How to build a General Intelligence: An interpretation of the biology

Figure 1: Our interpretation of the Thalamocortical system as 3 interacting sub-systems (objective, subjective and executive). The structure of the diagram indicates the dominant direction of information flow in each system. The objective system is primarily concerned with feed-forward data flow, for the purpose of building a representation of the actual agent-world system.  The executive system is responsible for making desired future agent-world states a reality. When predictions become observations, they are fed back into the objective system. The subjective system is a circular because its behaviour depends on internal state as much as external. The subjective system builds a filtered, subjective model of observed reality, that also represents objectives or instructions for the executive. This article will describe how this model fits into the structure of the Thalamocortical system.

Authors: David Rawlinson and Gideon Kowadlo

This is part 4 of our series on how to build an artificial general intelligence (AGI).

• Part 1: An overview of hierarchical general intelligence
• Part 2: Reverse engineering (the physical perspective - cells and layers - and the logical perspective - a hierarchy).
• Part 3: Circuits and pathways; we introduced our canonical cortical micro-circuit and fitted pathways to it.

In this article, part 4, we will try to interpret all the information provided so far. We will try to fit what we know about the biological general intelligence to our theoretical expectations.

Systems

We believe cortical activity can be usefully interpreted as 3 integrated systems. These are:

• Objective system
• Subjective system
• Executive system

So, what are these systems, why are they needed and how do they work?

Objective System

We theorise that the purpose of the objective system is to construct a hierarchical, generative model of both the external world and the actual state of the agent. This includes internal plans & goals already executed or in progress. From our conceptual overview of General Intelligence we think that this representation should be distributed, compositional and therefore robust and able to immediately model novel situations instantly and meaningfully.

The objective system models varying timespans depending on the level of abstraction, but events are anchored to the current state of the world and agent. Abstract events may cover long periods of time - for example, “I made dinner” might be one conceptual event.

We propose that the objective system is implemented by pyramidal cells in layers 2/3 and by spiny excitatory cells in layer 4. Specifically, we suggest that the purpose of the spiny excitatory cells is primarily dimensionality reduction, by performing a classifier function, analogous to the ‘Spatial Pooling’ function of Hawkins’ HTM theory. This is supported by analysis of C4 spiny stellate connectivity: “... spiny stellate cells act predominantly as local signal processors within a single barrel...”. We believe the pyramidal cells are more complex and have two functions. First, they perform dimensionality reduction by requiring a set of active inputs on specific apical (distal) dendrite branches to be simultaneously observed before the apical dendrite can output a signal (an action potential). Second, they use basal (proximal) dendrites to identify the sequential context in which the apical dendrite has become active. Via a local competitive process, pyramidal cells learn to become active only when observing a set of specific input patterns in specific historical contexts.

The output of pyramidal cells in C2/3 is routed via the Feed-Forward Direct pathway to a “higher” or more abstract cortical region, where it enters in C4 (or in some parts of the Cortex, C2/3 directly). In this “higher” region, the same classifier and context recognition process is repeated. If C4 cells are omitted, we have less dimensionality reduction and a greater emphasis on sequential or historical context.

We propose these pyramidal cells only output along their axons when they become active without entering a “predicted” state first. Alternatively, interneurons could play a role in inhibiting cells via prediction to achieve the same effect. If pyramidal cells only produce an output when they make a False-Negative prediction error (i.e. they fail to predict their active state), output is equivalent to Predictive Coding (link, link). Predictive Coding produces an output that is more stable over time, which is a form of Temporal Pooling as proposed by Numenta.

To summarize, the computational properties of the objective system are:

1. Replace simultaneously active inputs with a smaller set of active cells representing particular sub-patterns, and
2. Replace predictable sequences of active cells with a false-negative error coding to transform the output into a simpler sequence of prediction errors

These functions will achieve the stated purpose of incrementally transforming input data into simpler forms with accumulating invariances, while propagating (rather than hiding) errors, for further analysis in other Columns or cortical regions. In combination with a tree-like hierarchical structure, higher Columns will process data with increasing breadth and stability over time and space.

The Feed-Forward direct pathway is not filtered by the Thalamus. This means that Columns always have access to the state of objective system pyramidal cells in lower columns. This could explain the phenomenon that we can process data without being aware of it (aka “Blindsight”); essentially the objective system alone does not cause conscious attention. This is a very useful quality, because it means the data required to trigger a change in attention is available throughout the cortex. The “access” phenomenon is well documented and rather mysterious; the organisation of the cortex into objective and subjective systems could explain it.

Another purpose of the objective system is to ensure internal state cannot become detached from reality. This can easily occur in graphical models, when cycles form that exclude external influence. To prevent this, we believe that the roles of feed-forward and feed-back input must be separated to break the cycles. However, C2/3 pyramidal cells’ dendrites receive both feed-forward (from C4) and feed-back input (via C1).

One way that this problem might be avoided is by different treatment of feed-forward and feed-back input, so that the latter can be discounted when it is contradicted by feed-forward information. There is evidence that feed-forward and feedback signals are differently encoded, which would make this distinction possible.

We speculate that the set of states represented by the cells in C2/3 could be defined only using feed-forward input, and that the purpose of feedback data in the objective system is restricted to improved prediction, because feedback contains state information from a larger part of the hierarchy (see figure 2).

Figure 2: The benefit of feedback. This figure shows part of a hierarchy. The hierarchy structure is defined by the receptive fields of the columns (shown as lines between cylinders, left). Each Column has receptive fields of similar size. Moving up the hierarchy, Columns receive increasingly abstract input with a greater scope, being at the top of a pyramid of lower Columns whose receptive fields collectively cover a much larger area of input. Feedback has the opposite effect, summarizing a much larger set of Column states from elsewhere and higher in the hierarchy. Of course there is information loss during these transfers, but all data is fully represented somewhere in the hierarchy.

So although the objective system makes use of feedback, the hierarchy it defines should be predominantly determined by feed-forward information. The feed-forward direct pathway (see figure 3) enables the propagation of this data and consequently the formation of the hierarchy.

Figure 3: Feed-Forward Direct pathway within our canonical cortical micro-circuit. Data travels from C4 to C2/3 and then to C4 in a higher Column. This pattern is repeated up the hierarchy. This pathway is not filtered by the Thalamus or any other central structure, and note that it is largely uni-directional (except for feedback to improve prediction accuracy). We propose this pathway implements the Objective System, which aims to construct a hierarchical generative model of the world and the agent within it.

Subjective System

We think that the subjective system is a selectively filtered model of both external and internal state including filtered predictions of future events. We propose that filtering of input constitutes selective attention, whereas filtering of predictions constitutes action selection and intent. So, the system is a subjective model of reality, rather than an objective one, and it is used for both perception and planning simultaneously.

The time span encompassed by the system includes a subset of both present and future event-concepts, but as with the objective system, this may represent a long period of real-world time, depending on the abstraction of the events (for example, “now” I am going to work, and “next” I will check my email [in 1 hour’s time]).

It makes good sense to have two parallel systems, one filtered (subjective) and one not (objective). Filtering external state reduces distraction and enhances focus and continuity. Filtering of future predictions allows selected actions to be maintained and pursued effectively, to achieve goals.

