How to build a General Intelligence: Circuits and Pathways

Figure 1: Our headline image is from the Cognitive Consilience: An atlas of key pathways cross-referenced to supporting literature articles. The complexity and variety of routing within the brain can be appreciated with this beautiful illustration. Note in particular the specialisation of cortical cells and the way this affects their interactions with other cells in the cortex and elsewhere in the brain. Explore this fantastic resource yourself.

By David Rawlinson and Gideon Kowadlo

This is part 3 of our series “how to build an artificial general intelligence” (AGI). Part 1 was a theoretical look at General Intelligence (follow the link if you don’t know what General Intelligence is).

We believe that the Thalamo-Cortical system is the origin of General Intelligence in people. In Part 2 we presented very broadly how the Thalamo-Cortical system is structured and organised. We applied some core concepts, such as hierarchy, to help us describe the system.

We also looked at the cellular structure of the Cortex and in particular introduced Pyramidal cells.

This article is again about what we can learn from reverse-engineering the Thalamo-Cortical system, but this time from its connectivity, which we present in terms of circuits and pathways.

Pathways and Circuits

A pathway is a gross pattern of sequential connectivity between brain regions - for example, if part A is highly connected to part B, and activity in A is followed by activity in B, we say there exists a pathway between A and B. Cells in the Thalamo-Cortical system are connected to each other in quite restricted and specific ways, so these pathways are quite informative.

Circuits are more specific and precise details of both connectivity and functional interaction between neurons. In computational neuroscience there exists a concept called the Canonical Cortical Micro-Circuit. The specifics of this circuit are not widely agreed, because (a) the Cortex is complex and (b) many of the evidence-gathering exercises are statistical observations (e.g. “X% of outputs from A and Y% of output from B projects to region C”) which may obscure fundamental functional or topological features. For example, outputs from A and B may project to cells with exclusive roles, but physically co-located in C. Statistical, regional approaches will not capture such distinctions.

In the neuroscience literature there’s a frustrating habit of selectively reporting supporting details while ignoring others. Perhaps this is simply because it's impossible to describe any part exhaustively. In particular, there is a lot of contradictory information about Cortical circuits. But the research can still shed some light on what is happening. Just don’t expect all sources to be consistent or complete!

Key Cortical Pathways

There are several widely-cited and well established cortical pathways (i.e. routes with at least one end in the Cortex). To understand these, it is important to remember both the physical and logical structure of the Cortex as described in the previous article. Physically, the cortex is made of layers, and logically, Columns within the Cortex form a hierarchy.

The hierarchy defines a structure made of Columns, and determines which Columns interact. Pathways describe the patterns of interaction between cells within a Column, and between Columns. We assume that all Columns are functionally identical prior to training.

Cells within Columns are usually identified by both the physical location of cell bodies within particular Layers in the Column, and by the morphology (shape) of the cell. Data flow to and from Cortical cells is largely restricted to a handful of core pathways that begin and terminate in particular cell types in specific cortical layers.

There are many descriptions of cortical pathways and circuits in the literature. We will first introduce just 4 well-established cortical-cortical pathways, and then some thalamo-cortical pathways. Note that although the existence of these pathways is unambiguous, their purpose and function is poorly understood. They appear to be consistent across various somatosensory regions of the cortex, especially in comparison to variations in other brain tissues.

Hawkins’ Hierarchical Temporal Memory (HTM) introduces 3 of the 4 pathways in a single, coherent scheme and relates them to a general intelligence algorithm. We will borrow this terminology and describe them in detail below. Their names describe the direction of data flow and the routing used:

Feed-Forward Direct Pathway: C2/3 → C4 → C2/3
Feed-Forward Indirect Pathway: C5 → Thalamus → C4 → C2/3
Feed-Back Direct Pathway #1: from C6 → C1

We are also interested in a second Feed-Back Direct “pathway” implemented by cortically projecting C6 pyramidal cells with axons that terminate in both C6 and C1 in hierarchically lower regions (see here for a diagram).

Feed-Back Direct Pathway #2: from C6 → C6

Note that cells in all cortical layers (except, perhaps, C4) receive input via their dendrites in C1. In other words, feedback from C6 to C1 is then used as input to many layers. Feedback from C6 to C6 is generally not input for other layers.

In neuroscience, Feed-Forward usually means the flow of data away from external sources such as sensors (towards greater abstraction, if you believe in a cortical hierarchy). Feed-Back means the opposite - data flow towards regions that have direct interaction with external sensors and motors.

