners’s device is commonly called a booby trap.  One day, while the home owner is away, an intruder enters the home and is killed.  Is the home owner responsible?  You betcha!  The home owner may claim they are not responsible because they did not pull the trigger directly, but in the end they made a choice to apply extreme prejudice to any intruder and they developed a device to execute that delayed choice.  The homeowner’s booby trap did not kill the intruder, the home owner did.  Every action of any technology, including any act of construction, any act of repair, or any act of use, is ultimately the extended action of human beings.

No technology is a perfect fit for any problem and all technologies come with trade-offs associated with their use.  Even survival comes with its own set of trade-offs.  It is the responsibility of human beings to understand their problems to the best of their abilities, to understand the trade-offs associated with the technology options before them, and to choose appropriate technologies wisely.  Trade-off balancing does not happen on its own.  Humans are the ultimate arbiters of which technology problems they choose to live with.

If all humans were to vanish from this Universe tomorrow, there would be no human problems of any kind.  Human technologies would instantly cease being human technologies and would merely exist as artifacts of matter like any other.  At the same instance of Universal human extinction, all “problems” would also similarly vanish.

This is not merely an academic exercise in ethics.  The implications of failing to understand this point can be tremendous.  If the home owner in my delayed choice example would have understood his culpability ahead of time, would he have been so eager to create his intruder-killing device?  The lack of understanding of the concept of delayed choice leads, in business, law and in politics, to a class of problem called moral hazards.  Failure to understand this critical point about technology, in particular computing technology, can cause some people to impart “magical” qualities to technologies which the technologies do not have, which can skew expectation, and can lead to project and business failure.

No, no, no.  The only kinds of problems which exist in this world are people problems, by definition.  If you doubt that, then find a way to kill all of humanity right now and watch all problems simply vanish away the moment before you and I cease to be.

Current Activities

I know, I know, this blog has been a little quiet.  I have been involved lately with fiction writing and maybe even a little game design.

Two short stories:

• “My Wife”
• “The Frog of Truth”

Now writing chapter 4 of my novel:

• “Mercedes 10″

The I/T architecture ethics book?  I am in research stage with cognitive informatics, decision theory and behavioral economics. Understanding how and why human beings make the decisions they do is critical in software application design as well as business design, yet I am astounded at how little this set of related phenomena is understood by software and business designers.  I would say that, in some cases, the lapse is downright criminal.

Cognitive Entropy and Cognitive Informatics

In regards to my thoughts on cognitive irreversibility, I think the extant research favors the term, “cognitive entropy“.  I have a lot of reading to do, but I am not yet sure if my particular thoughts have been explicitly addressed.

An interesting paper, here:

http://psycnet.apa.org/index.cfm?fa=buy.optionToBuy&id=2009-22869-014

Also, here:

http://portal.acm.org/citation.cfm?id=1020230

Apparently, the IEEE has an interest group on Cognitive Informatics.  See also, the International Journal of Cognitive Informatics and Natural Intelligence,

http://www.enel.ucalgary.ca/IJCINI/

Barry Schwartz and the “Paradox of Choice”

This could be an interesting read in both topics of cognition and behavioral economics,

http://en.wikipedia.org/wiki/Paradox_of_choice

(note the section, “Why we suffer“, which corresponds to some of the notions I was throwing around in my previous article)

http://www.amazon.com/The-Paradox-of-Choice-ebook/dp/B000TDGGVU

A review and synopsis of the book can be found here,

http://andreagandino.com/journal/2009/cognitive-entropy-the-paradox-of-choice/

In short, we humans are not very fond of possessing too many choices.

Cataloging Cognitive Phenomena Using Reversibility Criteria

As you can probably tell by yesterday’s post, my thesis is still young, not quite formed, and has a few holes.  As an exercise, I am considering cataloging cognitive behavior (especially economic behavior) in terms of reversibility.  I am wondering if the results of this exercise, which might have a physical basis in neurobiology, could result in a kind of “periodic table of elements” for human behavior. Could it help point the way to better understanding of the neurobiology of various behavioral mechanisms?  This would be a mult-dimensional map, including:

• Degree of irreversibility
• Irreversibility seeking versus irreversibility maintenance
• Irreversibility recognition (do you know it when you see it?)

Consider some examples:

• Defense of property (irreversibility maintenance)
• Economic transaction (irreversibility seeking)
• Social bonding (irreversibility seeking while bonding, maintenance afterwards)
• Obsessive compulsive disorder (irreversibility seeking, but unable to recognize it when it occurs)
• Schizophrenia (brain randomness: low irreversibility/bias, high irreversibility seeking behavior)

Consider the potential mapping along these dimensions:

	Reversible ←	→ Irreversible
Successful recognition ↑
Recognition failure ↓

Indeed, I am having a difficult time expressing what is on my mind.  I will try to read more literature on behavioral economics to see if others have already tread these waters and also to see if I can get some hints on how to express my thoughts better on this topic.  Of course, this work might be a nothing but a snipe hunt, but I think I might at least learn something from the exercise.