In addition to events the agent can control, it is important to be aware of negative outcomes outside the agent’s control. Therefore the state of the subjective system must include events with both positive and negative reward outcomes. There is a big difference between a subjective model and a goal-oriented planning model. The subjective system should represent all outcomes, but preferentially select positive outcomes for execution.

The subjective system represents potential future states, both internal and external. It does not necessarily represent reality; it represents a biased interpretation of intended or expected outcomes based on a biased interpretation of current reality! These biases and omissions are useful; they provide the ability to “imagine’ future events by serially “predicting” a pruned tree of potential futures.

More speculatively, differences between the subjective and objective systems may be the cause of phenomena such as selective awareness and “access” consciousness.

Figure 4: Feed-Forward Indirect pathway, particularly involved in the Subjective system due to its influence on C5. The Thalamus is involved in this pathway, and is believed to have a gating or filtering effect. Data flows from the Thalamus to C4, to C2/3, to C5 and then to a different Thalamic nuclei that serves as the input gateway to another cortical Column in a different region of the Cortex. We propose that the Feed-Forward Indirect pathway is a major component of the subjective system.

Figure 5:  The inhibitory micro-circuit, which we suggest makes the subjective system subjective! The red highlight shows how the Thalamus controls activity in C5 by activating inhibitory cells in C4. The circuit is completed by C5 pyramidal cells driving C6 cells that modulate the activity of the same Thalamic nuclei that selectively activate C5.

The subjective system primarily comprises C5 (where subjective states are represented) and the Thalamus (which controls subjectivity), but it draws input from the objective system via C2/3. The latter provides context and defines the role and scope (within the hierarchy) of C5 cells in a particular column. Between each cortical region (and therefore every hierarchy level), input to the subjective system is filtered by the Thalamus (figure 5). This implements the selection process. The Feed-Forward Indirect pathway includes these Thalamo-Cortical loops.

We suggest the Thalamus implements selection within C5 using special cells in C4 that are activated by axons (outputs) from the Thalamus (see figure 6). These inhibitory C4 cells target C5 pyramidal cells and inhibit them from becoming active. Therefore, thalamic axons are both informative (“this selection has been made”) and executive (the axon drives inhibition of selected C5 pyramidal cells).

Figure 6: Thalamocortical axons (afferents) are shown driving inhibitory cells in C4 (leftmost green cell) that in turn inhibit pyramidal cells in C5 (red). They also provide information about these selections to other layers, including C2/3. When a selection has been made, it becomes objective rather than subjective, hence provision of a copy to C2/3. Image source.

Note that selection may be a process of selective dis-inhibition rather than direct control: Selection alone may not be enough to activate the C5 cells.  Instead, C5 pyramidal cells likely require both selection by the Thalamus, and feed-forward activation via input from C2/3. The feed-forward activation could occur anywhere within a window of time in which the C5 cell is “selected”.  This would relax timing requirements on the selection task, making control easier; you only need to ensure that the desired C5 cell is disinhibited when the right contextual information arrives from other sources (such as C2/3). This also ensures C5 cell activation fits into its expected sequence of events and doesn’t occur without the right prior context.

C5 also benefits from informational feedback from higher regions and neighbouring cells that help to define unique contexts for the activation of each cell.

We suggest that C5 pyramidal cells are similar to C2/3 pyramidal cells but with some differences in the way the cells become active. Whereas C2/3 cells require both matching input via the apical dendrites and valid historical input to the basal dendrites to become active, C5 cells additionally need to be disinhibited for full activation to occur.

As mentioned in the previous article, output from C5 cells sometimes drives motors very directly, so full activation of C5 cells may immediately result in physical actions. We can consider C5 to be the “output” layer of the cortex. This makes sense if the representation within C5 includes selected future states.

Management of C5 activity will require a lot of inhibition; we would expect most of the input connections to C5 to be inhibitory because in every context, for every potential outcome, there are many alternative outcomes that must be inhibited (ignored). At any given time, only a sparse set of C5 cells would be fully active, but many more would be potentially-active (available for selection).

Given predictive encoding and filtering inhibition, it would be common for few pyramidal cells to be active in a Column at any time. Separately, we would expect objective C2/3 pyramidal activity to be more consistent and repeatable than subjective C5 pyramidal activity, given a constant external stimulus.

Executive System

So far we have defined a mechanism for generating a hierarchical representation and a mechanism for selectively filtering activity within that representation. In our original conceptual look at general intelligence, we also desired that filtering predictions would be equivalent to action selection. But if we have selected predictions of future actions at various levels of abstraction within the hierarchy, how can we make these abstract prediction-actions actually happen?

The purpose of the executive system is to execute hierarchical plans reliably. As previously discussed, this is no trivial matter due to problems such as vanishing agency at higher hierarchy levels. If a potential future outcome represented within the subjective system is selected for action, the job of the executive system is to make it occur.

We know that we want abstract concepts at high levels within the hierarchy to be faithfully translated into their equivalent patterns of activity at lower levels. Moving towards more concrete forms would result in increasing activity as the incremental dimensionality reduction of the feed-forward hierarchy is reversed.

Figure 7: Differences in dominant direction of data flow between objective and executive systems. Whereas the Objective system builds increasingly abstract concepts of greater breadth, the Executive system is concerned with decomposing these concepts into their many constituent parts. so that hierarchically-represented plans can be executed.

We also know that we need to actively prioritize execution of a high level plan over local prediction / action candidates in lower levels. So, we are looking for a cascade of activity from higher hierarchy levels to lower ones.

Figure 8: One of two Feed-Back direct pathways. This pathway may well be involved in cascading control activity down the hierarchy towards sensors and motors. Activity propagates from C6 to C6 directly; C6 modulates the activity of local C5 cells and relevant Thalamic nuclei that activate local C5 cells by selective disinhibition in conjunction with matching contextual information from C2/3.

It turns out that such a system does exist: The feed-back direct pathway from C6 to C6. Cortex layer 6 is directly connected to Cortex layer 6 in the hierarchy levels immediately below. What’s more, these connections are direct, i.e. unfiltered (which is necessary to avoid the vanishing agency problem). Note that C5 (the subjective system) is still the output of the Cortex, particularly in motor areas. C6 must modulate the activity of cells in C5, biasing C5 to particular predictions (selections) and thereby implementing a cascading abstract plan. Finally, C6 also modulates the activity of Thalamic nuclei that are responsible for disinhibiting local C5 cells. This is obviously necessary to ensure that the Thalamus doesn’t override or interfere with the execution of a cascading plan already selected at a higher level of abstraction.

Our theory is that ideally, all selections originate centrally (e.g. in the Thalamus). When C5 cells are disinhibited and then become predicted, an associated set of local C6 cells is triggered to make these C5 predictions become reality.

These C6 cells have a number of modulatory outputs to achieve this goal:

• They remove local competition by directly inhibiting C5 cells in the local Column (see link, link, link)
• They drive C6 cells in the hierarchy level below, in Columns that have a Feed-Forward relationship with the local column. In effect, the execution of the task is delegated to a set of hierarchically-lower C6 cells that are recruited to the task.
• They modulate cells in the Thalamus that provide input to the local Column, to ensure that the desired C5 cells in the local column are disinhibited.

