NOTE-352: K-Lines: A Theory of Memory

Contribution

The memo replaces the store-and-retrieve picture of memory with reinstatement. To remember is to reactivate some of the agents that were active when something worked. It adds the three constraints that make this more than a slogan:

• Level band. Reactivate only an intermediate band of levels.
• Recursion. Attach mainly to earlier memories.
• Control. Let a third agency decide what is memorable.

It also offers a distributed reading of frames.

Key insight

A memory should make you approach the present the way you approached a past success. It should not make you see the past again, and it should not hand you the old conclusion. A K-line therefore sets the middle of the processing hierarchy. Low levels stay free to report what is actually there, and high levels stay free to pursue the current goal. Partial reinstatement in that band is what turns recall into analogy.

Assumptions

• Society of Mind. The mind is many partially autonomous agents in large “Divisions”. In the simplest version each agent is active or quiet, a total mental state is the set of active agents, and a partial state fixes only some of them.
• Mostly upward flow. Agents take inputs from below or the side and send outputs up or sideways, so information moves “only upwards, on the whole”. Seen from any agent, this gives a “P-pyramid”, which is “an illusion of an agent’s perspective”.
• Cross-exclusion groups. Mutual inhibition within small groups means that forcing one member on resets the group, and a forced state persists. This is built-in short-term memory, seen from outside as “dispositions”.
• Neocortex only. The theory is meant to cover the common properties of neocortex, not the brain at large (Note 6).
• No neurological detail is claimed. It is “just another form of information processing theory, emphasizing control structure and data flow” (Note 4).

Key results

These are proposals, stated as principles. There are no proofs and no simulations.

• K-node assignment and K-line attachment. When a part G declares a mental event in P memorable, a new K-node is made. Its K-line makes excitatory attachments to every currently active P-agent. No negative connections are needed, because cross-exclusion suppresses the rivals (Note 7).
• Level-band principle.
  • Lower limit: the K-line must not reach far below its level.
  • Upper limit: it must not reach close to its own level.
  • So it spans “only a band of levels somewhere below that of PK”. The author concedes the upper limit is “much less clear” (Note 9).
• K-recursion principle. “Attach the new K-line … to just the currently-active K-nodes.” K-agents lie near P-agents of the same level, and over development the preference shifts from P to K.
• Crossbar problem.
  • The answer is sought in the Society of Mind, not in clever coding: if most agents need not talk to each other, local crossbars suffice.
  • The scale is a few thousand P-nets of a few thousand agents each (Note 10).
  • Sparse random subset codes on shared “M-lines” are offered as a wiring economy (Note 10). For example, 10-line subsets of a 100-line bundle, after Mooers’s Zatocoding and Willshaw, Buneman and Longuet-Higgins.
• K-knowledge.
  • Logical: K-lines at the same level combine disjunctively. With “drop out” cross-exclusion, conflicts default upward, which the memo credits to Papert’s theory of Piagetian conservation.
  • Abstract: accumulated instances plus the cancelling of conflicts yield “the extraction of common, non-conflicting properties”.
  • Procedural: K-lines interacting across levels could partly instantiate one frame with others.
• Learning and reinforcement. No “simplistic, centralized reward mechanism” can suffice, because deciding what is memorable “requires too much intelligence”. Credit assignment across strategy and tactic time scales defeats recency. Minimal learning needs three nets, G, K and P, “in which the first controls how the second learns to operate the third”.
• Tacit and explicit knowledge. In a society theory the puzzle is how any knowledge becomes explicit. “Most knowledge stays more or less where it was formed, and does its work there.” “Self-awareness is a complex, constructed illusion.”

Claims

id	claim	strength	support
C1	The function of a memory is to re-create a partial mental state, not to retrieve a stored description	assertion, framed as a thesis	stated and illustrated (the concert novice and professional); no evidence
C2	Re-activating only an intermediate level band yields analogy: the present seen as an instance of the past	moderate as a design argument	the lower-limit argument is clear; the upper limit is admitted to be “much less clear” (Note 9)
C3	With cross-exclusion, accumulating instances on one K-node produces abstraction by cancelling conflicts	moderate as a mechanism sketch	follows from the stated drop-out rule; not simulated
C4	K-lines with weak lower fringes implement frames with default assignments	moderate	Note 9 gives the correspondence; the upper-fringe half is speculative
C5	Global recency-based reinforcement cannot solve human credit assignment; at least three nets are needed	weak to moderate	argued from multi-time-scale strategies; no model
C6	The crossbar problem needs no general solution, because most agents need not communicate	assertion	follows from the Society of Mind premise, which is assumed
C7	Attitudes and feelings precede propositions in development, and concrete recollection needs the most expertise	assertion	analogy (the concert); no developmental data

