Before Alan Turing we did not know about a single machine that could emulate other machines. Turing opened the door to developing a machine which can have multiple Ways to Think
We have multiple ways of estimating distances - remembering typical sizes, overlaps, context, binocular vision, perceived speed. Each method is imperfect but taken together we can usually avoid serious mistakes. We effortlessly switch between methods choosing the more appropriate one for each different situation. This might also apply to how we think.
Panalogy (a word invented by Minsky meaning parallel analogy, also discussed in Chapter 6)
Panalogies are corresponding features of different meanings which are connected to the same parts of one larger structure. Minsky suggests that our brain architecture has evolved structures which make it easy to link knowledge fragments in this way.
- Whenever you think about your Self, you are reflecting about a panalogy of mental models of yourself
- Sight is intertwined with memory, we fill in huge chunks from memory when "seeing" (without realising)
- We rarely make entirely new ideas, instead we modify existing ideas
How do people learn so rapidly?
Sometimes we learn new tricks from a single exposure (whereas a dog may require hundreds of lessons). A difference engine could be converted into a copying machine, so the structure in long term memory becomes the same as the one in short term memory
Minsky thinks our minds are like computers in this respect. Short term memory is expensive and limited.
"... a blow to the head can cause a person to lose all memory of what happened before and including that accident ... transfer to long term memory may take a day or more and require sleep"Other reasons why long term memories may require much time and processing:
Retrieval - it may have to be linked to an existing panalogy, otherwise how could it be retrieved?
Credit Assignment - to be useful it would need to be linked to other relevant panalogies
Real Estate problem - finding a place for new memories would not be simple (might involve destruction)
Copying complex descriptions - hard to think of plausible schemes for making complex, linked memories
Learning involved many varied skills, such as:
- Adding new If --> Do --> Then rules
- Changing low level connections
- Making new subgoals for goals
- Choosing better search techniques
- Changing high level descriptions
- Making new Suppressors and Censors
- Making new Selectors and Critics
- Linking older fragments of knowledge
- Making new kinds of analogies
- Making new models and virtual worlds
Behaviourism is limited. Learning complex things cannot be explained by reinforcement or if-do rules
Minsky speculates that we might use higher level processes to decide what to learn from each incident by reflecting on our recent thoughts. These processes could be used to make such "credit assignments":
- choosing how to represent a situation will affect which future ones will seem similar
- learn only the parts of your thinking that helped, and forgot those which were irrelevant
- connect each new fragment of knowledge so that you can access it when it is relevant
We need more research about what kind of credit assignments infants can make, how children develop better techniques, how long such processes persist and the extent to which we can control them
Transfer of learning to other realms:
Some children seem to transfer their learning to other realms, while others don't
Transfer of learning might be superior for those who make better credit assignments. To gain more from each experience, it would not be wise for us to remember too many details but only those aspects that were relevant to our goals. Also what we learn from an experience might be more profound if we assign credit to the earlier choices we made that selected our winning strategy
Creativity and Genius
Genius consists of unusual combinations of otherwise common ingredients:
- fortunate mental accidents
- learn how to praise self internally
- intense positive attention from parents
- isolation from other children
- mental management
- enduring discomfort when replacing a Way to Think
- selecting which new idea to develop
Memories and Representations
Minsky defines representation to mean any structure inside one's brain that one can use to answer some questions
What distinguishes us from other animals? It is our ability to treat ideas as though they were things, our ability to conceptualise. Minsky argues that there must be representation structures (networks) inside our brains. Knowledge fragments don't have meanings unless they are linked. He discusses various possible ways to represent knowledge:
- Describing events as stories or scripts
- Describing structures with semantic networks
- Using trans-frames to represent actions
- Using frames to embody commonsense knowledge
- Learning by building "knowledge Lines"
Connectionist and Statistical Representations
He contrasts two different ways to represent an apple, through a semantic (symbolic) network and a connectionist network
Connectionist networks (based on numbers showing strength of associations) can learn to recognise many important types of patterns, without any need for a person to program them. But number based networks have limitations. Every relationship is reduced to a number or strength so there remains almost no trace of the evidence that led to it, eg. the number 12 could represent all sorts of things
I see the popularity (of Connectionist Networks). in recent years, as having retarded the search for higher level ideas about human psychological machinery... research on commonsense thinking kept advancing until about 1980, but then it was clearly recognised that further progress would need ways to acquire and organise millions of fragments of commonsense knowledge. That prospect seemed so daunting that most researchers decided to try, instead, to invent machines that could learn, by themselves, all the knowledge that they would need - in short, to invent new kinds of "baby machines" ...How do we learn new representations?
Quite a few of these learning machines did indeed learn to do some useful things, but none of them went on to develop higher-level reflective Ways to Think - and I suspect that this was mainly because they tried to represent knowledge in numerical terms....
... I do not mean to suggest that such networks are not important ... it seems safe to assume that many of the low level processes in our brains must use some form of Connectionist Networks (pp. 289-91)
Kant 1787: ... experience and sensory knowledge is only part of knowledge ... cognition adds new knowledge
Minsky thinks we are born with primitive forms of structures like K-lines, Frames and Semantic Networks which are then built on to create representations
Which representations to use for which purposes?
A dialogue between different approaches to the best way to represent knowledge:
Mathematician: It is always best to express things with logic
Connectionist: No, logic is far too inflexible to represent commonsense knowledge. Instead, you ought to use Connectionist Networks
Linguist: No, because Connectionist Nets are even more rigid. They represent things in numerical ways that are hard to convert to useful abstractions. Instead, why not simply use everyday language - with its unrivaled expressiveness
Conceptualist: No, language is much too ambiguous. You should use Semantic Networks instead - in which ideas get connected by definite concepts!
Statistician: Those linkages are too definite and don't express the uncertainties we face, so you need to use probabilities
Mathematician: All such informal schemes are so unconstrained that they can be self contradictory. Only logic can ensure us against those circular inconsistencies