Executive Training

No, this is not a personal development course for CEOs. This section checks whether C6 cells can learn to replay specific action sequences via C5 activity. This is an essential feature of our interpretation, because only C6 cells participate in a direct, modulatory feedback pathway.

We propose that C6 pyramidal neurons are taught by historical activity in the subjective system. Patterns of subjective activity become available as “stored procedures” (sequences of disinhibition and excitatory outputs) within C6.

Let’s start by assuming that C6 pyramidal cells have similar functionality to C2/3 and C5 pyramidal cells, due to their common morphhology. Assume that C5 cells in motor areas are direct outputs, and when active will cause the agent to take actions without any further opportunity for suppression or inhibition (see previous article).

In other cortical areas, we assume that the role of C5 cells is to trigger more abstract “plans” that will be incrementally translated into activity in motor areas, and therefore will also become actions performed by the agent.

To hierarchically compose more abstract action sequences from simpler ones, we need activity of an abstract C5 cell to trigger a sequence of activity in more concrete C5 cells. C6 cells will be responsible for linking these C5 cells. So, activating a C6 cell should trigger a replay of a sequence of C5 cell activity in a lower Column. How can C6 cells learn which sequences to trigger, and how can these sequences be interpreted correctly by C6 cells in higher hierarchy levels?

C6 pyramidal cells are mostly oriented with their dendrites pointing towards the more superficial cortex layers C1,...,C5 and their axons emerging from the opposite end. Activity from C5 to C6 is transferred via axons from C5 synapsing with dendrites from C6. Given a particular model of pyramidal cell learning rules, C6 pyramidal cells will come to recognize patterns of simultaneous C5 activity in a specific sequential context, and C6 interneurons will ensure that unique sets of C6 pyramidal cells respond in each context.

So how will these C6 cells learn to trigger sequences of C5 cells? We know that the axons of C6 cells bend around and reach up into C5, down to the Thalamus and directly to hierarchically-lower C6 cells. At all targets they can be excitatory or inhibitory.

All we need beyond this, is for C6 axons to seek out axon target cells that become active immediately after the originating C6 cell is stimulated by active C5 cells. This will cause each C6 cell to trigger C5 and C6 cells that are observed to be activated afterwards. Note that we require the C6 cells themselves be organised into sequences (technically, a graph of transitions).

Target seeking by axons is known as “Axon Guidance” and C6 pyramidal cells’ axons do seem to target electrically active cells by ceasing growth when activity is detected. We have not yet found biological evidence for the predicted timing.

C6 axons can also target C4 inhibitory cells (evidence) and Thalamic cells, which again is compatible with our interpretation, as long as they are cells that become active after the originating C6 cell. If we want to “replay” some activity that followed a particular C6 cell, then all the cells described above should be excited or inhibited to ensure that the same events occur again. Activating a C6 cell directly should reproduce the same outcome as incidental activation of the C6 cell via C5 - a chain of sequential inhibition and promotion will result. Note that the same learning rule could work to discover all axon targets mentioned.

Collectively, the C6 cells within a Column will become a repertoire of “stored procedures” that can be triggered and replayed by a cascade of activity from higher in the hierarchy or by direct selection via C5. C6 cells would behave the same way whether activated by local C5 cells, or by C6 cells in the hierarchy level above. This allows cascading, incremental execution of hierarchical plans.

C6 cells do not need to replace sequences of C5 cell activity with a single C6 cell (i.e. label replacement for symbolic encoding), but they do need to collectively encode transitions between chains of C5 cells, individually trigger at least 1 C5 cell and collectively allow a single C6 cell to trigger a sequence of C6 cells in both the current and lower hierarchy regions.

C6 interneurons can resolve conflicts when multiple C6 triggers coincide within a column. We can expect C6 interneurons to inhibit competing C6 pyramidal cells until the winners are found, resulting in a locally consistent plan of action.

As with layers C2/3 and C5, C6 inhibitory interneurons will also support training C6 pyramidal cells for collective coverage of the space of observed inputs, in this case from C5 and C2/3.

Bootstrapping

Now we are only left with a bootstrapping problem: How can the system develop itself? Specifically, how do the sequences of C5 activity come to be defined so that they can be learned by C6?

We suggest that conscious choice of behaviour via the Thalamus is used to build the hierarchical repertoire from simple primitive actions to increasingly sophisticated sequences of fine control. Initially, thalamic filtering of C5 state would be used to control motor outputs directly, without the involvement of C6. Deliberate practice and repetition would provide the training for C6 cells to learn to encode particular sequences of behaviour, making them part of the repertoire available to C6 cells in hierarchically “higher” Columns.

Initially, concentration is needed to perform actions via direct C5 selections; these activities need to be carefully centrally coordinated using selective attention. However, when C6 has learnt to encode these sequences, they become both more reliable and require less effort to execute, requiring only a trigger to one C6 cell.

After training, only minimal thalamic interventions are needed to execute complex sequences of behaviour learned by C6 cells. Innovation can continue by combining procedures encoded by C6 with interventions via the Thalamus, that can still excite or inhibit C5 cells. However, in most other cases C6 training is accelerated by the independence of Columns: When a C6 cell learns to control other cells within the Column, this learning remains valid no matter how many higher hierarchy levels are placed on top. By analogy, once you’ve learned to drink from a cup, you don’t need to relearn that skill to drink in restaurants, at home, at work etc.

As C6 learns and starts to play a role in the actions and internal state of the agent, it becomes important to provide the state of C6 to the objective and subjective systems as contextual input.

Axons from C6 to other, hierarchically lower Columns take two paths: To C6, and to C1. We propose that the copy provided to C1 is used as informational feedback in C2/3 and C5 pyramidal cells (these axons synapse with Pyramidal cell Apical dendrites). We suggest the copy to C6 allows C6 cells to execute plans hierarchically, by delegating execution to a number of more concrete C6 cells. Therefore, the feedback direct pathway from C6 to C6 is part of the executive system. These axons should synapse on cell bodies, or nearby, to inhibit or trigger C6 activation artificially (rather than via C5).

Interpretation of the Thalamus

Rather than as merely a relay, we propose that a better concept of the Thalamus is as a control centre. It’s job is to centrally control cortical activity in C5 (the subjective system). Abstract activity in C5 is propagated down the hierarchy by C6, and translated into its concrete component states, eventually resulting in specific motor actions. Therefore, via this feedback pathway the filtering performed by the Thalamus assumes an executive role also.

We believe that filtering predictions of oneself performing an action or experiencing a reward is the mechanism by which objectives and plans are selected. We believe there is only one representation of the world in our heads. There is no separate “goal-oriented” or “action-based” representation. This means that filtering predictions is the mechanism of behaviour generation. Note that in a hierarchical system, you can simultaneously select novel combinations of predictions to achieve innovation without changing the hierarchical model.

Our interpretation of the Thalamus depends on some theoretical assumptions about how general intelligence works. Crucially, we believe there is no difference between selective awareness of externally-caused and self-generated events, except some of the latter have agency in the real world via the agent’s actions. This means that selective attention and action selection can both be consequences of the same subjective modelling process.

But where does selection actually occur?