Direct pathways are so-called because data is routed directly from one cortical column or region to another, without a stop along the way. Indirect pathways are routed via other structures. The  “Feed-Forward-Indirect” pathway described by Hawkins is routed via the Thalamus.

Figure 2, derived from a Hawkins/Numenta publication, shows graphically how information flows between columns and between layers within columns, as part of these 3 pathways according to the HTM theory. As mentioned before, the community is welcome to contribute by updating and adding to the figure.

Hawkins assigns specific roles to these pathways, but we will be re-interpreting them in the next article.

Figure 2: Routing of 3 core pathways, based on a diagram from the HTM/CLA White Paper. Note the involvement of specific cortical layers with each pathway, and the central role of the Thalamus. The names of the pathways indicate direct (cortex-to-cortex) and indirect (cortex-thalamus-cortex) variants, with direction being either forward (away from external sensors and motors, towards increasing abstraction) or backward (towards more concrete regions dealing with specific sensor/motor input). 

The role of the Thalamus

Let’s recap: The Cortex is composed of Columns, organised into a hierarchy. Cells pass messages directly to other Columns that are higher or lower in the hierarchy. Messages may also be transmitted indirectly between Columns, via the Thalamus.

The Thalamus is often viewed as having a gating or relaying function. The Thalamus is particularly associated with control of attention.

This section will describe indirect pathways involving the Thalamus. Figure 3 is a reproduction of a figure from Sherman and Guillery (2006) that has two new features of interest. These authors use the terminology “first order” to denote cortical regions receiving direct sensor input and “higher order” to denote cortical regions receiving input from “first order” cortical regions. This corresponds with the notion of hierarchy levels 1 and 2.

The Thalamus is a significant part of the “Feed-Forward Indirect” pathway. This pathway originates at Cortex layer 5 and propagates to a nucleus in the Thalamus. There, the nucleus may react by transmitting a (presumably corresponding) signal to one or more other Cortical Columns, in a different region. In some theories of cortical function, the target Column is conceptually “higher” in the hierarchy. The Thalamic input enters the Cortex via Thalamic axons terminating in Cortex layer 4 and is then propagated to Cortex Layer 5 where the pathway begins again.

Figure 3 also shows Cells in Columns in Cortex layer 6 fairly accurately form reciprocal modulatory connections to Thalamic nuclei that provide input to the Column via C4 and C5! Therefore, a Column within the Cortex has influence on data that it receives from the Thalamus. In effect, the Cortex is not a passive recipient but works with the Thalamus to control its input. The figure also depicts C6 cells projecting to C6 in lower regions (our second feedback pathway).

Figure 3: Pathways between cortical columns in different regions, showing layer involvement in each pathway and the role of the Thalamus. Sherman and Guillery use the terminology “first order” to denote cortical regions receiving direct sensor input and “higher order” to denote cortical regions receiving input from lower (e.g. “first order”) cortical regions. This corresponds with the notion of hierarchy levels 1 and 2. Note that in addition to the 3 pathways shown in the previous figure, we see additional direct feedback pathways and reciprocal feedback from Cortex layer 6 to the Thalamic nuclei that stimulate the cortical region. Image source.

Motor output

At this point it is interesting to look at how the Cortex can influence or control behaviour, particularly the generation of motor output. There are two pathways that allow the cortex to influence or control behaviour:

Cortical Control: Basal Ganglia → Thalamus → Cortex → Motors
Cortical Influence: Cortex → Basal Ganglia → Motors

Note that in both cases, the origin of action selection is the Basal Ganglia. In the first case, the Basal Ganglia control signals emitted by the Thalamus, with these signals in turn affecting activity within Cortex layer 5 (C5). C5, particularly in motor areas, has been studied in detail. 10-15% of the cells in these areas are very large pyramidal neurons known as Betz cells, that can be observed to drive muscles very directly with few synapses in between. These cells are more prevalent in primates and are especially important for control of the hands. This makes sense given that manual tasks are typically more complex and require greater dexterity than movements by other parts of the body. The human Cortex is believed to be crucial for innovative and sophisticated manual tasks such as tool-making.

Within the Cortical layers, C5 seems to be uniquely involved in motor output. Figure 4 shows some of the ways output from Pyramidal cells in C5 project output to areas of the brain associated with motor output and control. In contrast, pyramidal cells in C2/3 predominantly project to other areas of the cortex and are not directly involved in control.

