Empiricism in the foundations of cognition

The first step in our intellectual history is behaviorism of Skinner and his followers. We are aware that there are important predecessors to this school of thought (Walker 1992), but we do not concentrate on them as they are not as strongly theoretically founded and do not have to systematically answer challenges to their theoretical commitments. For us, the debate starts to be genuinely about empiricism at the moment when there is a serious contender that rejects empiricist assumptions. While the first part of our philosophical story relies heavily on debates about forms of language acquisition, it serves only to illustrate general foundations of associationist learning strategies and their legitimate criticism. We first introduce basic building blocks of behaviorism. Then we take up Chomsky’s critique of behaviorism, beginning with his 1959 review of Skinner’s Verbal Behaviour and continuing through the 1970s and see how Skinner and Quine altered their position in its light.

Let us start with the basic tenets of behaviorism. First, behaviorism holds that mental entities are not explanatory. This is not tantamount to a claim to eliminate the mental domain. Instead, it is meant to expel the mental as an explanatory category in psychology. Hence, the notorious ‘black box’ argument: whatever may go on inside the black box of our heads, i.e., a subject’s mental processes, is irrelevant to explanation. Second, the radical empiricism of behaviorism is restricted to externally observable inputs (known as stimuli) and outputs (known as behavior). It subscribes to a strict empiricist understanding of what constitutes credible scientific entities. If it is to be scientific, psychology must search for correlations between stimuli and outputs. Correlations identified by researchers do not necessarily correspond with traditional psychological notions. Third, behaviorism rejects traditional (or folk) categories of psychological explanation based on thoughts, attitudes, and other psychological states. Explanations of behavior must fundamentally be based on the observed data. The work of a behaviorist should thus proceed in a piecemeal fashion. Scientists need to focus on particular sets of stimuli and on particular behaviors. The correlation of behavior with desired outcomes is at the heart of the behaviorist explanation for learning. Learning consists of links between stimulus and behavior governed by operant conditioning. In the strict empiricist fashion, it is the history of encounters with a given phenomenon that shapes up any individual future performance on the relevant task. It is shaped by the tendency to repeat behavior that is rewarded in a particular situation, and to refrain from behavior that is punished. This process gives us the principles of positive and negative reinforcement—the principles that will become crucial in artificial neural networks several decades later. The strategy envisioned by the behaviorist theory of the mind aims at a creation of an implicit list of probabilistic correlations of stimuli and responses—and little else (we will return to this point in Sect. 1.4 and later).

In philosophy, behaviorism is most famously represented by Quine. In fact, Quine and Skinner were close associates. On their picture, psychological processes are not based on internal representations or models. Instead, they take place in the purely physical space of interactions. These processes are therefore not far away from later connectionist views on categorization and other cognitive tasks, where interactions between neurons and input/output relations are all that matters.

In his work, Quine aimed to build a theory of language learning based on conditioning. When a child is presented with a red ball, for example, the child might be rewarded for uttering the word “red”. The basis for learning language (and hence everything else) is ostentation, or constant pointing to objects and naming them (Quine 1950). Our primary aim is not an exegesis of Quine. Rather, we concentrate on an argumentative exchange between Quine and his prime opponent Chomsky as it is this exchange that forces Quine to abandon some of his early strict empiricist inclinations, and implicitly embrace some nativist presuppositions.

As Chomsky was recently involved in exchanges with deep learning advocates (Norvig 2017), we want to point out that a predecessor for these exchanges took place in the 1950s and afterward—namely, Chomsky’s debates with Quine. Interestingly, various recently employed arguments on deep learning (see Sect. 3 below) resemble those employed in their early debate.

Chomsky questions the innocence of the notion of stimulus upon which the Quine account for language learning rests. To borrow his example, suppose we seat a subject in front of a red chair and wait for an utterance. If the subject says “red” (or “chair”, or “red chair”), we praise the subject’s correct utterance. Our cheering is meant to reinforce such utterances following presentation with such stimuli. But suppose the subject says, “It smells funny in here”. Then, the stimulus of the utterance must be the smell of the room (and the bombarding of the subject’s olfactory receptors). In that case, what counts as stimulus depends on what the subject utters. This leads to a bigger worry. By invoking the method of ostension, one tacitly introduces an intentional vocabulary. Such vocabulary needs to specify what a speaker intends to point to through an ostensive gesture. It must also identify the link between the referent object and the co-occurring utterance. This, Chomsky (1967) argues, is nothing other than a retreat to mentalist explanation. To make sense of the response, we must refer to the stimuli. Locating the stimuli requires knowing what the response is actually about. The intensionality of the utterance presents a further problem. The same stimulus may invoke many responses under various modes of presentation. Without recourse to mentalism, this many-to-one relation cannot be resolved at all.

These arguments first appeared in Chomsky’s review of Skinner’s Verbal Behaviour and apply, mutatis mutandis, to Quine’s position. They are all part of a more general problem wherein utterance and stimuli seem to be independent of one another. For instance, we often utter names when their bearers are not present. Broader considerations about context play a crucial role in establishing the existence of any relation between utterance and its referent. Quine seems to be aware of this complication and in his later writings claims that reduction of many-to-one relation can be achieved via a specific mechanism within human subjects: “…learning depends indeed on both the public currency of the observation sentences and on a preestablished harmony of people’s private scales of perceptual similarity.” (1995, 254). It is worthy of noticing that metaphorical language of preestablished harmony and private scales neither offer a satisfactory explanation nor present a firm empiricist stance.

That more is needed in explication of the complex relation between worldly inputs and linguistic outputs is clear to many observers. Nulty describes the situation in no uncertain terms: “The typical empirical perspective on learning the referents of single terms is that of an overwhelming problem space in which the novitiate language learner must find the correct connections for words and objects from a practically infinite number of possible couplings” (2005, 377).

Empiricists have often tried to save themselves from falling into the infinity abyss by invoking dispositional accounts, linked to probability theory. Chomsky quotes Quine’s characterization of language as a “complex of present dispositions to verbal behavior, in which speakers of the same language have perforce come to resemble one another” (Chomsky 1968, p. 57, quoting Quine). He then notes that we can treat dispositions as probabilities: “[p]resumably, a complex of dispositions is representable as a set of probabilities for utterances (responses) in certain definable circumstances or situations” (ibid.).

