This is part 2 of the CM Topologies series. Part 1 is here

So you’ve read the CM paper and are ready for some RNN wiring. Great!

First off, quick recap of the algorithm for training C and M:

Algorithm 1: Train C and M in Alternating Fashion

  1. Initialize C and M and their weights.
  2. Freeze M’s weights such that they cannot change while C learns.
  3. Execute a new trial by generating a finite action sequence that prolongs the history of actions and observations. Actions may be due to C which may exploit M in various ways (see Sec. 5). Train C’s weights on the prolonged (and recorded) history to generate action sequences with higher expected reward, using methods of Sec. 5.
  4. Unfreeze M’s weights, and re-train M in a “sleep phase” to better predict/compress the prolonged history; see Sec. 4.
  5. If no stopping criterion is met, goto 2

So C is using M somehow. But how exactly ? Let’s review the different ways M and C can be connected together.

  • Black box:

A simple way would be to just consider M as a separate black box, send activations to M, and read off the results. This is rather limited because C does not have access to the internals of M.

  • Connected as normal units. (this is the approach described in the paper)

Consider an RNN C (with typically rather small feasible search space) as in Sec. 5.2. We add standard and/or multiplicative learnable connections (Sec. 3.1) from some of the units of C to some of the units of the typically huge unsupervised M, and from some of the units of M to some of the units of C. The new connections are said to belong to C. C and M now collectively form a new RNN called CM, with standard activation spreading as in Sec. 3.1. The activations of M are initialized to default values at the beginning of each trial. Now CM is trained on RL tasks in line with step 3 of algorithm 1, using search methods such as those of Sec. 5.2 (compare Sec. 1). The (typically many) connections of M, however, do not change—only the (typically relatively few) connections of C do.

We connect C and M together and consider the whole as a new RNN. We expect the weights of C to change, and the many weights of M to remain stable.

What does that mean? It means that now C’s relatively small candidate programs are given time to “think” by feeding sequences of activations into M, and reading activations out of M, before and while interacting with the environment. Since C and M are general computers, C’s programs may query, edit or invoke subprograms of M in arbitrary, computable ways through the new connections. Given some RL problem, according to the AIT argument (Sec. 2.1), this can greatly accelerate C’s search for a problem-solving weight vector w, provided the (time-bounded [147]) mutual algorithmic information between w and M’s program is high, as is to be expected in many cases since M’s environment-modeling program should reflect many regularities useful not only for prediction and coding, but also for decision making.

C can reuse the concepts encoded in M, and thus plan and act in a higher level of abstraction. (as high as M permits)

This simple but novel approach is much more general than previous computable, but restricted, ways of letting a feedforward C use a model M (Sec. 1.3.1)[301, 273][245, Sec. 6.1], by simulating entire possible futures step by step, then propagating error signals or temporal difference errors backwards (see Section 1.3.1). Instead, we give C’s program search an opportunity to discover sophisticated computable ways of exploiting M’s code, such as abstract hierarchical planning and analogy based reasoning. For example, to represent previous observations, an M implemented as an LSTM network (Sec. 1.2) will develop high-level, abstract, spatio-temporal feature detectors that may be active for thousands of time steps, as long as those memories are useful to predict (and thus compress) future observations [62, 61, 189, 79]. However, C may learn to directly invoke the corresponding “abstract” units in M by inserting appropriate pattern sequences into M. C might then short-cut from there to typical subsequent abstract representations, ignoring the long input sequences normally required to invoke them in M, thus quickly anticipating a few possible positive outcomes to be pursued (plus computable ways of achieving them), or negative outcomes to be avoided.

C has access to the whole repertoire of concepts of M, and can directly activate them where necessary. This might be the mechanism at work while planning, imagining or reasoning.



Some remarks/questions:

  • What would happen if C makes a connection to a unit of M, and that unit changes or dies ?
  • Where should C makes its connection to M ? In the input, output or hidden layer ? Or all of these ? Can the learning algorithm figure out where C should connect ?
  • Why C and M should be trained in an alternating fashion ? What are the sources of instabilities if M and C are trained at the same time ?