Figure 4: Pyramidal cells in C5 project output to areas of the brain associated with motor output and control. In contrast, pyramidal cells in C2/3 predominantly project to other areas of the cortex and are not directly involved in control. Image source.

The second way that the Cortex can influence motor output is via the Basal Ganglia. In this case, we propose that the Cortex might provide contextual information to assist the Basal Ganglia in its direct control outputs, but we found no evidence that the Cortex is able to exert control over the Basal Ganglia.

We suggest Cortical influence over the Basal Ganglia is less interesting from a General Intelligence perspective, because the hierarchical representations formed within the Cortex are not exploited, and execution is performed by more ancient brain systems not associated with General Intelligence qualities.

For the rest of this article series, we will ignore control pathways that do not involve the Cortex,  and will focus on direct control output from Cortex layer 5.

Action Selection

It is widely believed that action selection occurs within the flow of information from Cortex through the Basal Ganglia, a group of deep, centralised brain structures adjacent to the Thalamus. There are a number of theories about how this occurs, but it is generally believed to involve a form of Reinforcement Learning used to select ideas from the options presented by the Cortex, with competitive mechanisms for clean switching and conflict resolution.

A major output of the Basal Ganglia is to the Thalamus; one prevailing theory of this relationship is that the Basal Ganglia controls the gating or filtering function performed by the Thalamus, effectively manipulating the state of the Cortex in consequence. The full loop then becomes Cortex → Basal Ganglia → Thalamus → Cortex (see Wikipedia for a good illustration, or figure 5).

As discussed above, this article will focus on motor output generated directly by the Cortex.

Figure 5: Pathways forming a circuit from Cortex to Basal Ganglia to Thalamus and back to Cortex. Image Source.

Canonical Cortical Circuit

We now have the all the background information needed to define a “Canonical Cortical micro-Circuit” at a cellular level. All the information presented so far has been relatively uncontroversial, but this circuit is definitely our interpretation, not an established fact. However, we will present some evidence to (inconclusively) support our interpretation.

Figure 6: Our interpretation of the canonical cortical micro-circuit. Only a single cortical region or Column is shown. Arrow endings indicate the type of connection - driver, modulator or inhibitor. The numbers 2/3, 4, 5, and 6 refer to specific cortical layers. Each shape represents a set of cells of a particular type, not an individual cell. Self-connections and connections within each set are not shown, but often exist. Shapes T and B refer to Thalamus and Basal Ganglia, not broken down into specific cell layers or types. Data enters the diagram at 4 points, labelled A-D, but does not exit; in general the system forms a circuit not a linear path. Note that shape T occurs twice, because the circuit receives data from only one part of the Thalamus but projects to two areas in forward and backward directions.

Diagram Explanation

We will use variants of the diagram shown in figure 6 to explain our interpretation of cortical function. In this diagram, only a single Cortical region or Column (used interchangeably here) is shown. In later diagrams, we will show 3 hierarchy levels together so the flow of information between hierarchy levels is apparent.

In these diagrams, shapes represent a class of Neurons within a specific Cortical Layer. The numbers 2/3, 4, 5 and 6 refer to the Cortical layers in which these cell classes occur. The shapes labelled T and B refer to the Thalamus and Basal Ganglia (internal cell types and layers are not shown). Arrows on the diagram show the effect of each connection, either driving (providing information or input that causes another cell to become active), modulation (stimulating or inhibiting the activity of a target cell) or inhibition (exclusively inhibiting the activity of a target cell).

If you want more detail on the thalamic end of the thalamocortical circuitry, an excellent source is this paper by Sherman.

There are many interneurons (described in the previous article) that are not shown in this diagram. We chose to omit these because we believe they are integral to the function of a layer of pyramidal cells within a Column, rather than an independent system. Specifically, we suggest that inhibitory interneurons implement local self-organising and local competitive functions (e.g. winner-take-all), ensuring sparse activation of the cell types represented by shapes in our diagram (C2/3, C4, C5, and C6). The self-organising behaviour also ensures that cells within each column optimise coverage of observed input patterns given a finite cell population. Inclusion of the interneurons would clutter the diagram without adding much explanatory value.

We also omit self-connections within a class of cells represented by a shape. These self-connections likely provide context and contribute to learning and exclusive activity within the class, but don’t make it easier to understand circuits in terms of cortical layers and hierarchy levels.

Excitatory Circuit

Figure 7 shows a multilevel version of the cortical circuit, similar to the multi-level figure from Sherman and Guillery (figure 3). We can now understand where the inputs to the circuit come from, in terms of other layers and external Sensors (S) and Motors (M). Note that Motors are driven directly from C5.