This follows Skinner’s approach, in which “the probability that a verbal response of given form will occur at a given time is the basic datum to be predicted and controlled” (Skinner 1957, p. 27). “The response Quiet! is reinforced through the reduction of an aversive condition, and we can increase the probability of its occurrence by creating such a condition that is, by making a noise” (ibid., p. 35). All of this greatly depends on the notion of resemblance¹ of the context for utterance, something Skinner spends a great deal of time with (as does Quine 1969a, b).

The aim is ultimately to eliminate intensional idioms and replace them with probabilities determined by frequencies. Skinner invents the world ‘tact’ to describe spontaneous behavior in the presence of non-verbal stimuli (such as the presence of a dog, which could prompt the response, “Oh look, a doggy!”). In his discussion, he observes:

[i]t may be tempting to say that in a tact the response “refers to,” “mentions,” “announces,” “talks about,” “names,” “denotes,” or “describes” its stimulus. But the essential relation between response and controlling stimulus is precisely the same as in echoic, textual, and intraverbal behavior. We are not likely to say that the intraverbal stimulus is “referred to” by all the responses it evokes, or that an echoic or textual response “mentions” or “describes” its controlling variable. The only useful functional relation is expressed in the statement that the presence of a given stimulus raises the probability of occurrence of a given form of response. (ibid., p. 82)

This quote is a clear illustration of the attempt to reduce intensional idioms to an extensional notion (counting the frequency of occurrence). But the easiness with which behaviorists move from referential relations to purely statistical ones, occurring in-between utterances, remains very problematic. Quine’s radical metaphysical physicalism makes things even worse. For him, physicalism means that, ultimately, there is nothing in the universe other than atoms moving in the void. Given this underlying approach, behaviorism is a natural fit, as stimuli and responses are observable parts of the natural world, unlike mental entities. It follows that, fundamentally, the probabilities mentioned above are not linking stimuli with sentences, but only one set of physical events with another.² In Sect. 2.4 we will demonstrate how analogical strategies based on statistics are used in machine learning to achieve human-level cognitive capacities.

Chomsky attacks Quine’s behaviorism by arguing that the frequency of an utterance following some occurrence is effectively zero:

“...assuming ‘circumstances’ and ‘situations’ to be defined in terms of objective criteria, as Quine insists, it is surely the case that almost all entries in the situation-response matrix are null. That is, in any objectively definable situation, the probability of my producing any given sentence of English is zero, if probabilities are assessed on empirical grounds” (Chomsky 1975, pp. 310–311).

Hence, he concludes, the probability of producing a sentence in Japanese is the same as the probability of producing a sentence in English (i.e., zero) (Chomsky 1975, p. 311). This consequence follows from Quinean naturalism, as utterances in various languages are just physical phenomena and as such cannot be distinguished from each other on any non-physical ground. This makes statistical analysis a very unsuitable starting point for the endeavour of linking utterances in various languages to their referents or causes. The discussion presented here is a probabilistic restatement of the poverty of stimulus argument.³ This argument standardly states that children do not learn language by stimulus, response, and reward only as these elements are not sufficiently structured to fix correct utterances (of words and sentences). Put differently, children acquire language far too quickly for acquisition to be a matter of finding a proper probabilistic matrix of stimuli and responses.⁴

Interestingly, Quine does not respond by denying that a disposition towards particular types of linguistic behavior should be qualified probabilistically. He instead argues that Chomsky is focused on the wrong probabilities because the probability of a disposition toward uttering a particular sentence is conditioned by very specific circumstances—and hence, not zero at all:

I am puzzled by how quickly he [Chomsky] turns his back on the crucial phrase “in certain definable ‘circumstances.’” Solubility in water would be a pretty idle disposition if defined in terms of the absolute probability of dissolving, without reference to the circumstance of being in water. … Verbal dispositions would be pretty idle if defined in terms of the absolute probability of utterance out of the blue. I, among others, have talked mainly of verbal dispositions in a very specific circumstance: a questionnaire circumstance, the circumstance of being offered a sentence for assent or dissent or indecision or bizarreness reaction. (Quine 1972, pp. 444–445).

Similar attack is waged by MacCorquodale:

Chomsky seems not to grasp the difference between the overall probability of occurrence of an item in a speaker's verbal repertoire, which is the frequency with which it occurs in his speech over time without regard to his momentary circumstances, and the momentary probability of a given response in some specified set of circumstances. (See, for example, Chomsky, 1959, p. 34). The two probabilities are very different. The overall probability that any speaker will say, for example, 'mulct', is very low; it occurs rarely in comparison with such responses as 'the' or 'of. The probability that he will say 'mulct' may become momentarily extremely high, as when he sees the printed word. Of the two, overall probability is a typically linguistic concern, while momentary probability shifts are, in a sense, the very heart of the psychologists' problem, since they reflect the relation between speech and its controlling variables. Under what conditions does an organism speak an item from his repertoire? Simply knowing the repertoire tells us precisely nothing about that (MacCorquodale, 1970, p. 88).

Overall, if conditional probabilities are to offer a solution, we need to consider how they are obtained. Obtaining conditional probability through counting occurrences of pairs of words, phrases, and sentences would require immense amounts of data.⁵ Given that this requirement is unrealistic in the case of human cognitive functioning, it seems some additional apparatus is needed to provide the relevant conditional probabilities. But reference to any such apparatus would, as Chomsky charged, also seem to reintroduce a priori notions and thereby weaken the empiricist inspirations of the Quinean project. As we will see, this debate on prior conditions for processing closely resembles debates on setting parameters within deep learning networks, which is addressed in the final parts (Sect. 4.2) of the paper.

Alongside the above-mentioned issues, additional problems appear within the behaviorist paradigm. One of the most critical involves the notion of abstraction. Humans can abstract from concrete particulars to categories and then apply these categories to other exemplars (we might move from a golf ball, say, to the notion of a ‘sphere’ then on to a ball ornament for a Christmas tree). Abstraction must employ mechanisms other than a basic form of induction over instances. Indeed, it is hard to see how any form of induction could lead to a formulation of claims about abstract objects. This is yet another restatement of the poverty of stimulus argument—no amount of exposure to stimuli gets us to abstract notions. Quine, of course, is aware of this difficulty. He thus supplements his account for language learning with the conditional notion of ‘analogical synthesis’ (for discussion, see Gibson 1987). Chomsky responds that this notion is empirically empty; it can only serve as a serious fix to the problem of abstraction if we are provided with an exact account of what it is and how it is used (1969, p. 56). Section 3.2 will recount how the problematic issue of abstraction has resurfaced, decades later, within the debate on deep learning.