For a number of practical reasons, action and attentional selection should be centralized functions. For one thing, the reward criteria for selecting actions are of much smaller dimension than the cortical representations - for example, the set of possible pain sensations are far more limited than the potential external causes of pain. We essentially need to compare the reward of all potential actions against each other, rather than an absolute scale.

It is also important that conflicts between items competing for attention or execution are resolved so that incompatible plans are replaced by a single clear choice. Conflict resolution is difficult to do in a highly parallel & distributed system; instead, it is preferable to force all alternatives to compete against each other until a few clear winners are found.

Finally, once an action or attentional target is selected, it should be maintained for a long period (if still relevant), to avoid vacillation. (See Scholarpedia for a good introduction to the difficulties of conflict resolution and the importance of sticking to a decision for long enough to evaluate it).

We believe the Thalamus plays this role via its interactions with the Cortex. It interacts with the Cortex in two ways. First, the Thalamus selectively dis-inhibits particular C5 cells, allowing them to become active when the right circumstances are later observed objectively (i.e. via C2/3, which is not subjective).

Second, the Thalamus must also co-operate with the Feed-Back cascade via C6.  While the Thalamus generates new selections by controlling C5, it must also permit the execution of existing, more abstract Thalamic selections by allowing cascading feedback activity to override local selections. Together, these mechanisms ensure that execution of abstract plans is as easily accomplished as simpler, concrete actions.

Interpretation of the Basal Ganglia

The Basal Ganglia are involved in so many distinct functions that they can’t be fully described within this article. They consist of a set of discrete structures located adjacent to the Thalamus.

In our model, selection is implemented by the Thalamus manipulating the subjective system within the Cortex. We propose that the selections themselves are generated by the Basal Ganglia, which then controls the behaviour of the Thalamus.

Crucially, we believe the Striatum within the Basal Ganglia uses reward values (such as pleasure and pain) to make adaptive selections. In other words, the Basal Ganglia are responsible for picking good actions, biasing the entire Thalamo-Cortical system towards futures that are expected to be more pleasant for the agent.

However, to make adaptive choices it is necessary to have accurate context and predictions (candidate actions). The hierarchical model defined within the Cortex is an efficient and powerful source for this data, and in fact, this pathway (Cortex → Basal Ganglia → Thalamus → Cortex) does exist within the brain (see figure 9 below).

Thanks to studies of relevant disorders such as Parkinson’s and Huntingdon’s, it is known that this pathway is associated with behaviour initiation and selection based on adaptive criteria.

Figure 9: Pathways forming a circuit from Cortex to Basal Ganglia to Thalamus and back to Cortex. Image source.

Lifecycle of an idea

Using our interpretation of biological general intelligence, we can follow the lifecycle of an idea from the conception to execution. Lets walk through the theorized response to a stimulus, resulting in an action.

Although the brain is operating constantly and asynchronously, we can define the start of our idea as some sensory data that arrives at the visual cortex. In this example, it’s an image of an ice-cream in a shop.

Objective Modelling

Sensor data propagates unfiltered up the Feed-Forward Direct pathway, activating cells in C4 and C2/3 in numerous cortical areas as it is transformed into its hierarchical form. The visual stimuli become a rich network of associated concepts, including predictions of near-future outcomes, such as experiencing the taste of ice-cream. These concepts represent an objective external reality and are now active and available for attention.

Subjective Prediction

Activity within the Objective system triggers activity in the Subjective system. Some C5 cells become “predicted”, but are inhibited by the Thalamus. These cells represent potential future actions and outcomes. Things that, from experience, we know are likely to occur after the current situation.

The Cortex projects data from C2/3 to the Striatum where it is weighted according to reward criteria. A strong response to the flavour of the frozen treat percolates through the Basal Ganglia and manipulates the activity of the Thalamus.

Between the Thalamus and the Cortex, an iterative negotiation takes place resulting in the selection (via dis-inhibition) of some C5 cells. The Basal Ganglia have learned which manipulations of the Thalamus maximize the expected Reward given the current input from Cortex.

The way that the Thalamus stimulates particular C5 cells is somewhat indirect. The path of activity to “select” C5 cells in layer n is C5[n-1] →  Thalamus → C4[n] → C5[n]. The signal is re-interpreted at each stage of this pipeline - that is, connections do not carry a specific meaning from point to point. Therefore, you can’t just adjust one “wire” to trigger a particular C5 cell. Rather, you must adjust the inhibition of input to many C4 → C5 cells until you’ve achieved the conditions to “select” a target C5 cell. Many target C5 cells might be simultaneously selected.

In addition to requiring disinhibition, C5 cells also wait for specific patterns of cell activity in C2/3 prior to becoming “predicted”. This means that it’s very difficult to select a C5 cell that is not “predicted”; it simply doesn’t have the support to out-compete its neighbours in the column and become “selected”. This prevents unrealistic outcomes being “selected”, or output commencing, before the right circumstances have arrived to match the expectation.

Eventually, a subset of C5 cells become “predicted” and “selected”, representing a subjective model of potential futures for the agent in the world. In this case, the anticipated future involves eating ice-cream.

Execution

When C5 cells become active, they in turn drive C6 pyramidal cells that are responsible for causing the future represented by “contextual, selected & predicted” C5 cells. In this case, C6 cells are charged with executing the high-level plan to “buy some ice-cream and eat it”.

The plan is embodied by many C5 cells, distributed throughout the hierarchy; each represents a subset of the “qualia” relating to the eating of ice-cream. C6 cells begin to interpret these C5 cells into concrete actions, via the C6-C6 Feed-Back Direct pathway. Crucially, they no longer require the Thalamus to modulate the input that makes C5 cells “selected”. Instead, C6 cells stimulate C5 and C6 cells in hierarchically-lower Columns directly, moving them to “selected” status and allowing them to become active as soon as the corresponding Feed-Forward evidence arrives to match.

C6 cells also modulate relay cells in the Thalamus, guiding the Thalamus to disinhibit C5 cells in lower hierarchy regions. This helps to ensure the parts of the decomposed plan are executed as intended. In turn, these newly selected “lower” C5 cells drive associated C6 cells, and the plan cascades down the hierarchy.

Note that the plan is also flowing in the “forward” direction, as it incrementally becomes reality rather than expectation. As motor actions take place, they are sensed and signalled through the Feed-Forward pathways. When C5 cells become “selected”, this information becomes available to higher columns in the hierarchy, if not filtered. This also helps the Feed-Forward Indirect pathway and C6 cells to keep track of activity and execute the plan in a coordinated manner.

At the lowest levels of the hierarchy, the plan becomes a sequence of motor activity, which is activated by C5 cells directly, and also by other brain components that are not covered by our general intelligence model.

A few moments later, the ice-cream is enjoyed, triggering a release of Dopamine into the Striatum and reinforcing the rewards associated with recent active Cortical input. Delicious!

Summary

In the previous articles we explored the characteristics of a general intelligence and looked at some of the features we expected it to have. In part 2 and part 3 we reviewed some relevant computational neuroscience research. In this article we’ve described our interpretation of this background material.

We presented a model of general intelligence built from 3 interacting systems - Objective, Subjective and Executive. We described how these systems could learn and bootstrap via interaction with the world, and how they could be implemented by the anatomy of the brain. As an example, we traced an experience from sensation, through planning and to execution.