Figure 7: The cortical micro-circuit across several levels of Cortex with involvement of Thalamus and Basal Ganglia. The red highlight shows a single excitatory ‘circuit’. See text for details.

The red path in figure 7 shows our excitatory “canonical circuit”: Data flows from the Thalamus to spiny stellate (star-shape in figures) cells in C4 (see source and source), from where it propagates to pyramidal cells in C2/3, and then to pyramidal cells in C5. C6 is known as the multiform layer, but also contains many pyramidal cells of unusual proportions and orientations. C6 cells are driven by C5, and in turn modulate the Thalamus. Note that C6 cells within a region modulate the same Thalamic nuclei that provide input to that region of Cortex.

Inhibitory Circuit

A second, inhibitory circuit exists alongside our excitatory circuit. In addition to providing input to the Cortex via C4, axons from the Thalamus also drive inhibitory Parvalbumin-expressing (PV) neurons in C4 (shown as circles in the diagram). These inhibitory neurons make up a large fraction of all the cells in C4, and inhibit pyramidal cells in C5 (see source or source ).

This means that the input from the Thalamus can be both informative and executive. It is executive in that it actually manipulates the activity of layer 5 within the Cortex, and informative by providing a copy of the signal driving the manipulation to C4. Figure 8 shows our inhibitory circuit. We believe this circuit is of critical importance because it provides a mechanism for the Thalamus to centrally manipulate the state of the Cortex, specifically layer 5 and 6 pyramidal cells. This hypothesis will be expanded in the next article.

Figure 9 catalogues inhibitory cells, notably showing the cells used in our inhibitory circuit.

Figure 8: The inhibitory micro-circuit. The red highlight shows how the Thalamus controls activity in C5 within a Column by activating inhibitory cells in C4. The circuit is completed by C5 pyramidal cells driving C6 cells, which in turn modulate the activity of the same Thalamic nuclei that selectively activates C5. Each shape denotes a population of cells of a specific type within a single Column, excluding ‘T’ and ‘B’ that refer to the Thalamus and Basal Ganglia respectively.

Figure 9: Inhibitory interneurons in the Cortex. Of particular interest are the “PV” cells that are driven by axons from the Thalamus terminating in layer 4 and in turn inhibit pyramidal cells in layer 5. Image source. 

Pathways and the Canonical Circuits

Now let’s look at how pathways emerge from our cortical micro-circuit. Figures 10, 11, 12 show the Feed-Forward Direct, Feed-Forward Indirect and first Feed-Back pathways respectively. We also include another direct, Feed-Back pathway terminating at C6 (figure 13). Feed-back direct pathways terminating at C1, where many fibres are intermingled, are harder to interpret than feedback terminating directly at C6. Pyramidal neurons from many layers have dendrites in C1.

Figure 10: Feed-Forward Direct pathway within our canonical cortical micro-circuit.

Figure 10 highlights the Feed-Forward direct pathway. Signals propagate 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. Although activity from C2/3 propagates to C5, it does not ascend the hierarchy via this route: C5 in one Column does not directly connect to C5 in a higher Column, only via an indirect pathway (see below).

Figure 11: Feed-Forward Indirect pathway.

Figure 11 highlights the Feed-Forward Indirect pathway. The Thalamus is involved in this pathway, and may have a gating or filtering effect. Data flows from the Thalamus to C4, to C2/3, to C5 and then to a different Thalamic nucleus that serves as the input gateway to another cortical Column in a different region of the Cortex.

Figure 12: The first of two Feed-Back Direct pathways.

Figure 12 highlights the first type of Feed-Back Direct pathway. This pathway may be more concerned with provision of broader and more abstract (i.e. hierarchically higher) contextual information to be used in the Feed-Forward pathways for better prediction. This suggestion is supported by evidence that axons from C6 via C1 synapse with apical dendrites of pyramidal cells in C2/3, C5 and C6, in hierarchically lower regions.

Figure 13 highlights the second of two Feed-Back Direct pathways. This pathway might be involved in cascading control activity down the hierarchy towards sensors and motors - the next article will expand on this idea. Activity propagates from C6 to C6 directly. C6 modulates the activity of local C5 cells and relevant Thalamic nuclei that drive local C5 cells. Note that connections from a Column to the Thalamus are reciprocal; feedback from C6 to the Thalamus targets the same nuclei that project axons to C4.

Figure 13: The second of two Feed-Back Direct pathways.