Under the argumentative pressure, Quine later appears to change his position. He denies that the introduction of mechanisms other than the correlation of stimulus and response amounts to an abandonment of behaviorism. For him, “empiricism of this modern sort, or behaviorism broadly so called, comes of the old empiricism by a drastic externalisation. The old empiricist looked inward upon his ideas; the new empiricist looks outward upon the social institution of language…. Externalised empiricism or behaviorism sees nothing uncongenial in the appeal to innate dispositions to overt behavior, innate readiness for language learning” (1969a, p. 58). In another paper from the same year, he is more specific:

Language aptitude is innate; language learning, on the other hand, in which that aptitude is put to work, turns on intersubjectively observable features of human behavior and its environing circumstances, there being no innate language and no telepathy. … Chomsky says 'I postulate a pre-linguistic (and presumably innate) 'quality space' with a built-in distance measure'. But 'postulate' is an odd word for it, since a quality space is so obviously a prerequisite of learning, and since distances in a quality space can be compared experimentally. Quine (1969b, p. 306).

Yet the notion of the linguistic aptitude remains unexplained and unless we learn more, its collapse toward Chomskian model of the mind is likely. In any case, this is indeed a very different empiricism than the traditional one we started from.⁶ We have moved from the original Humean notions where nothing is said about the content of the mind to the Kantian picture with inner mechanisms and structures that shape up incoming stimuli. It marks the beginning of a tendency of hybridization that we will observe repeatedly. Hybridization is a process of enriching the originally pure empiricist position with nativist or explicit representational elements in the face of mounting criticism. While hybridization is not a problem in itself, it makes the defense of the original empiricist position significantly less compelling.⁷ This argument will come to the fore in the Sect. 3 with the discussion of the resource inefficiency of deep learning.

While essential in taking scientific psychology off the ground, behaviorism, with its extreme empiricist tendencies, became unattractive. It did not deliver on its original promises to provide a firm background for reliable connections between stimuli and responses. Given the computational difficulties of such a task, behaviorists faced ever increasing problems with restraining the domain over which probabilistic operations link inputs and outputs. More importantly, Chomsky has forced empiricists to acknowledge that even if such a domain exists, it has to take into account inner mechanisms of the mind. This is an important sign of demise of behaviorist original aspirations. Instead of a pristine black box, we now have a black box with functions, parameters and inner structures. Once these are postulated, it is only a matter of time till we learn more about their precise values and mutual dependencies. The result is a model of the mind, where it would be odd not to label its inner workings as mental. Once this is allowed, the behaviorist program is over. A parallel story can be told about empiricism. As soon as Quine moves from input-based dependencies to amend his approach with “innate dispositions” (see Sect. 1.5), he is significantly weakening its status. One might still insist (as he does) that we are just altering empiricism. After all, ever since Kant, internal constraints on the mind’s processing of external inputs have been widely acknowledged (Van Cleve 1999). The question to be answered is how such constraints are to be mapped out for empiricism to deserve its name. As the reader will see in the end of this paper, there is growing evidence that chances to preserve a credible version of empiricism are rather bleak.

Partly due to its inability to respond to the Chomskian criticism and partly due to unavailability of viable alternatives, two decades following the 1960s have seen an almost complete rejection of empiricist approaches. Computationalism has ruled supreme, with its emphasis on explicit inner representations that are both installed in early computers and postulated in biological minds.⁸ Computationalism equates the mind with a device that employs algorithms on symbols endowed with meaning. Its starting point is psychological. It takes for granted traditional psychological mental states (beliefs, desires etc.) and asks which processes are such that allow for these states to process worldly inputs and culminate in appropriate actions. Its view of the mind is compositional and it takes concepts to be the basic building blocks of cognition (Pylyshyn and Demopoulos 1986). Concepts are combined in accordance with algorithmic rules to create novel contents that can be cognitively utilized in various ways (stored in memory, used in arguments, acted upon, etc.). There are several persistent difficulties with computationalism. We have little idea how tenets of computationalism relate to brain’s neural underpinnings of cognition, how concepts get their meanings or how syntactic brain operations preserve semantic mental relations. Yet these are not the topics to be addressed in our paper. We only want to contrast the explicitly representationalist story of computationalism with both that of its behaviorist predecessor and its current successors in machine learning. In a stark contrast to both of them, defenders of computationalism declare a strongly anti-empiricist stance, with concepts as representations being the precondition of processing of empirical inputs and delivering relevant outputs. Where do concepts come from is an uneasy question, yet some version of nativism seems acceptable for most in the field (Fodor 1998; Marcus 2018a). Given various successes of computationalism across many domains, it looked for quite some time that empiricism is a completely untenable position. Yet with the onset of connectionism in the 1980s, nativist assumptions have gradually lost their upper hand.

It is indeed remarkable that these simple networks, often containing just a few dozen nodes and a corresponding number of connections, have achieved significant results across various fields of cognition. Three elements of connectionism are worthy of special attention. Connectionist networks are very simple, with no explicit conceptual structure of their own. They are also modelled on the brain’s wiring, thereby overcoming an objection of the disconnection of computationalism from its underlying biological substrate. Their flexibility allows for the experimental testing of various models of cognitive processing. Finally, in their thorough associationism, they are as close to an implementation of empiricist ideas as possible.⁹ In diverse categorization and recognition tasks, artificial neural networks have surpassed expectations and performed very —often approximating or even outperforming human subjects. While remarkable in their own right, their success has led their defenders to rethink the basic elements of cognitive mechanisms (for a more detailed exposition, see Rumelhart 1998). As networks achieve their results primarily by being trained on large amount of input data, one revolutionary consequence of their success is implicit support for empiricism.

Advocates of connectionism stress that networks are modelled on the actual (albeit extremely simplified) biology of the brain. The functionality of both neurons and network nodes depends on their interconnectivity and processing of incoming signals. Points of similarity do not end at the level of architecture. They are also manifest at the cognitive level: like human minds, neural networks are well equipped to work with degraded, context dependent, and multivariate inputs. Unlike in classical computational models where the damage of a single component often leads to devastating consequences for the overall task performance, the failure of one or even several network nodes is hardly noticeable. The output might be slightly more inaccurate, yet it remains reliable. Networks also master situations where the damage occurs not within their architecture, but with the stimulus. Just as human beings can perceive and recognize objects under visually challenging conditions, connectionist networks produce correct outputs for partially occluded, degraded, noisy, or otherwise altered stimuli.