Concepts

• Agent. A unit that recognises configurations of a few associates and changes state. In the simplest model it is active or quiet.
• Partial mental state. A specification of the states of some agents. Compatible partial states can be held at once.
• P-pyramid. The hierarchy below a given agent, from its perspective.
• K-node, K-line. A memory unit, and its “wire having potential connections to every Agent in the P-pyramid”.
• Level band. The intermediate span of levels a K-line may reach.
• Cross-exclusion. Mutual inhibition within small agent groups.
• K-pyramid. The K-nodes’ structure, mirroring the P-pyramid with activity flowing downward. Together they make a local “counterclockwise spiral” of computation.
• Disposition. “A momentary range of possible behaviors” (Note 2).
• M-lines. A shared bundle on which K-lines are simulated by sparse random subsets (Note 10).

Connections

• Plain Talk. It extends Plain Talk (LIT-399), its reference [1]. It “complements” that paper, maps its c-lines onto K→P connections, and cites it for P-net internals and the claim that not all P-nets need to communicate.
• Frames. It extends the frame memo (LIT-410), its reference [3], with K-nodes as distributed frames and weak lower fringes as defaults (Note 9).
• Named sources. Papert, for the basic idea and the conservation principle. Winston’s near-miss learning, whose emphasis links are “identified with K-lines to members of cross-exclusion groups”, while its prevention pointers have no easy equivalent (Note 11). Doyle’s truth-maintenance in and out flags. Hebb and Marr. Mountcastle on excitatory cortical inputs. Miller, for the resemblance to the older notion of “redintegration”.
• Jokes memo. The jokes memo (LIT-408) cites this paper for partial mental states and for the claim that memories are not all “made of the same stuff”.

Bearing on the record

• Re-enactment accounts of memory.

  • Damasio’s convergence zones (LIT-385, Deferred) make recall the reconstruction of fragments under feedback from convergence zones.
  • The somatic-marker hypothesis (LIT-384) re-enacts learned body states to bias decisions.

  Both share this memo’s re-enactment shape and its dispositions-first emphasis. Neither is a K-line theory, and the pairing is mine. A reading of LIT-385 should check whether Damasio cites Minsky.

• Sparse distributed memory. Note 10’s sparse random-subset coding comes from the associative-memory line (Willshaw et al. 1969) that sparse distributed memory belongs to. The record reads SDM through Bricken and Pehlevan (LIT-269, NOTE-242). There retrieval is reconstruction by similarity of address. Here it is reinstatement of the agents that were active. These are different mechanisms with a common coding trick.

• Consciousness. “Self-awareness is a complex, constructed illusion” sits on the illusionist side of the record’s consciousness readings. The memo asserts it in one paragraph and does not argue it.

• Boundary. No THEORY in the record should rest on this memo. It carries no machine-learning practice instruction, and no anthology topic holds it.

Limitations

• Half-built, by its author’s account. “This ends the constructive part of this essay.” The P→K association, how events in P get tied to goals, is never specified.
• Not simulated. Nothing shows that level-band reinstatement produces useful analogy, or that cross-exclusion cancelling produces sensible abstraction, rather than mush.
• Memory only grows. Connections are only added, never removed. Note 12 lists possible remedies and settles on none.
• Neocortex only. The scope is limited to neocortex (Note 6).
• No sequence. Sequential activity is “suppressed” throughout (Note 11).

Open questions

• What is the P→K connection that relates perceptual events to goals?
• Does reinstating a middle band of a processing hierarchy, in any built system, give transfer to new cases without copying old answers? This would test C2.
• How are K-lines pruned or revised (Note 12)?

Corrections

• none to a seeded skim (there was no seed).
• Citation. The brief is right that the article is in Cognitive Science 4(2):117–133 (1980), and right on the memo, AIM-516, June 1979. Crossref holds two DOIs for the article: the Wiley DOI 10.1207/s15516709cog0402_1, issued April 1980, and an Elsevier-era DOI 10.1016/s0364-0213(80)80014-0, issued June 1980. The filing uses the Wiley one, the brief’s “likely” DOI, which resolves. The DSpace title is “K-Lines: A Theory of Memory”.
• Which text was read. The memo, not the journal article. The journal version was not reachable.