Summary

We’ve presented some additional, detailed perspectives on the organisation and function of circuits and pathways within the Thalamo-Cortical system and presented our interpretation of the canonical cortical micro-circuit.

So what’s the point of all this information? What do these circuits and pathways do, and why are they connected this way? How do they work?

It might seem that we’ve stopped short of really trying to interpret all this information and that’s because we are, indeed, holding back. After having spent so much time presenting background information, the next article finally attempts to understand why the thalamocortical system is connected in the ways described here, and how this system might give rise to general intelligence. 

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.

AGI Experimental Framework: A platform for AGI R&D

By Gideon Kowadlo and David Rawlinson

Introduction

We’ve been building and testing AGI algorithms for the last few years. As the systems become more complex, we have found it ever more difficult to run meaningful experiments. To summarise, the main challenges are:

• testing a version of the algorithm repeatedly and over some range of parameters or conditions,
• scaling it up so that it can run quickly,
• debugging: the complexity of the ‘brain’ makes visualising and interpreting its state almost as hard as the problem itself! 

Platforms for testing AIs already exist, such as the Arcade Learning Environment. There are also a number of standard datasets and frameworks for testing them. What we want is a framework for understanding the behaviour of an AI that can be applied successfully to any problem - it is supposed to be an Artificial General Intelligence, after all. The goal isn't to advance the gold-standard incrementally; instead we want to better understand the behaviour of algorithms that might work reasonably well on many different problems.

Whereas most AI testing frameworks are designed to facilitate a particular problem, we want to facilitate understanding of the algorithms used. Further, the algorithms will have complex internal state and be variably parameterised from small instances on trivial problems to large instances - comprising many computers - on complex problems. As such there will be a lot of emphasis on interfaces that allow the state of the algorithm to be explored.

These design goals mean that we need to look more at the enterprise and web-scale frameworks for distributed systems, than test harnesses for AIs. There's a huge variety of tools out there: Distributed filesystems, cloud resourcing (such as Elastic Compute), and cluster job management (e.g. many scientific packages available in Python). We'll design a framework with the capability to jump between platforms as available technologies evolve.

Developing distributed applications is significantly harder than single-process software. Synchronization and coordination is harder (c.f. Apache Zookeeper), and there's a lot of crud to get right before you can actually get to the interesting bits (i.e. the AGI). We're going to try to get the boring stuff done nicely, so that others can focus on the interesting bits!

Foundational Principles

• Agent/World conceptualisation
• For AGI, we have developed a system based around Experiments, with each Experiment having Agents situated in a World.
• Reproducible
• All data is persisted by default so that any experiment can be reproduced from any time step.
• Easy to run and use
• Minimal setup and dependencies.
• No knowledge of the implementation is required to implement a custom module (primarily the intelligent Agent or World in which it operates).
• Highly modular (Scalability)
• Different parts of the system can be customised, extended or overridden independently.
• Distributed architecture (Scalability)
• Modules can be run on physically separated machines, without any modification to the interactions between modules (i.e. the programmer’s perspective is not affected by scaling of the system to multiple computers).
• Easy to develop
• Code is open source.
• Code is well documented.
• All API’s well documented and using standard protocols (at the moment RESTful, in future could be websockets or other).
• Explorable / Visualisable
• High priority placed on debugging and understanding of data rather than simply efficiency and throughput. We don’t yet know what the algorithm should look like!
• All state is accessible, relations are can be explored.
• Execution is on demand (step-by-step) or automatic (until criteria, or batches of experiments completed).
• It must be easy for anyone to build a UI client that can explore the state of all parts of the system.

Conceptual Entities

We have defined a number of components that make up an experiment. We refer to these components as Entities, and give them a specific interface.

• World
• The simulated environment within which all the other simulated components exist.
• Agent 
• The intelligent agent itself. It operates within a World, and interacts with that World and (optionally) other Agents via a set of Sensors and Actuators.
• Sensor
• A means by which the Agent senses the world. The output is a function of a subset of the World state. For example, a unidirectional light sensor may provide the perceived brightness at the location of the sensor.
• Actuator
• A means by which an Agent acts on the World. The output is a simulated physical action. For example, a motor rotating a wheel.
• Experiment
• The Experiment Entity is a container for a World, and a set of Agents (each of which have a set of Sensors and Actuators), and an Objective Function which determines the terminating condition of the experiment (which may be a time duration).
• Laboratory
• A collection of Experiments that form a suite to be analysed collectively. This may be a set of Experiments that have similar setups with minor parameter variations.
• ObjectiveFunction
• The objective function computes metrics about the World and/or Agents that are necessary to provide Supervised Learning or Reinforcement Learning signals. It might instead provide a multivariate Optimization function. The ObjectiveFunction is a useful encapsulation because it is often easy to separate objective measurements from the AI that is needed to achieve them.