Despite their operational successes, there is a limited way to ascribe individual features or recognizably human categories to a particular node.¹⁰ Representations of features relevant to performance of a given categorization task are widely distributed. This has been an exciting piece of news to all defenders of anti-representationalism because it seems to indicate that the artificial systems have no need for robust, semantically evaluated representations. If they prove successful across a variety of tasks, networks might shed a light on human minds as well. Maybe our minds are also devoid of representations and the contrary assumptions have been held due to our inability to shed off folk psychological intuitions. Behind debates on representationalism lies a deeper question about the smallest building blocks of complex mental processes. While computationalism is fully committed to the existence of explicit representations (and has to face the worries introduced by their presence), connectionism follows in the footsteps of Skinner and Quine to exclude explicit representational features from the mind.

While the debate on the nature of representations within networks would lead us away from our main target, we want to at least indicate what is at stake and what positions were adopted by researchers on this issue. In classical computationalism, all representations are fully explicit and, whether they are taken as innate or implemented, there is no issue about their identity and specific role. Conversely, neural networks lack explicit representations, and the debate has emerged on how to account for their absence. Some authors (Rougier 2009, O’Brien and Opie 2009) argue that given the causal role representations play in any explanation of cognitive processes, it is reasonable to assume there are representations within the networks, albeit in an implicit form. Other authors argue that such an assumption rests on a confusion. Darwiche (2018) puts this point succinctly:

Architecting the structure of a neural network is ‘function engineering’ not ‘representation learning,’ particularly since the structure is penalized and rewarded by virtue of its conformity with input-output pairs. The outcome of function engineering amounts to restricting the class of functions that can be learned using parameter estimation techniques…. The practice of representation learning is then an exercise of identifying the classes of functions that are suitable for certain tasks.

Whatever the outcome of this debate, it is important to see how the influence of connectionism extends beyond the field of computer science and changes the landscape within cognitive science, psychology and philosophy. While it does not promote the radical behaviorist ‘black box’ approach to the mind that recognizes knowledge of inputs and outputs only and ignores any of its inner workings, it makes the mind significantly less transparent than its opponents would like to have it. Were the connectionist program successful, Occam’s razor would make postulations of true representations at the psychological level of explanation unlikely. Psychological terms and their adoption at the level of scientific mentalism would have to go, and the doors of empiricism would remain wide open.

While remarkably successful, connectionist networks are also prone to problems. We will only briefly mention some of them as the real target of our endeavor are analogous problems in the newest incarnation of networks within the deep learning paradigm. From the onset of their application on cognitive tasks, networks have been under attack on their incapacity to solve their target tasks properly. Notorious are exchanges between defenders and opponents of networks on the topics of compositionality and systematicity in natural languages (Fodor and Pylyshyn 1988; Aizawa 2003).

Even with a space for skepticism regarding the notion of systematicity (cf. Johnson 2004), troubles modelling compositional and systemic linguistic intuitions in connectionist networks seem to stem from the absence of any basic building blocks that could be combined to form larger linguistic units.

While we cannot delve into these discussions more deeply, we would at least like to connect them to some other issues that have been identified within the connectionist’s domain. These include difficulties with generalization from concrete examples to a highly abstract level, issue with general rules or output invariance (see Mayor et al. 2014 for an overview).

All these difficulties seem to originate from an overemphasis on empiricism. Any move from the concrete to an abstract level has been difficult for empiricists since early on and the technological advances do not come to a cheap rescue.

We are not arguing that connectionist networks are a priori unable to deal with the set of problematic cases that were used as a weapon against them early on. On the contrary, many ingenious solutions to these challenges have been offered and some are in use till now. For the thesis of the paper, the problem lies elsewhere. While various solutions might have worked, they achieved satisfactory results only because the novel architectures introduced dedicated modules and specific structures within the networks to deal with various mentioned difficulties. Their introduction meant abandonment of the original empiricist credential. Once auxiliary dedicated structural features were in, the tangibility of the claim that networks operate purely on the inputs and arrive at abstractions on their own lost its credibility. While these amalgamated results of newer networks satisfied the majority of connectionist community, because the emphasis has focused on achieving the target efficiency, empiricist undertone has been lost. None argues that complex connectionist networks with memory slots and explicit categorization modules offer a strong support to empiricism.

While the connectionist program has continued mostly uninterrupted, its philosophical significance has temporarily ceased. This was largely because the debate over the nature of connectionism has moved from philosophical quarrels about the building blocks of cognition into practical concerns about optimal architecture for various tasks that systems were trained for. Yet, very recent developments in the field (since around 2015) have once again galvanized disputes about suitable approaches to cognition based on complex neural networks. The last decade has seen a revolution in artificial intelligence (AI) based on a variety of sophisticated network architectures that are often grouped together under the rubric of ‘deep learning’ (DL). The term usually refers to very large-scale networks consisting of tens of thousands of nodes with a multitude of mutually connected layers and additional dedicated features. While starting off from basic connectionist principles, deep learning resulted from a skillful combination of important architectural insights and technological advances that moved the field significantly forward.

Deep networks entered public imagination by proving remarkably successful compared to all other competitors. They achieved superhuman results in a variety of domains, such as image recognition and highly sophisticated game playing (checkers, chess, Go, and Starcraft), and very good results in a number of other fields (a list can be found at https://​deepindex.​org/​). Experts recognise that without hardware advances, the field could not have achieved these successes. Geffner (2018) admits that: “The recent successes have to do with the gains in computational power and the ability to use deeper nets on more data.” (p.2) In his extensive historical review, Schmidhuber (2015) describes how techniques essential for the current string of deep network accomplishments are founded on principles and theories that have been known for decades. Similar observations have been made by Darwiche (2018) who credits Oren Etzioni (see ibid, ft. 9) with the thought about not-so-novel theoretical foundations of deep learning. Whether hardware powers bear crucial responsibility for the DL successes is debatable, yet it is likely that a brute force is not solely responsible for its successes. Without specific architectural advances that we comment on in the next section, the field would not achieve what it did. Darwiche (ibid.) concurrently speaks of the employment of better statistical methods for data fitting in various current approaches. Skeptically, he also assigns a role for the hype surrounding the deep learning to the downgrading of the measurement of success. For example, in the domain of language translation, early challenges in machine translation quantified the success rate of a system to translate a previously unknown text to a foreign language and back while novel approaches are fine with giving us a reasonable enough estimate of what the target text in foreign language is about. The so-called gist translation would fail the early criteria on translation success miserably, though it became the recent new standard of achievement.