Let’s assume that our understanding of biology is approximately correct. We can use this as inspiration to build an artificial general intelligence with a similar architecture and test whether the systems behave as described in these articles.

The next article in this series will look specifically at how these concepts could be implemented in software, resulting in a system that behaves much like the one described here.

How to Build a General Intelligence: Reverse Engineering

Figure 1: The physical architecture of general intelligence in the brain, namely the Thalamo-Cortical system. The system comprises a central hub (the Thalamus) surrounded by a surface (the Cortex, shown in blue). The Cortex is made up of a number of functionally-equivalent units called Columns. Each Column is a stack of cell layers. Columns share data, often via the central hub. The hub filters the data relayed between Columns.

Authors: Rawlinson and Kowadlo

This is part 2 of our series on how to build an artificial general intelligence (AGI). This article is about what we can learn from reverse-engineering mammalian brains. Part 1 is here.

The next few articles will try to interpret some well-established neuroscience in the context of general intelligence. We’ll ignore anything we believe is unrelated to general intelligence, and we’ll simplify things in ways that will hopefully help us to think about how general intelligence happens in the brain.

It doesn’t matter if we are missing some details, if the overall picture helps us understand the nature of general intelligence. In fact, excluding irrelevant detail will help, as long as we keep all the important bits!

These articles are not peer reviewed. Do assume everything here is speculation, even when linked to a source reference (our interpretation may be skewed). There isn’t space to repeatedly add this caveat throughout these articles.

1. Physical Architecture

First we’ll review and interpret the gross architecture of the brain, focusing on the Thalamo-Cortical system, which we believe is primarily responsible for general intelligence.

The Thalamo-Cortical system comprises a central hub (the Thalamus and Basal Ganglia), surrounded by a thin outer surface (the Cortex). The surface consists of a large number of functionally-equivalent units, called Columns. The Cortex is wrinkled so that it’s possible to pack a large surface area into a small space.

Why are the units called Columns? It’s the physical structure of their connectivity patterns. Cells within each Column are highly interconnected, but connections to cells in other Columns are fewer and less varied. Columns occupy the full thickness of the surface, approximately 6 distinct layers of cells. Since these layers are stacked on top of each other and are loosely connected between stacks, we have the appearance of a surface made of Columns.

Confusingly, there are both Macro and Micro-Columns, and these terms are used inconsistently. In these articles we will simply say ‘Column’ when referring to a Macro-Column as defined in a previous post.

In the previous article we described the ideal general intelligence as a structure made of many identical units that have each learned to play a small part in a distributed system. These theoretical units are analogous to Columns in real brains.

Columns can be imagined to be independent units that interact by exchanging data. However, data travelling between Columns often takes an indirect path, via the central hub.

The hub filters messages passed between Columns. In this way, the filter acts as a central executive that manages the distributed system made up of many Columns.

We believe this is a fundamental aspect of the architecture of general intelligence.

In mammalian brains, the filter function is primarily the role of the Thalamus, although its actions are supported and modulated by other parts, particularly the Basal Ganglia.

Other brain components, such as the Cerebellum, are essential for effective motor control but maybe not essential for general intelligence. They are not within the scope of this article.

2. Logical Architecture

The Cortex has both a physical structure (a layered surface, partitioned into columns) and a logical structure. The logical structure is a hierarchy - a tree-like structure that describes which columns are connected to each other (see figure 2).

Connections between columns are reciprocal: “Higher” columns receive input from “Lower” columns, and return data to the same columns. This scheme is advantageous: Higher columns have (indirect) input from a wider range of sources; lower columns use the same resources to model more specific input in greater detail. This occurs naturally because each column tries to simplify the data it outputs to higher columns, allowing columns of fixed complexity to manage greater scope in higher levels, as data is incrementally transformed and abstracted.

Only Columns in the lowest levels receive external input and control external outputs (albeit, often indirectly via subcortical structures).

Figure 2: The logical architecture of the cortex, a hierarchy of Columns. The hierarchy gives us the notion of “levels”, with the lowest level having external input and output, and higher levels being increasingly abstract. The logical architecture is superimposed on the physical architecture. Note that inter-Column connections may be gated by the central hub (not shown).

Note that there are not necessarily fewer columns in each hierarchy level; there may be, but this is not essential. However, abstraction increases and scope broadens as we move to higher hierarchy levels.

We can jump between the physical and logical architectures of the Cortex. Moving over the surface implies moving within the hierarchy. It also implies that moving between areas we will observe responses to different subsets of input data. Moving to higher hierarchy levels implies an increase in abstraction. We can observe this effect in human brains, for example by following the flow of information from the processing of raw sensor data to more abstract brain areas that deal with language and understanding (see figure 3).

Figure 3: Flow of information across the physical human Cortex also represents movement higher or lower in the logical hierarchy (increasing or decreasing abstraction). In fact, we can observe this phenomenon in human studies. Different parts of the hierarchy are specialised to conceptual roles such as understanding what, why and where things are happening. Image source. 

One final point about the logical architecture. The hierarchical structure of the Cortex is mirrored in the central hub, particularly in the Thalamus and Basal Ganglia, where we see the topology of the cortical Columns preserved through indirect pathways via central hub structures (figure 4).

Figure 4: Data flows between different Columns within the Cortex either directly, or via our conceptual “central hub”. Our hub includes Basal Ganglia such as the Striatum, and the Thalamus. Throughout this journey the topology of the Cortex is preserved. Image source.

3. Layers and Cells

Each Column has approximately 6 distinct “layers”. Like every biological rule, there are exceptions; but it suffices for the level of detail we require here. The layers are visual artefacts resulting from variations in cell type, morphology, connectivity patterns and therefore function between the layers (figure 5).

Figure 5: Various stainings showing variation in cell type and morphology between the layers of the Cortex. Image source.

The Cortex has only 5 functional layers. Structurally, it has 6 gross layers; but one layer is just wires; no computation. In addition, the functional distinction between layers 2 and 3 is uncertain, so we will group them together. This gives us just 4 unique functional layers to explain.

We will use the notation C1 ... C6 to refer to the gross layers:

• C1 - just wiring, no computation; not functional
• C2/3 (indistinct)
• C4
• C5
• C6 (known as the “multiform” layer due to the variety of cell types)

The cortex is made of a veritable menagerie of oddly shaped cells (i.e. Neurons) that are often confined to specific layers (see figure 6). Neurons have a body (soma), dendrites and axons. Dendrites provide input to the cell, and reach out to find that data. Axons transmit the output of the cell to places where it can be intercepted by other cells. Both dendrites and axons have branches.

Figure 6: Some of the cell types found in different cortical layers. Image source. 

An important feature of the Cortex is the presence of specialized Neuron cells with pyramidal Soma (bodies) (figure 7). Pyramidal cells are predominantly found in C2/3, C5 and C6. They are very large and common cells in these layers.

Pyramidal cells do not behave like classical artificial neurons. We agree with Hawkins' characterisation of them. Pyramidal cells have two or three dendrite types: Apical (proximal or distal) dendrites and Basal dendrites. Distal Apical dendrites seem to behave like a classical integrate-and-fire neuron in their own right, requiring a complete pattern of input to “fire” a signal to the body of the cell. In consequence, each cell can respond to a number of different input patterns, depending on which apical dendrites become active from their input.

