Malfunction of artificial neural networks, such as those that include analogy priming, might cause those networks to make mistakes about the nature of reality and display analogous behaviour to human brains and minds which are not normal. This may be especially relevant to understanding creativity, as well as the development of childhood onset neuropsychiatric illnesses. The following are three such types of defect in computational models of reasoning – well known in the artificial intelligence field – which can arise in analogy priming. (For general references, see Rumelhart & McClelland [1986], Haykin [1998], Rojas and Feldman [1996], and Muller et al. [1995].)

Example 1

In the first example (Fraser Reference Fraser1998), a network was trained on pictures of a landscape with and without a tank present. In those with the tank, it was placed in all sorts of different locations, obscured to various degrees, shot from different angles. The landscapes without the tank were similarly photographed from various angles, and some pictures from each set were used to train the network, the data being provided to indicate whether or not the tank was present. The goal was to see if the network could “learn” the task of telling whether or not a tank was present in a picture it had not seen – a very simple form of analogical reasoning based on previous pictures but without the need to explicitly isolate the tank as an object in the scene, which hopefully would lead to the ability to recognize arbitrary objects in arbitrary scenes. The pictures not used in training were used as tests – and the results were spectacular, with essentially 100% correct discrimination between pictures with and without tanks present. How the network had done this was stored in all the weights of connections between nodes, and the general feeling was that whatever was going on, it would likely not be understandable by a human being. However, on examination, it turned out that what the network had done was effectively to sum up the brightness of each pixel in the photograph. The pictures with the tank had been taken on a different day than the pictures without, and the network had discovered a significant difference between the two sets of pictures – one set had been taken on a cloudier, and therefore darker, day than the other. The network had learned to classify the pictures by analogy, but had used the total brightness of the scene rather than the presence or absence of the tank.

This “error” might be seen as a basis for unexpected creativity in neural networks, in which new perspectives result from dropping prejudices. For example, in the above example, the input consisted of the raw data. If instead the network had received the input as a list of possible tank-like features (perhaps extracted on the basis of some more logic-based algorithm such as matching features in the scene to features on photographs of tanks), it might have counted up the number of tank-like features and their quality and made a different discrimination. Dropping the pre-classification of features in the scene, in a sense, opened up the “creative realization” that the two scenes were of different brightness – something that typical humans might well (and indeed in this case did!) miss. Additionally, it also shows how selective prejudices can sharpen cognition by making some features stand out.

On the other hand, this alternate solution to the problem might explain how autistic children make social misjudgements – perhaps using unusual aspects of a social scene. For example, if the qualities of the tank were like the intricacies of facial expressions, the processing of which appears to be impaired in autism (Schultz Reference Schultz2005), then excluding this information would lead to some of the social errors that autistic children make in missing facial expressions.

Example 2

Networks that have too many connections between too many neurons often do not work well (Müller et al. Reference Müller, Reinhardt and Strickland1995; Rojas & Feldman Reference Rojas and Feldman1996). This is perhaps not surprising, since it essentially means that almost no weights (connections) are close to zero. Given the high apoptosis rate in the developing brain, one might wonder whether or not any mental disorders are associated with defects in this apoptotic process. Indeed, autism is associated with unusually large brain size (Courchesne et al. Reference Courchesne, Redcay and Kennedy2004). Perhaps future therapies for autism could be based upon restoring normal apoptotic mechanisms during infancy.

Contrasting mechanisms of neurogenesis, neural sprouting, and new synapse formation would also be important in regulating neural network performance. Abnormalities in those new connections and activity, because of either genetic or environmental issues, could lead to problems such as structural non-uniformity in computational models (Rumelhart & McClelland Reference Rumelhart and McClelland1986).

Example 3

An efficient neural network must appropriately switch between flexible and stable states (Haykin Reference Haykin1998; Rumelhart & McClelland Reference Rumelhart and McClelland1986). The stable state of a neural network might be akin to a focused state. Perhaps difficulties in reaching and maintaining stable states in children's brains manifest as the lack of focus and hyperactivity of attention deficit/hyperactivity disorder (ADHD) (American Psychological Association 2000). Perhaps understanding overactive brain circuits may also inform our understanding of abnormally active cortex in epilepsy. Alternatively, states that are too stable may appear like the psychomotor retardation of depression (Sadock & Sadock Reference Sadock and Sadock2004). Perhaps treatments such as electroconvulsive therapy in adults are a sort of “reset,” helping the brain out of a state of excessive stability. Thinking about mentally ill brains as connectionist neural networks which have an impaired ability to attain, maintain, and switch between stable states may lead to novel therapies aimed at augmenting these brain mechanisms.

Analogy does indeed lie at the heart of the acquisition of human cognition, as Leech et al. posit. Connectionist models of the neural networks in brains may help explain how the acquisition of cognitive skills in humans actually works. In addition, apparent errors in the development and maintenance of these networks, which may be modelled computationally, may mimic aspects of mental illness and lead to improved and alternative treatments. This kind of innovative approach may be especially helpful to understand and treat infants and children who are learning critical cognitive skills, yet are not necessarily able to communicate their problems clearly.