Since radical behaviourists first described learning theory, it has undergone a significant change of approach. In the paper “Has the Wheel Turned Full Circle? Fifty Years of Learning Theory, 1946-1996”, Mackintosh (1997) summarises what has been done in the field of learning theory since 1946. His main argument relates to the importance of the associative learning theory. He claims that with appropriate modifications, it is the most powerful approach that science has.
In the presented review, I aim to critically discuss this paper. I start with criticising Mackintosh’s approach of treating cognitive and associative processes as two distinct. Then, I examine Mackintosh’s point of view on animal learning. I conclude that Machintosh’s (1997) main claim about the importance of the associative learning is supported by subsequent research. Nevertheless, it is important not to neglect either associative, or cognitive approach, because they are not fundamentally different, and, at least in humans, both of them play a significant role.
Cognitive versus associative approaches to learning
Mackintosh (1997) starts his paper with a critique of radical behaviourism. In the 1960s, Pavlovian and Instrumental conditionings were main theories to explain animal and human behaviour. However, with cognitivism coming into the picture, learning theory has dramatically improved. Animal theorists started paying more attention to animal cognition, and nothing but good came from it, as researchers have discovered many “rules of associations” that allowed them to see the wider and more accurate picture of how animal learning takes place.
It seems like Mackintosh (1997) thinks that radical behaviourism is outdated. However, it is important not to neglect the effect radical behaviourism had on the field of psychology. Even with taking a cognitive viewpoint into account, reward and punishment still play an important role in shaping one’s behaviour. Moreover, there are implications of behaviourism theory in studying substance abuse (Alcaro, Huber & Panksepp, 2007). Therefore, it is unreasonable to completely abandon classical behaviourism. It is obvious though that Mackintosh acknowledges behaviourism to some extent, as associative learning, in fact, is an improved version of behaviourism.
In the first section of the review, Mackintosh (1997) says that his “own preference is to keep the terms distinct, using “cognitive’’ to refer to processes or operations lying outside the scope of associative learning theory” (p. 882). Therefore, he sees cognitive and associative learning as two distinct processes. He contrasts associative and cognitive explanations of animal behaviour in the next sections. However, it seems that at least in humans, associative and cognitive learning processes are not two distinct, but the two ends of one continuum, with the degree of control being a factor that varies across them (see McLaren et al., 2014). Indeed, there is strong evidence that associative learning occurs both in animals and humans; therefore, associative learning is fundamental. At its highest form, associative learning becomes cognitive. Therefore, according to the modern point of view, treating them as distinct processes, like Mackintosh did, seems not to be the right approach.
Do animals rely primarily on cognition or association?
In the next section, Mackintosh (1997) states that many psychologists try to explain all animal behaviour in terms of reasoning. He considers it not to the right approach and argues against the claim that animals have a theory of mind. However, some modern research suggests otherwise. As such, following Gallup’s (1970) study, Ari and D’Agostino (2016) observed evidence for self-awareness and self-recognition in apes. Nevertheless, other researchers have argued that the evidence regarding self-recognition in animals that we have to date are not very convincing – for example, that recognising self in the mirror does not necessarily implies awareness (see Anderson & Gallup, 2015). Therefore, the theory of mind in non-human primates is still a controversial topic, with more research needed.
Mackintosh’s (1997) cites Thorndlike (1911) on the claim that the laws of association are similar in all vertebrates and maybe even invertebrates. Indeed, neuroscientific studies have shown that even Aplysia shows associative learning (Hawkins, Abrams, Carew and Kandel, 1983; Hawkins, Carew and Kandel, 1986). He makes a strong claim that it is important to always try to understand the behaviour of an animal from the associative viewpoint. As such, he rejects O’Keefe and Nadel’s (1978) point of view on spatial navigation in rats and instead suggests that more plausible (though imperfect) account is the one by Deutsch. He then argues that it is important not to reject associative account if it is not obvious how to apply it. He claims that different animals solve problems differently, depending on their capability for stimulus representations. To understand animal’s behaviour, it is important to know how they represent stimuli, and, knowing that, it is almost always possible to apply associative account. However, after stating this, he mentions that there are cases that associative theory cannot explain and which show that animals have higher cognitive capabilities. He uses Premack’s chimpanzee, Sarah, and Pepperberg’s parrot, Alex as examples of such phenomena. It seems unclear why Mackintosh abruptly introduces these examples, after arguing above that the majority of animals are not capable of cognitive analysis. Moreover, Mackintosh does not give a clear explanation of why these particular animals are capable of cognitive analysis, while others are not.