Architecture

To enforce good design principles, the architecture is multi-layered and highly modular. Multiple layers (also known as multi-tier architecture) allows you to work with concepts that are at the appropriate level of abstraction, which simplifies development and use of the system.

Each entity is a module. Use of particular entities is optional and extensible. A user will inherit the entities that they choose, and implement the desired functionality. Another modularisation occurs with the AGIEF Nodes. They communicate via interprocess conventions so that components can be split between multiple host computers.

Interprocess communication occurs via a central interface called the Coordinator, which is a single point of contact for all Entities and the shared system state. This also enables graphical user interfaces to be built to control and explore the system.

These concepts are expanded in the sections below.

Design Considerations

The various components of the system may have huge in-memory data-structures. This is an important consideration for persisting state, distributed operation, and ability to visualise the state.

Processing to update the state of Worlds and Agents will be compute-intensive. Many AI methods can easily be accelerated by parallel execution. Therefore, the system can be broken down into many computing nodes, each tasked with performing a specific computational function on some part of the shared system state. We hope to support massively parallel hardware such as GPUs in these compute nodes.

We will write the bulk of the framework and initial algorithm implementations in Java. Others can extend on this, or develop against the framework in other languages. We will also write a graphical user interface using web technologies that will allow easy management of the system.

Perspectives on the system design

The architectural layers are shown in the diagram below.

Figure 1: 'Architectural Layers'

Each layer is distinct, with strict separation. No layer has access to the layers above, which operate at a higher level of abstraction.

• State:
• State persistence: storage and retrieval of state of all parts of the system at every time step. This comprises the shared filesystem.
• Interprocess:
• Communications between all modules running in the system, locally and/or across a network.
• Provides a single point of contact via a local interface, to any part of the system (which may be running in different physical locations), for both control signals and state.
• Experiment:
• Provides all of the entities that are required for an experiment. These are expanded shortly.
• UI:
• The user interface that an experimenter uses to run experiments, debug and visualise results.
• The typical features would be:
• set up parameters of an experiment,
• run, stop, step through an experiment,
• save/load an experiment,
• visualise the state of any part of the experiment.
• Specific Experiments:
• This is defined by the person experimenting with the system. For example, a specific Agent that seeks light, a specific World that contains a light source, and an objective function that defines the time span for operation.

Another perspective on the design is to view the Services and Entities and their lines of communication. The diagram is colour coded to indicate Layers, as per the diagram above.

Figure 2: 'Services and Entities'

The Coordinator and Database are services. The Coordinator is shown at the centre, as described earlier (Architecture section), being the primary point of contact for Entities and potentially other clients such as a Graphical User Interface.

A similar perspective is shown in an expanded diagram below that illustrates the Database API module and the distributed implementation of the Coordinator in the Interprocess layer, enabling Entities to run on separate machines. This is just one possible configurations; there can be multiple slaves, each with multiple entities.

Each bounding box indicates what we refer to as an AGIEF Node (or node for short). The node comprises a process that provides a context for execution of one or more entities, or other clients such as the GUI.

Figure 3: 'AGIEF Nodes'

We looked at popular No-SQL web storage systems (basically key-value stores) which are very convenient and flexible due to the inherently dynamic, software-defined schemas and HTTP interfaces. However, we have a relatively static schema for our data, on which we will build utilities for managing experiments and visualising data. In addition, relational databases such as MySQL and PostgreSQL are beginning to offer HTTP interfaces as well. Whether we pick a NoSQL or Relational Database, we will require a HTTP interface.

A third perspective is the data model that represents the system in its entirety. This is the model implemented in the database.

Figure 4: 'Data Model'

The data model stores the entire system state, including hierarchy and relationship between entities, as well as the state of each entity. With a RESTful API exposing the database, we have a shared filesystem accessible as a service, essential for distributed operation and restoring the system at any point in time.

Future Work

We will shortly be releasing an initial version of our framework and we'll post about the technology choices we've made, and some alternatives. We'll include a demonstration problem with the initial release and then start rolling out some more exciting algorithms and graphics, including lots of AI methods from the literature (we have hundreds in our old codebase ready to go).