Hawkins suggests that the Basal dendrites provide a sequential or temporal context in which the Pyramidal cell can become active. Output from the cell along its axon branches  only occurs if the cell observes particular instantaneous input patterns in a particular historical context of previous Pyramidal cell activity.

Within one layer of a Column, Pyramidal cells exhibit a self-organising property that results in sparse activation. Only a few Pyramidal cells respond to each input stimulus. The Pyramidal cells are powerful pattern and sequence classifiers that also perform a dimensionality-reduction function; when active, the activity of a single Pyramidal cell represents a pattern of input over a period of time.

The training mechanism for sparsity and self-organisation is local inhibition. In addition to Pyramidal cells, most of the other Neurons in the Cortex are so-called “Interneurons” that we believe play a key role in training the Pyramidal cells by implementing a competitive learning process. For example, Interneurons could inhibit Pyramidal cells around an active Pyramidal cell ensuring that the local population of Pyramidal cells responds uniquely to different input.

Unlike Pyramidal cells, which receive input from outside the Column and transmit output outside the Column, Interneurons generally only work within a Column. Since we consider Interneurons play a supporting role to Pyramidal cells, we won’t have much more to say about them.

Figure 7: A Pyramidal cell as found in the Cortex. Note the Apical and Basal dendrites, hypothesised to recognise simultaneous and historical inputs patterns respectively. The complete Pyramidal cell is then a powerful classifier that when active represents a particular set of input in a specific historical context. Image source. 

Summary

That's all we feel is necessary to say about the gross physical structure of the Thalamo-Cortical system and the microscopic structure of its cells and layers. The next article will look at the circuits and pathways by which these cells are connected, and the computational properties that result.

How to build a General Intelligence: What we think we already know

Authors: D Rawlinson and G Kowadlo

This is the first of three articles detailing our latest thinking on general intelligence: A one-size-fits-all algorithm that, like people, is able to learn how to function effectively in almost any environment. This differs from most Artificial Intelligence (AI), which is designed by people for a specific purpose. This article will set out assumptions, principles, insights and design guidelines based on what we think we already know about general intelligence. It turns out that we can describe general intelligence in some detail, although not enough detail to actually build it...yet.

The second article will look at how these ideas fit existing computational neuroscience, which helps to refine and filter the design; and the third article will describe a (high-level) algorithm that is, at least, not contradictory to the design goals and biology already established.

As usual, our plans have got ahead of implementation, so code will follow in a few weeks after the end of the series (or months...)

FIGURE 1: A hierarchy of units. Although units start out identically, they become differentiated as they learn from their unique input. The input to a unit depends on its position within the hierarchy and the state of the units connected to it. The hierarchy is conceptualized as having levels; the lowest levels are connected to sensors and motors. Higher levels are separated from sensors and motors by many intermediate units. The hierarchy may have a tree-like structure without cycles, but the number of units per level does not necessarily decrease as you move higher.

Architecture of General Intelligence

Let’s start with some fundamental assumptions and outline the structure of a system that has general intelligence characteristics.

It Exists

We assume there exists a “general intelligence algorithm” that is not irreducibly complex. That is, we don’t need to understand it in excruciating detail. Instead, we can break it down into simpler models that we can easily understand in isolation. This is not necessarily a reasonable assumption, but there is evidence for it:

Units

A general intelligence algorithm can be described more simply as a collection of many simpler, functionally-identical units. Again, this is a big assumption, but it is supported by at least two pieces of evidence. First, it has often been observed that the human cortex has quite uniform structure across areas having greatly varying functional roles. Second, this structure has revealed that the cortex is made up of many smaller units (called columns, at one particular scale). It is reasonable to decompose the cortex in this way due to high and varied intra-column connectivity and limited variety of inter-column connectivity. The patterns of inter and intra column connectivity are very similar throughout the cortex. “Columns” contain only a few thousand neurons organized into layers and micro-columns that further simplify understanding of the structure. That’s not overwhelmingly complex, although we are making simplifying assumptions about neuron function.

Hierarchy

Our reading and experimentation has suggested that hierarchical representation is critical for the types of information processing involved in general intelligence. Hierarchies are built from many units connected together in layers. Typically, only the lowest level of the hierarchy receives external input. Other levels receive input from lower levels of the hierarchy instead. For more background on hierarchies, see earlier posts. Hierarchy allows units in higher layers to model more complex and abstract features of input, despite the fixed complexity of each unit. Hierarchy also allows units to cover all available input data and allow combinations of features to be jointly represented within a reasonable memory limit. It’s a crucial concept.

Synchronization

Do we need synchronization between units? Synchronization can simplify sequence modelling in a hierarchy by restricting the number of possible permutations of events. However, synchronization between units may significantly hinder fast execution on parallel computing hardware, so this question is important. A point of confusion may be the difference between synchronization and timing / clock signals. We can have synchronization without clocks, but in any case there is biological evidence of timing signals within the brain. Pathological conditions can arise without a sense of time. In conclusion we’re going to assume that units should be functionally asynchronous, but might make use of clock signals.

Robustness

Your brain doesn’t completely stop working if you damage it. Robustness is a characteristic of a distributed system and one we should hope to emulate. Robustness applies not just to internal damage but external changes (i.e. it doesn't matter if your brain is wrong or the world has changed; either way you have to learn to cope).

Scalability

Adding more units should improve capability and performance. The algorithm must scale effectively without changes other than having more of the same units appended to the hierarchy. Note the specific criteria for how scalability is to be achieved (i.e. enlarge the hierarchy rather than enlarge the units). It is important to test for this feature to demonstrate the generality of the solution.

Generality

The same unit should work reasonably well for all types of input data, without preprocessing. Of course, tailored preprocessing could make it better, but it shouldn’t be essential.

Local interpretation

The unit must locally interpret all input. In real brains it isn’t plausible that neuron X evolved to target neuron Y precisely. Neurons develop dendrites and synapses with sources and targets that are carefully guided, but not to the extent of identifying specific cells amongst thousands of peers. Any algorithm that requires exact targeting or mapping of long-range connections is biologically implausible. Rather, units should locally select and interpret incoming signals using characteristics of the input. Since many AI methods require exact mapping between algorithm stages, this principle is actually quite discriminating.

Cellular plausibility

Similarly, we can validate designs by questioning whether they could develop by biologically plausible processes, such as cell migration or preferential affinity for specific signal coding or molecular markers. However, be aware that brain neurons rarely match the traditional integrate-and-fire model.

Key Insights

It’s surprising that in careers cumulatively spanning more than 25 years we (the authors) had very little idea how the methods we used everyday could lead to general intelligence. It is only in the last 5 years that we have begun to research the particular sub-disciplines of AI that may lead us in that direction.

Today, those who have studied this area can talk in some detail about the nature of general intelligence without getting into specifics. Although we don’t yet have all the answers, the problem has become more approachable. For example, we’re really looking to understand a much simpler unit, not an entire brain holistically. Many complex systems can be easily understood when broken down in the right way, because we can selectively ignore detail that is irrelevant to the question at hand.

From our experience, we've developed some insights we want to share. Many of these insights were already known, and we just needed to find the right terminology. By sharing this terminology we can help others to find the right research to read.