In the same section, Mackintosh (1977) describes “matching-to-sample” discrimination in different animals and discusses whether animals can show a transfer to new stimuli. Apparently, chimpanzees, monkeys, dolphins, sea lions, and birds show a good transfer to new stimuli in such paradigm, while pigeons do not (however, even the pigeons were eventually able to show transfer – see Colombo, Cottle, & Frost, 2003). The evident problem with those studies might be that animal simply chooses the stimulus based on the recency effect, that is, it might simply select the stimulus that was last presented. On the other hand, if it is not the case, it seems that those animals that show a good transfer are capable of solving problems closer to the cognitive end of the continuum, while those that do not show a good transfer rely on an associative end. Therefore, there is a possibility that dual-processes account is applicable to animals.
Peak shift in pigeons
Pigeons do not show evidence of representing relationships between two or more stimuli (Mackintosh, 1997; however, see Colombo, Cottel, & Frost, 2003). Mackintosh (1997) agrees with Spence’s (1937) position and claims that even though pigeons show transposition, it is simply because they detect the elements that the stimuli have in common. For example, suppose that pigeons are trained to discriminate between two stimuli that differ in brightness. If pigeons are trained to peak S+ instead of S-, then when presented with S+ and even brighter stimulus, they would peak a brighter one. This is a classic peak shift data. After stating that, Mackintosh describes another experiment, where the stimuli did not vary on a real physical dimension and pigeons still showed a peak shift. As Mackintosh claims, peak shift data demonstrates associative learning in pigeons. He concludes that pigeons are not capable of relational learning.
However, Lazareva, Wasserman, and Young (2004) have argued otherwise. In their study, four pigeons were trained to discriminate between one, two, or four pairs of circles. Testing included five new stimuli. What the authors found was that the more stimuli pigeons were trained on, the better was their relational responding. Additionally, transposition increased if the stimuli looked more different. These results are not consistent with Spence (1937) theory, and, consequently, with Mackintosh’s (1997) claim. To sum up, some modern research shows that even pigeons might be able to move further on the continuum in the direction of cognitive learning.
Peak shift in humans
Mackintosh (1997) used Wills and Mackintosh (1998) experiment to demonstrate peak shift in humans. He concluded that normally, if stimuli vary across a real physical dimension, humans show rule use, but when the stimuli are too complex, presented too rapidly, and the rule could not be verbalised, then humans show peak shift. He sees this as a demonstration of associative learning. Similar to this result, contemporary experiments have demonstrated a peak shift data in humans using face dimensions (for example, Spetch, Cheng and Clifford, 2004)
However, Howard (2000) argued in favour of an alternative explanation of some of these results. He claims that explicit instructions might hugely affect the results. When humans are told to respond to S, results that seem like peak shift could occur simply because participants confuse S+ for S. Indeed, in Wills and Mackintosh (1998) experiment, participants were told to remember the appearance of the stimulus for future reference. Therefore, they might have still been using a rule, but simply confusing S+ for S. Generalization gradient can be explained by the fact that the further away from S, the less similar the stimulus look, and, therefore, participants are less likely to confuse it for S. Howard (2000) says that in such experiments, if the group, as opposed to the individual, data is presented, then there is little evidence for generalisation gradient and peak shift. This critique is not to neglect the occurrence of those phenomena; rather, it is a future direction for improving the design of peak shift experiments to show a better demonstration of associative learning in humans. Modern experiments on peak shift in humans involve asking participants whether they were using a rule. This seems to be a massive improvement. Normally, those who are not using or able to verbalise the rule are showing peak shift that is similar to pigeons, while those who are using the rule show transposition and relational learning (Jones & McLaren, 1999).
Mackintosh (1997) concludes the section with experiments that demonstrate that humans are capable of both rule-use and implicit learning (that is, both transposition and peak shift). Importantly, subsequent and more sensitive tests of implicit learning supported its existence in humans (for example, Guo et al., 2013).
Overall, if people are showing peak shift under cognitive load, but given time and opportunity show rule-based behaviour, then there are two possible explanations for thins phenomena. It could either be cognitive, where people are confusing S for S+ in what seems like peak-shift. Alternatively, there might be a dual-processes explanation (McLaren et al., 2014), according to which peak shift is an example of associative learning, which, with an appropriate degree of control, can become cognitive learning. There is still a debate in the literature of which account is correct.
Conclusion
To conclude, Mackintosh (1997) makes a convincing argument that an associative account can explain much of animal and human behaviour. Even though Mackintosh (1997) does not reject cognitivism completely, he obviously prefers the associative account of animal and human learning. However, at least in humans, the dual-processes account can explain learning much better (McLaren et al., 2014). That is why, when considering the importance of associative learning theory, it is necessary not to neglect cognitivism. In fact, associative and cognitive accounts are not qualitatively different. It may even be the case that if appropriately trained, animals are able to show some higher cognitive capabilities, but normally rely on the associative end of the continuum. It is still a debated theory; therefore, more research is required. All in all, even though Mackintosh’s paper has some problems, it is an excellent summary of evidence of the demonstrations of associative learning.