We are not in a position to adjudicate a debate about the causes for DL successes. Instead, our investigation is interested in its philosophical significance. Despite the notorious triumphs of DL and vocal voices to the contrary, we believe there are good reasons for skepticism about the kind of support deep networks provide for empiricism. Our aim in this part of the paper is twofold. While we want to show that current debates on the nature of cognition via deep network architecture echo some of the core issues from the Chomsky/Quine debate, our purpose is to go further. We aspire to provide additional arguments for skepticism with regards to the overall tangibility of empiricism.

There are many types of deep networks, differing in architectures and the presence of specialized submodules. Instead of a simple forward-feed model where information flows from input nodes through internal nodes towards output ones, current more complex nets include backward loops and dedicated memory modules. Because the field of deep networks is multifaceted and the optimization tasks remain of a prominent interest to the majority of researchers, only few nets are built with intentionally empiricist principles in mind. Many researchers are aware of the distinctions between bottom-up purely data-driven networks and more complex setups that includes internal modules, various gates or encoder/decoder architecture (see Baroni 2020). All of these additional architectural features have been utilized to enhance networks’ performance. As we have already indicated, our primary concern lies with the first, seemingly simpler group of networks, devoid of dedicated internal modules. There are various ways to categorize these distinct types of nets. Darwiche (2018) speaks of model-based and function-based approaches to AI, with the latter determined solely by inputs. Geffner (2018) distinguishes between solvers and learners. While the solvers compute their outputs in accordance with a model, learners are empiricist in nature, driven by the data. Geffner further divides learners into two sub-groups: deep learners and deep reinforcement learners. It is the last group that is of most interest to us as it utilizes “a non-supervised method that learns from experience, where the error function depends on the value of states and their successors” (2018, p. 2), while standard deep learning regularly relies on supervision. An analogous distinction of model-free and model-based systems is drawn in more details by Lake et al.:

The statistical pattern recognition approach treats prediction as primary, usually in the context of a specific classification, regression, or control task. In this view, learning is about discovering features that have high-value states in common—a shared label in a classification setting or a shared value in a reinforcement learning setting—across a large, diverse set of training data. The alternative approach treats models of the world as primary, where learning is the process of model building. The difference between pattern recognition and model building, between prediction and explanation, is central to our view of human intelligence. (Lake et al. 2017, p. 3)

In this text we will follow the terminology of Lake et al. and speak of model-free and model-based systems, with model-free approach being our primary concern.

The quotes above might make one believe that both systems are essential for understanding cognition. However, a part of the commotion surrounding the recent advances of deep learning stems from a conviction of some of the researchers that there are good reasons to concentrate exclusively on the model-free variants while regarding model-based approaches as secondary. If so, then the combination of ever larger data sets and brute force computing brings upon the final justification of empiricism. When Silver et al. (2017) describe their famous AlphaGo Zero system, they use unequivocal language:

… our results comprehensively demonstrate that a pure reinforcement learning approach is fully feasible, even in the most challenging of domains: it is possible to train to superhuman level, without human examples or guidance, given no knowledge of the domain beyond basic rules. (emphasis ours)

They then continue by contrasting the performance of their system with that of human beings:

… humankind has accumulated Go knowledge from millions of games played over thousands of years, collectively distilled into patterns, proverbs and books. In the space of a few days, starting tabula rasa, AlphaGo Zero was able to rediscover much of this Go knowledge, as well as novel strategies that provide new insights into the oldest of games. (emphasis ours).

The reader should take into account straightforward gestures toward empiricism in both quotes. Some researchers even argue that model-free systems can do all the work that we expect from any cognitive system. Ng endorses a full (present!) replacement of human cognitive capacities by artificial systems: “If a typical person can do a mental task with less than one second of thought, we can probably automate it using AI either now or in the near future.” (Ng 2016).

It is difficult to take these quotes seriously. While partly justified by practical successes of their systems, these quotes rely on strong philosophical assumptions which have been sufficiently scrutinized neither by the computer science community nor by the external experts. To defend empiricism in the form of the model-free approaches requires a thorough philosophical exercise. At least one philosopher has offered his methodical support for these optimistic judgments. In his 2018 paper, Buckner argues that most direct philosophically support for empiricism comes with convolutional deep networks. While relying on the already mentioned hardware advances, three characteristics have combined to make convolutional DNs strikingly powerful: network depth, convolution, and pooling (Buckner 2018, p. 5350). Originally, networks’ depth referred to the number of its layers. The greater number of layers, the deeper the network. In accordance with this assessment, a straightforward analysis might take the early connectionist attempts as fairly shallow and recent developments as significantly deep. Yet, the current complexity of the architecture might call for a more detailed assessment of depth criterion. Several suggestions from the field go beyond a simple counting of layers. Schmidhuber (2015) highlights the role of credit assignment paths in tracing the causal origins of a given output, while Sun et al. (2016) concentrate their effort on the effectiveness of margin bounds. On all approaches, the depth reflects a measure of the complexity of network processes, and proves an ever-increasing hierarchical complexity of network architecture. From the perspective of empiricism, it is important to note that the greater depth permits of a more nuanced analysis of the input data. In combination with other two mentioned features (convolution and pooling), layers at different depth can focus on distinct features of the input. The invention of convolution altered the foundations of neural networks substantially. Convolution refers to filtering algorithms that select certain input features as belonging to a particular category. For example, filters may discern the presence of an edge or a colour at a particular location in a visual object recognition task. By doing so, convolution extracts from the image specific features that lead to its ultimate classification. Crucially, the combination of depth and convolution allows for detection of invariances that are common across various target objects from a given category but might not be straightforwardly detectable at each instance of object presentation. For example, in detection of faces, eyes might be once fully visible, once seen only partially from an angle and once completely hidden, yet the network learns that despite these noticeable differences, one category of objects (a face) is always present. Finally, the pooling mechanism assesses detected lower-level features and sends an affirmation of their presence to a next layer. It works by averaging over the results of convolution, or by using a particular down-sampling operation such as max-pooling (for details, see the cited piece by Buckner). From the perspective of empiricism, pooling allows for capture of categories across various layers of complexity. Analogously to the processing of images in brain’s visual areas, networks can work their way out from the detection of the simplest features like lines and shadows all the way to the high-level categories like faces. Thanks to these three characteristics of depth, convolution and pooling, networks can learn to classify a set of objects despite a great number of nuisance variations in their presentation.