We’re looking for a stackable building block, not the perfect monolith

We must find a unit that can be assembled into an arbitrarily large - yet still functional - structure. In fact, a similar feature was instrumental in the success of “deep” learning: Networks could suddenly be built up to arbitrary depths. Building a stackable block is surprisingly hard and astonishingly important.

We’re not looking to beat any specific benchmark

... but if we could do reasonably well at a wide range of benchmarks, that would be exciting. This is why the DeepMind Atari demos are so exciting; the same algorithm could succeed in very different problems.

Abstraction by accumulation of invariances

This insight comes from Hawkins’ work on Hierarchical Temporal Memory. He proposes that abstraction towards symbolic representation comes about incrementally, rather than as a single mapping process. Concepts accumulate invariances - such as appearance from different angles - until labels can correctly be associated with them.  This neatly avoids the fearful “symbol grounding problem” from the early days of AI.

Biased Prediction and Selective Attention are both action selection

We believe that selective bias of predictions and expectations is responsible for both narrowing of the range of anticipated futures (selective ignorance of potential outcomes) and the mechanism by which motor actions are generated. A selective prediction of oneself performing an action is a great way to generate or “select” that action. Similarly, selective attention to external events affects the way data is perceived and in turn the way the agent will respond. Filtering data flow between hierarchy units implements both selective attention and action selection, if data flowing towards motors represents candidate futures including both self-actions and external consequences.

The importance of spatial structure in data

As you will see in later parts of this article series, the spatial structure of input data is actually quite important when training our latest algorithms. This is not true of many algorithms, especially in Machine Learning where each input scalar is often treated as an independent dimension. Note that we now believe spatial structure is important both in raw input and in data communicated between units. We’re not simply saying that external data structure is important to the algorithm - we’re claiming that simulated spatial structure is actually an essential part of algorithms for dynamically dividing a pool of resources between hierarchy units.

Binary data

There's a lot of simplification and assumption here, but we believe this is the most useful format for input and internal data. In any case, the algorithms we’re finding most useful can't easily be refactored for the obvious alternative (continuous input values). However, continuous input can be encoded with some loss of precision as subsets of bits. There is some evidence that this is biologically plausible, but it is not definitive. Why binary? Dimensionality reduction is an essential feature of a hierarchical model; it may be that sparse binary representations are simply a good compromise between data loss and qualities such as compositionality:

Sparse, Distributed Representations

We will be using Sparse, Distributed Representations (SDRs) to represent agent and world state [RE  ]. SDRs are binary data (i.e. all values are 1 or 0). SDRs are sparse, meaning that at any moment, only a fraction of the bits are 1's (active). The most complex feature to grasp is that SDRs are distributed: No individual bit uniquely represents anything. Instead, data features are jointly represented by sets of bits. SDRs are overcomplete representations - not all bits in a feature-set are required to “detect” a feature, which also means that degrees of similarity can also be expressed as if the data were continuous. These characteristics also mean that SDRs are robust to noise - missing bits are unlikely to affect interpretation. .

Predictive Coding

SDRs are a specific form of Sparse (Population) Coding where state is jointly represented by a set of active bits. Transforming data into a sparse representation is necessarily lossy and balances representational capacity against bit-density. The most promising sparse coding scheme we have identified is Predictive Coding, in which internal state is represented by prediction errors. PC has the benefit that errors are propagated rather than hidden in local states, and data dimensionality automatically reduces in proportion to its predictability. Perfect prediction implies that data is fully understood, and produces no output. A specific description of PC is given by Friston et al but a more general framework has been discussed in several papers by Rao, Ballard et al since about 1999. The latter is quite similar to the inter-region coding via temporal pooling described in the HTM Cortical Learning Algorithm.

Generative Models

Training an SDR typically produces a Generative Model of its input. This means that the system encodes observed data in such a way that it can generate novel instances of observed data. In other words, the system can generate predictions of all inputs (with varying uncertainty) from an arbitrary internal state. This is a key prerequisite for a general intelligence that must simulate outcomes for planned novel action combinations.

Dimensionality Reduction

In constructing models, we will be looking to extract stable features and in doing so reduce the complexity of input data. This is known as dimensionality reduction, for which we can use algorithms such as auto-encoders. To cope with the vast number of possible permutations and combinations of input, an incredibly efficient incremental process of compression is required. So how can we detect stable features within data?

Unsupervised Learning

By the definition of general intelligence, we can’t possibly hope to provide a tutor-algorithm that provides the optimum model update for every input presented. It’s also worth noting that internal representations of the world and agent should be formed without consideration of the utility of the representations - in other words, internal models should be formed for completeness, generality and accuracy rather than task-fulfilment. This allows less abstract representations to become part of more abstract, long-term plans, despite lacking immediate value. It requires that we use unsupervised learning to build internal representations.

Hierarchical Planning & Execution

We don’t want to have to model the world twice: Once for understanding what’s happening, and again for planning & control. The same model should be used for both. This means we have to do planning & action selection within the single hierarchical model used for perception. It also makes sense, given that the agent’s own actions will help to explain sensor input (for example, turning your head will alter the images received in a predictable way). As explained earlier, we can generate plans by simply biasing “predictions” of our own behaviour towards actions with rewarding outcomes.

Reinforcement Learning

In the context of an intelligent agent, it is generally impossible to discover the “correct” set of actions or output for any given situation. There are many alternatives of varying quality; we don’t even insist on the best action but expect the agent to usually pick rewarding actions. In these scenarios, we will require a Reinforcement Learning system to model the quality of the actions considered by the agent. Since there is value in exploration, we may also expect the agent to occasionally pick suboptimal strategies, to learn new information.

Supervised Learning

There is still a role for supervised learning within general intelligence. Specifically, during the execution of hierarchical control tasks we can describe both the ideal outcome and some metric describing similarity of actual outcome to desired. Supervised learning is ideal for discovery of actions with agency to bring about desired results. Supervised Learning can tell us how best to execute a plan constructed in an Unsupervised Learning model, that was later selected by Reinforcement Learning.

Challenges Anticipated 

The features and constraints already identified mean that we can expect some specific difficulties when creating our general intelligence.

Among other problems, we are particularly concerned about:

1. Allocation of limited resources
2: Signal dilution
3: Detached circuits within the hierarchy
4: Dilution of executive influence
5: Conflict resolution
6: Parameter selection

Let’s elaborate:

Allocation of limited resources

This is an inherent problem when allocating a fixed pool of computational resources (such as memory) to a hierarchy of units. Often, resources per unit are fixed, ensuring that there are sufficient resources for the desired hierarchy structure. However, this is far less efficient than dynamically allocating resources to units to globally maximize performance. It also presupposes the ideal hierarchy structure is known, and not a function of the data. If the hierarchy structure is also dynamic, this becomes particularly difficult to manage because resources are being allocated at two scales simultaneously (resources → units and units → hierarchy structure), with constraints at both scales.

In our research we will initially adopt a fixed resource quota per hierarchy unit and a fixed branching factor for the hierarchy, allowing the structure of the hierarchy and resources per unit to be determined by data. This arrangement is the one most likely to work given a universal unit with constant parameters, as the number of inputs to each unit is constrained (due to the branching factor). It is interesting that the human cortex is a continuous sheet, and evidences dynamic resource allocation as neuroplasticity - resources can be dynamically assigned to working areas and sensors when others fail.