Given that convolutional networks arrive at their results solely with the help of the three above-mentioned features, they present for us a prototypical example of function-based, model-free architecture. As such, they bring in the approach that is the closest to the empiricist mind, thereby deserving a detailed critical scrutiny. In the remaining part we will demonstrate that model-free networks of this type face their own serious limitations, significantly weakening support for the thesis that artificial minds can be empiricist.

The first category of problems arises from quantitative requirements on deep learning. Data dependence is listed among the most perplexing issues in the field (Tan et al. 2018).

Let us compare the learning curve for human beings and networks on identical tasks. The often-cited successes of deep networks in game playing seem less impressive when considering how humans can achieve similar results with much less input—and much more quickly (see Fig. 3 in Lake et al. 2017). To attain the celebrated victory in the game of Go, the winning algorithm had trained on a number of games that would correspond to some 200,000 “human” years of playing the game—relying on extremely large data sets. Networks typically demand training sessions that last for tens (and hundreds) of thousands of trials. It is only after this amount of exposure that they are able to categorize objects from a given domain. This does not even remotely match the human ability to learn from just one example (Lake et al. 2015).¹² It is unclear (and improbable) whether increasing the number of nodes and layers or tweaking the internal structure of a network can remedy such a huge discrepancy.¹³

While the belief in the brute force paradigm of ever-increasing computational power to overcome resource inefficiency remains popular, there is an increasing skepticism about the approach. One of the most persistent objections is the lack of learning-to-learn strategies within networks (Lake et al. 2017). This limited ability to come forward with new learning strategies is what prevents model-free systems from reducing the size of data sets, required for training.

Yet, we also want to argue that, while indicating a significant gap between humans and networks, resource inefficiency is in itself not an argument against generic empiricism. If the crux of empiricist commitments relies on derivation of categories from pure data, we should not be surprised that immensely large data sets are required. On the other hand, an attempt to explain human minds in traditional empiricist terms loses its attractiveness due to the significant dissimilarity in efficiency of learning strategies between networks and humans.

Abstraction is a broad notion and when researchers report failures of networks to abstract, they often mean different things. The most common usage is that of obtaining invariant categorial information about a given target from its tokens. A related usage refers to a possibility to derive general rules from circumstantially distinct instances of certain phenomena. While the first notion of categorization is crucial for various discrimination tasks, the second is often invoked in language comprehension and production. Finally, there is a third layer of abstraction, that of uncovering causal relations between phenomena. It plays a dominant role in science, but is also prominent in folk theoretical explanations. We will briefly demonstrate that, within the field of model-free deep learning, processes of abstraction face difficulties at all three layers.

The general project of artificial intelligence was originally conceived as that of building artificial systems that attain general intelligence or at least solve generic problems (for the very early formulation, see Newell, Shaw and Simon, 1959). While it is hard to specify what the criterion of attaining general intelligence or generic problem-solving amounts to, one requirement seems obvious. It is a system’s transferability. A system can be assigned general intelligence or said to be capable of solving generic problems when its success to solve tasks in one domain can be transferred onto another domain. The scope of generality is difficult to set beforehand (should one system be able to solve quadratic equations, design reclining chairs and shoot basketball to count as a generic problem solver?), so some reasonable restrictions of the target domain is legitimate. Even with such restrictions in place, no system has demonstrated its ability to solve generic problems. However, given the input-driven architecture of model-free networks, their attainment of the general problem-solver’s goal seems particularly hopeless. Their training, focused on a particular task, makes transferability of the resulting function particularly troublesome. Geffner (2018, p. 4) describes “transferring useful knowledge from instances of one domain to instances of another” for model-free networks (“learners”, in his vocabulary) as particularly challenging. He also explains why that is the case: “Learners can infer the heuristic function h over all the states s of a single problem P in a straightforward way, but they cannot infer an heuristic function h that is valid for all problems. This is natural: learners can deal with new problems only if they have acquired experience on related problems.” (p. 3). Importantly, it is fair to say that some transferability is achievable in deep learning networks. However, as Marcus observes, the applicability of a learnt function is surprisingly restrictive: situations “in which the deep reinforcement learning system is confronted with scenarios that differ in minor ways from the ones on which the system was trained show that deep reinforcement learning’s solutions are often extremely superficial.” (2018, 8) He illustrates the point on various games that networks have excelled at. Networks that mastered Atari games were very different from those successful in Go. Differences within the domain of board games (board size and its symmetry, differences between games with perfect and imperfect information) are sufficient to block an algorithm’s efficacy.

Lake and Baroni (2018) make similar point in the domain of language comprehension, when they claim that recurrent networks “generalize well when the differences between training and test … are small [but] when generalization requires systematic compositional skills, RNNs fail spectacularly” (2018, 1). This is a very unwelcoming consequence of the model-free deep learning approach. Yet it is illustrative of the pitfalls of empiricism. If a system is fitted to a particular task by being endlessly trained on one kind of input data, there is only a diminished chance it would perform well on dissimilar kinds. If empiricist-based models are to bring problem-solving success, it is going to be necessarily restricted to a small subset of tasks. Transferability even within a single domain (say, board games), is virtually impossible to ensure and general problem solving remains an elusive dream.

The property of opacity is hotly debated within the domain of deep learning networks (Burrell 2016; Zednik 2019). Opacity refers to our inability to comprehend functional dependencies within the network. We can also say that it signifies the indeterminate way networks represent humanly recognizable features of input. Due to their immense distributive complexity, deep learning networks are epistemically almost impenetrable. Tinkering with networks is often a matter of trial-by-error as we have very little idea why they perform the way they do and how they learn what they do. Consequently, in comparison to systems with explicit instructions, networks suffer from low debuggability. When nothing is known about their inner dependencies, it is nearly impossible to fix possible errors. This observation points us once again to the ‘black box’ problem that we have repeatedly encountered before. While networks are not black boxes in the strict behaviorist sense, their inner workings are substantially obscured. Black boxing, it turns out, is not only epistemically significant, but it also has practical consequences.

Acknowledging epistemic and practical difficulties of opacity, we try to extend its significance to the debate on empiricism. We illustrate our point with notorious cases of adversarial attacks on deep learning networks. So far, the problem has been covered almost exclusively by the computer science community; philosophers and theoreticians of cognition have not taken a strong enough lesson from these effective efforts to block networks’ performance. Adversarial attacks are simple methods of stalling the success of networks by targeting the categorization process for the very domain in which they are supposed to be expert classifiers. The unusual behavior of networks under the attacks was first noted by Szegedy et al. (2013). Since then, it has been demonstrated that the issue is more widespread than originally conceived. Adversarial interventions either alter non-robust¹⁵ features of the target inputs (inputs still remain easily recognizable by humans) or introduce changes at the level completely imperceptible to human observers. During the attack, alterations of the input data substantially weaken or completely paralyze successful operation of a network.¹⁶ There are two versions of adversarial attacks, white box and black box ones (Huang et al. 2017). A white box attack requires knowledge of the network’s architecture. This knowledge expedites access to system vulnerabilities. While white box attacks are interesting, from the perspective of empiricism black box attacks are more intriguing. These occur with no prior knowledge of a particular network architecture. “Attackers” only know the training set and categorization task (e.g., the classification of a picture as a human face). Upon witnessing a black box attack, an uninitiated observer might be very puzzled by the behavior of the network that fails utterly, while all features of its operations remain apparently undisturbed. The fact that consolidated networks can be fooled by imperceptibly minor perturbances on the input side should cause alarm for any proponent of empiricism.

The scope of these disturbing outcomes has been placed under scrutiny by Ilyas et al. (2019). Their research team suggests a reversal of a common explanation of successful adversarial attacks. The previously held views explain vulnerability to attacks as resulting from overfitting or particular design flaws. The authors point out that non-robust features of the dataset, susceptible to the attacks, actually perform essential functions within the network. These features that humans are not capable of detecting, help to optimize the performance of the network. In an insightful experimental setup, Ilyas and his team separated humanly undetectable features within a given category (e.g., planes) that are vital for network’s classification and overlapped them onto images of a different category class (e.g., animals), preserving the original labels (names of airplanes). The network was then trained on these newly conjoined images with labels from the original category set. Although the exercise was deeply confusing in human terms, the trained network was eventually able to classify a standardly labelled dataset with both images and labels that corresponded to the robust human categories. (Trained on pictures of animals that contained undetectable features of planes, together with the plane labels, the network learnt successfully to classify planes.) As the researchers concluded, “this demonstrates that adversarial perturbations can arise from flipping features in the data that are useful for classification of correct inputs (hence not being purely aberrations)” (Ilyas et al. 2019, p. 2).

There is a further worrisome consequence of adversarial attacks. As Ilyas et al. (2019) and other teams have observed, these attacks are transferable across various architectures trained on identical data sets with unexpected ease. Authors suggest why: “different classifiers trained on independent samples from [the same] distribution are likely to utilise similar non-robust features” (ibid., p. 8). Networks find humanly indiscernible input features very instrumental in their image categorization.

These recently discovered properties of network make them not only vulnerable to malicious attackers, but also inform us about principled limitations of model-free networks. Crucially, these observations of adversarial attacks should be very unwelcoming to any defender of empiricism. If the thesis of empiricism consists of utmost reliance on human experiences to obtain the cognitive content, with networks we see those experiences as we know them, are actually secondary. Some other, humanly undetectable features of input turn out to be cognitively more important than experiences.

Quite a bit has changed in the outline and efficacy of neural networks over the last few decades. We are now facing a much more aspiring field, and the successes of deep networks across many areas have not gone unnoticed. As a consequence, there is renewed scholarly interest in capturing their philosophical significance, especially in their justification of empiricism. We have seen that an inclusion of convolution led attempts to resuscitate networks’ support for the theory. Buckner (2018) argues that, by an automated creation of categories solely from inputs, networks perform transformational abstraction, fully in line with the spirit of empiricism. Convolutional networks apparently support the original empiricist conviction of the real possibility to derive cognitive categories exclusively by extensive processing of a sufficient amount of empirical data. Yet, as evidenced by the adversarial attacks, as well as problems with abstraction and non-transferability of algorithms, processes within networks do not resemble those we encounter in the human mind. If a network aims only at optimizing performance on a given task, its creation of an unusual representational structure presents no major problem. The difficulties arise when the parable of optimizing features is stretched outside of its intended target domain.

Drawing conclusions about human cognitive structures and processes from networks presents an example of such a stretch. Debates on empiricism started off as queries about the constitution of the human mind. The fact that networks generalize in their own unique way provides little support for the resolution of disputes about the workings of the mind or (general) intelligence. Given how distinct their resulting functions are from human methods of abstraction, knowledge transfer and causal reasoning, it is impossible to use them as means of justifying philosophical empiricism. It is appropriate for the advocates of model-free deep networks to argue that their nets offer solutions to a particular problem, within a given data set, but not much beyond that.

Interestingly, even their defenders of the networks are aware that to match the human mind, additional structural features need to be fitted into the system. They speak of an introduction of various priors that influence the net’s functionality. Their notion of priors is importantly distinct from the one standardly used in probability theory. Priors encompass any factors that are fixed within the network and precede its actual operations. As will be noted below, this usage comes closer to that of I. Kant and his system of categories. Ilyas et al. (2019) offer “enforcing a prior over the features learned by the classifier” (p.10) as the method for approximating artificial networks to the human mind. Even Buckner (following Goodfellow et al. 2016) admits that to achieve a robust system of categorization, networks would have to adopt “infinitely-strong, domain-general priors”.¹⁹ The broadly-conceived priors might come in the form of modules and settings that amend existing model-free networks to match corresponding human capabilities more closely. Yet the addition of such priors is a far cry from empiricism. This admission is especially surprising for Buckner as it reverses his original commitments to empiricism. Once he allows “infinitely-strong” priors to play a prominent role within the system, it is hard to understand why such a system would deserve to be called empiricist in the first place.

Buckner is not the only one voicing concerns about prospects of purely empiricist approaches to categorizations. In fact, there is a line of criticism that denies the very starting point of the argument in the support of empiricism by deep network results. In his critical analysis, Marcus (2018b) has been very vocal in pointing out various obstacles to call networks model-free and, thereby, empiricist. He demonstrates how various crucial components of successful networks are not an outcome of pure processes of generalization, but instead result from a series of conscious decisions of designers and their endless tinkering with the nets’ architecture. From an inclusion of a search algorithm²⁰ to various decision trees and a chosen number of internal layers, a network is not pristine tabula rasa. Even the choice of a general network setup, its depth and connection density are decisions that come prior to its training and as such have little to do with the empirical content. Every such a choice of parameters goes contrary to the empiricist spirit of model-free architecture. In fact, the whole point of Marcus’ paper is to deny that model-free networks deserve their name. There are too many inherent designed features within them to call them model free.