Signal Dilution

As data is transformed from raw input into a hierarchical model, information will be lost (not represented anywhere). This problem is certain to occur in all realistic tasks because input data will be modelled locally in each unit without global oversight over which data is useful. Given local resource constraints, this will be a lossy process. Moreover, we have also identified the need for units to identify patterns in the data and output a simplified signal for higher-order modelling by other units in the hierarchy (dimensionality reduction). Therefore, each unit will deliberately and necessarily lose data during these transformations. We will use techniques such as Predictive Coding to allow data that is not understood (i.e. not predictable) to flow through the system until it can be modelled accurately (predicted). However, it will still be important to characterise the failure modes in which important data is eliminated before it can be combined with other data that provides explanatory power.

Detached circuits within the hierarchy

Consider figure 2. Here we have a tree of hierarchy units. If the interactions between units are reciprocal (i.e. X outputs to Y and receives data from Y) there is a strong danger of small self-reinforcing circuits forming in the hierarchy. These feedback circuits exchange mutually complementary data between a pair or more units, causing them to ignore data from the rest of the hierarchy. In effect, the circuit becomes “detached” from the rest of the hierarchy. Since sensor data enters via leaf-units at the bottom of the hierarchy, everything above the detached circuit is also detached from the outside world and the system will cease to function satisfactorily.

In any hierarchy with reciprocal connections, this problem is very likely to occur, and disastrous when it does. In Belief Propagation, another graphical model, this problem manifests as “double counting” and is avoided by nodes carefully ignoring their own evidence returned to them.

FIGURE 2: Detached circuits within the hierarchy. Units X and Y have formed a mutually reinforcing circuit that ignores all data from other parts of the hierarchy. By doing so, they have ceased to model the external world and have divided the hierarchy into separate components.

Dilution of executive influence

A generally-intelligent agent needs to have the ability to execute abstract, high-level plans as easily as primitive, immediate actions. As people we often conceive plans that may take minutes, hours, days or even longer to complete. How is execution of lengthy plans achieved in a hierarchical system?

If abstract concepts exist only in higher levels of the hierarchy, they need to control large subtrees of the hierarchy over long periods of time to be successfully executed. However, if each hierarchy unit is independent; how is this control to be achieved? If higher units do not effectively subsume lower ones, executive influence will dilute as plans are incrementally re-interpreted from abstract to concrete (see figure 3). Ideally, abstract units will have quite specific control over concrete units. However, it is impractical for abstract units to have the complexity to "micro-manage" an entire tree of concrete units.

FIGURE 3: Dilution of executive influence. A high-level unit within the hierarchy wishes to execute a plan; the plan must be translated towards the most concrete units to be performed. However, each translation and re-interpretation risks losing details of the original intent which cannot be fully represented in the lower levels. Somehow, executive influence must be maintained down through an arbitrarily deep hierarchy. 

Let’s define “agency” as the ability to influence or control outcomes. Lacking the ability to cause a particular outcome is a lack of agency over the desired and actual outcomes. By making each hierarchy unit responsible for the execution of goals defined in the hierarchy level immediately above, we indirectly maximise the agency of more abstract units. Without this arrangement, more abstract units would have little or no agency at all.

Figure 4 shows what happens when an abstract plan gets “lost in translation” to concrete form. I walked up to my car and pulled my keys from my pocket. The car key is on a ring with many others, but it’s much bigger and can’t be mistaken by touch. It can only be mistaken if you don’t care about the differences.

In this case, when I got to the car door I tried to unlock it with the house key! I only stopped when the key wouldn't fit in the keyhole. Strangely, all low-level mechanical actions were performed skillfully, but high level knowledge (which key) was lost. Although the plan was put in motion, it was not successful in achieving the goal.

Obviously this is just a hypothesis about why this type of error happens. What’s surprising is that it isn't more common. Can you think of any examples?

FIGURE 4: Abstract plan translation failure: Picking the wrong key but skilfully trying it in the lock. This may be an example of abstract plans being carried out, but losing relevant details while being transformed into concrete motor actions by a hierarchy of units.

In our model, planning and action selection occur as biased prediction. There is an inherent conflict between accurate prediction and bias. Attempting to bias predictions of events beyond your control leads to unexpected failure, which is even worse than expected failure.

The alternative is to predict accurately, but often the better outcome is the less likely one. There must be a mechanism to increase the probability of low-frequency events where the agent has agency over the real-world outcome.

Where possible, lower units must separate learning to predict and trying to use that learning to satisfy higher units’ objectives. Units should seek to maximise the probability of goal outcomes, given an accurate estimate of the state of the local unit as prior knowledge. But units should not become blind to objective reality in the process.

Conflict resolution

General intelligence must be able to function effectively in novel situations. Modelling and prediction must work in the first instance, without time for re-learning. This means that existing knowledge must be combined effectively to extrapolate to a novel situation.

We also want the general intelligence to spontaneously create novel combinations of behaviour as a way to innovate and discover new ways to do things. Since we assume that behaviour is generated by filtering predictions, we are really saying we need to be able to predict (simulate) accurately when extrapolating combinations of existing models to new situations. So we also need conflict resolution for non-physical or non-action predictions. The agent needs a clear and decisive vision of the future, even when simulating outcomes it has never experienced.

The downside of all this creativity is that there’s really no way to tell whether these combinations are valid. Often they will be, but not always. For example, you can’t touch two objects that are far apart at the same time. When incompatible, we need a way to resolve the conflict.

There’s a good discussion of different conflict resolution strategies on Scholarpedia; our preferred technique is selecting a solitary active strategy in each hierarchy unit, choosing locally to optimise for a single objective when multiple are requested.

Evaluating alternative plans is most easily accomplished as a centralised task - you have to bring all the potential alternatives together where they can be compared. This is because we can only assign relative rewards to each alternative; it is impossible to calculate meaningful absolute rewards for the experiences of an intelligent agent. It is also important to place all plans on a level playing-field regardless of the level of abstraction; therefore abstract plans should be competing against more concrete ones and vice-versa.

Therefore, unlike most of the pieces we've described, action selection should be a centralised activity rather than a distributed one.

Parameter Selection

In a hierarchical system the input to “higher” units will be determined by modelling in “lower” units and interactions with the world. The agent-world system will develop in markedly different ways each time. It will take an unknown amount of time for stable modelling to emerge, first in the lower units and then moving higher in the hierarchy.

As a result of all these factors it will be very difficult to pick suitable values for time-constants and other parameters that control the learning processes in each unit, due to compounded uncertainty about lower units’ input. Instead, we must allow recent input to each unit to determine suitable values for parameters. This is online learning. Some parameters cannot be automatically adjusted in response to data. For these, to have any hope of debugging a general intelligence, a fixed parameter configuration must work for all units in all circumstances. This constraint will limit the use of some existing algorithms.

Summary

That wraps up our theoretical overview of what we think a general intelligence algorithm must look like. The next article in this series will explain what we've learnt from biology’s implementation of general intelligence - ourselves! The final article will describe how we hope to build an algorithm that satisfies all these requirements.