One should not overlook differences between the critical stance of Marcus, who sees the entire enterprise of model-free approaches as misguided in its proclaimed ambitions and a modest admission of Buckner who is aware of the need to adjust parameters within networks. It might be useful to come back to the towering figure of Kant to appreciate these differences.

Empiricism has, at least since Kant, largely discarded its vision of the mind as an empty slate filled in only with empirical content and few rules of associations. This early conception, which, as we have seen, still has certain appeal among advocates of model-free approaches, was replaced with the empirical fine-tuning of a system that possesses strong initial setup. It looks like Quine is also relying on this reading of empiricism when he speaks of innate dispositions (Sect. 1.4). Initial parameters (“priors” in broad Marcusian sense) can be read as implicit models that streamline empirical content. For the original Kantian approach of the mind, the number of priors remained very small (he introduces twelve very abstract categories of human cognition, such as unity, necessity, and negation, see Kant 1781, A80/B106). It is worth of noting that Kantian categories have direct bearing on the content of cognition: they specify intrinsic limitations on how one can cognize. To use a metaphorical expression, Kant limits the space of possibilities for our “software”. The crux of the current debates on the empiricist foundations have shifted our attention on the “hardware” side as it also appears very important in shaping the cognitive capabilities of a system. As Marcus demonstrates, to match human cognition, networks call for a significant number of different priors. Chosen algorithm and network architectures are already setting limitations on their outcomes (Marcus 2018a). Without setting these parameters correctly, network would not function at all. There are also external adjustments to the network during the training process. Yet Marcus’ criticism is more dimensional. He sees virtually all tinkering by the network’s designers as moving against the spirit of model-free empiricism. We share his conviction that instead of model-free systems, deep networks present fine-tuned cognitive systems, full of various priors. Baroni illustrates the same point on the case of linguistic processing by networks:

Modern sequence-processing networks are complex systems, equipped with strong structural priors such as gates, encoding and decoding modules and attention. They should not be thought of as “tabulae rasae”, as they often were in early debates on connectionism. At the same time, the “innate” biases they encode are rather different from those assumed to shape human linguistic competence. Some researchers are trying to inject into modern networks priors closer to those traditionally postulated by linguists, such as a preference for hierarchical tree structures. … neural networks might solve complex linguistic tasks, but not in the way we expect them to be solved. (Baroni 2020, p. 3–4).

With inherent limitations of model-free networks that we have mapped, and empiricism largely out of the picture, there remains a further question about how we can ensure that networks resemble the human mind more closely. It is important to stress that a demise of universalist aspirations of model-free approaches should not be interpreted as a signal to return to purely representationalist systems. There is no desire to go back to the fully representational models in the spirit of expert systems or other fully explicit computational approaches. While they might have historically proved very successful in various domains, their inflexibility and thirst for explicit knowledge representations make them unlikely candidates for general cognitive systems.

After the failures of both fully representational and model-free systems, a middle road stays open in the domain of creating artificial cognizers. We do not want to commit ourselves to a particular approach that might be necessary for such a project. A conclusive answer to this outstanding problem requires substantive empirical research and while some of it has already been conducted, a lot more needs to be done. However, a general consensus seems to be emerging. In the words of Pearl (2018): “human-level AI cannot emerge solely from model-blind learning machines; it requires the symbiotic collaboration of data and models”. This view is supported by others in the field. Geffner (2018) defends an integration of model-free and model-based approaches (to use his terminology, “learners and solvers”). Analogous programmatic proclamations have been also made by Spelke and Kinzler (2007) and Marcus (2018b). Marcus specifically defends the need to include several “computational primitives” that help with various cognitive tasks deep learning handles only with difficulties. In his (2018a) paper he recalls a debate with Yann LeCun where he talked about the issue: “At my October 5, 2017 debate with Yann LeCun, I had an opportunity to draw up a preliminary list. The list I proposed was, roughly, the union of a set of computational primitives that I had advocated for in my book The Algebraic Mind (Marcus 2001), and a set of conceptual primitives drawn from Elizabeth Spelke’s work on cognitive development (Spelke 1994): representations of objects, structured, algebraic representations, operations over variables, a type-token distinction, a capacity to represent sets, locations, paths, trajectories, obstacles and enduring individuals, a way of representing the affordances of objects, spatiotemporal contiguity, causality, translational invariance, capacity for cost–benefit analysis”. The response of his opponent was highly unorthodox: “Provocatively, LeCun argued (when pushed by the moderator, David Chalmers) that none of these need be innate.”

These are not just ideological advocacy calls. On the contrary, they result from a long history of research on the mechanisms and content of human minds these authors conducted. Recently, Lake et al. (2017) argued that the inclusion of human-like modules, such as those for folk psychology or folk physics, is advisable for obtaining network performance proximal to that of humans. They argue that the presence of both approaches is critical for cognition: “The difference between pattern recognition and model building, between prediction and explanation, is central to our view of human intelligence” (Lake et al. 2017, p. 3) and as such it has to be preserved in artificial systems as well. The authors also believe that analogous enrichment of the networks facilitate learning efficiency in artificial systems: “if deep neural networks could adopt … compositional, hierarchical, and causal representations, we expect they could benefit more from learning-to-learn” (ibid., p. 17). All these claims serve as the final nail to the coffin of the empiricist program and very much continue the discussion on how to overcome the poverty of stimulus (Sect. 1.4). Hopefully, we have finally learnt that purely empiricist strategies can neither explain human behavior nor sanction human-level efficiency in machine learning.

Still, precise arguments about the form and scope of various cognitive models and forms of integrations with neural networks are open to further inquiry. They cannot be resolved by armchair theorizing, though some inspirations might be suggestive despite being only shallowly anchored in empirical evidence. One such a framing idea is put forward by Geffner (2018) when he argues for the need to integrate model-free and model-based approaches in machine learning because this integration is widely observed in operations of System 1 and System 2 of human minds (see Kahneman 2013). If this integration of dual systems of processing is successful in us, there is a good reason to presume it will be beneficial to artificial systems as well.