User:ACIre101/sandbox

Reverse Inference
Reverse inference, is a commonly employed practice within neurocognitive science. It involves inferring the presence of a particular cognitive process from a given neural signature, typically employed in order to support a given hypothesis. However, in recent years, various issues have arisen regarding the legitimacy of such inferences. Specifically, many have noted poor selectivity between a neural signature (e.g. activation in a given neural region) and the cognitive process deemed to be putatively engaged as a result. The following provides a brief overview of the positions and attempts made to resolve this issue.

Reverse Inference, and it’s use
Cognitive neuroscience commonly involves the manipulation of a particular psychological process (in light of the cognitive task at hand), and, upon observing the resultant neural activity (e.g. measured via fMRI) makes a ‘forward inference’ regarding what neural regions correlate with the given cognitive process. Reverse inference, in contrast, involves inferring the presence of a cognitive process from a given neural signature. Specifically, researchers may draw upon the results of studies external to the one at hand, in which a forward inference is derived between cognitive process X and brain area Z. Upon observing activation in brain area Z within the study at hand, they may then infer the presence of cognitive process X accordingly

These inferences have been employed across a variety of studies DeQuervain, 2004; Richeson et al., 2003). For example, Greene (2001) scanned subjects while they made decisions on a moral dilemma task. Specifically, conditions were classed as either moral-personal (pushing and subsequently killing the person from a bridge, halting the impending death of five individuals), or moral-impersonal (simply pulling a lever, though to produce identical effects). The moral-personal condition elicited greater activity within region BA 9 and 10 of the medial frontal gyrus and BA 31 in the posterior cingulate gyrus, with lesser activity in the BA 46 in the middle frontal gyrus (right) and BA 7/40 in the parietal lobe (bilateral) .  In light of external experiments which implicated emotional processing within the same former set of neural regions (see Greene, 2001) , they employed a reverse inference, suggesting that the moral-personal condition engaged emotional processes to a relatively greater extent, on account of the particular neural regions activated. Similiar inferences have been made inferring the presence of reward in the provision of punishment  , as well as pup suckling (relative to cocaine) from activity in regions such as the dorsal striatum. Indeed one neuroscientist even self diagnosed themselves as psychopathic, in light of links between hypo activity in the frontal and temporal lobes, and empathy scores.

Problems with Reverse Inference
However, one central problem has surfaced within the literature with regard to the legitimacy of such practices. Specifically, it has been pointed out that such inferences are not deductively valid. That is, any given neural region is often tied to many different cognitive processes (either given it’s multiple functionality in nature, or the weaknesses of methods so far produced to generate adequate constraints . That is, if neural region X correlates with mental process A, it does not follow that process A is active in light of the present activity of region X .  Indeed, multiple possible cognitive processes per neural region tends to be the rule rather than the exception for neural signatures .  For example, the left fusiform gyrus has been implicated in different cognitive processes related to both face-processing and word recognition

However, as Poldrack (2006) points out, it is also noteworthy that given that scientific progress does not rely on deductive reasoning alone reverse inference may be harnessed once necessary precaution is taken. Poldrack (2006) therefore prescribes that one may frame any such inference more formally in Bayesian terms (for box diagrams, see Poldrack, 2006), in order to avoid the pitfalls mentioned above.

From this it is clear that the ‘posterior probability’, namely, the likelyhood of the inference being correct, is directly related to A) the likelyhood of the neural region being active given the mental process,(i.e. consistency of forward inferences (e.g. Henson, 2005 ) and B) the ‘prior’, encompassing the likelyhood of the given process, based on data external to the neural signature. Conversely, the ‘base rate’ of the given neural region (neural activity, whether or not process is present) will be inversely related to the confidence of the inference, as many such cases may involve alternative cognitive processes .  Drawing upon data, either within large scale databases, (e.g. Neurosynth, or meta-analysis’ (see Hutzler, 2014 ), a particular value can then be computed (with >10 considered ‘strong’, (see Hutzler, 2014 ). As discussed below, the greater amount of data one can employ, the greater precision once can generate (and thus strength in probability), in estimating any given reverse inference.

Task conditions and the ‘prior’.
Some attempts to increase the confidence of any given reverse inference have targeted the potential utility of the ‘prior’, that is drawing upon information external to the neural signature at hand, in order to increase the strength of the inference  suggests, in light of the fact that not all mental processes have equal likelyhood of being active during the operation of a given cognitive task, taking the nature of the task into consideration would increase the precision of the inference. For example, while the left fusiform gyrus is tied to both, process related to the mental lexicon (such as visual word recognition), and face perception, the experimenter can also consider the nature of the task (e.g. the presentation of visual words), and conclude that word recognition processes are more likely to be at play.

Neuroselectivity
Big Data

Another means of resolving this issue may lie within relying upon large scale neuro databases (e.g. 'Neurosynth). Briefly, researchers are invited to insert their findings into this database, upon which both reported activations and the cognitive terms tied to such will be uploaded onto the database. Neurosynth can then synthesise all such data across thousands of studies, generating a cumulative result, and thereby produce a more accurate estimate (via probabilistic mappings) of the relation between any given neural region and a cognitive process (see Neurosynth.

Decoding Mental States

An alternative approach aims to make similiar progress with use of machine learning techniques (Haynes & Rees 2006, Norman et al. 2006). More specifically, this involves the employment of particular algorithms which, after processing an initial piece of data, can then ‘learn’ and detect patterns from this data, in order to make predictions based upon novel data. One key challenge has been to produce algorithms which would successfully generalise across participants (training on one and making predictions on the other), rather than simply within, in addition to a range of additional challenges. Although, results appear to be promising (see Poldrack & Yarkoni, 2016 for overview).

Cognitive Atlas Programme

It has also been argued that the reason for poor selectivity (and thus weak reverse inferences) may be due to the possibility that cognitive scientists have not carved the mind into the appropriate ‘pieces’. In this case Poldrack (2006 ; see also Poldrack & Yarkoni, 2016 ) prescribe that cognitive psychologists co-operate to form something akin to the Gene Consortium, that is, by each inserting annotated information regarding the particular nature of the task, the particular cognitive process, and it’s respective constituents deemed to be involved, forming a formal cognitive ontology (i.e. the Cognitive Atlas Programme )  The ambition is to amalgamate all such entries in order to generate a nodal network, specifying the full range of cognitive tasks (e.g. the sternberg task ) previously employed as well as the full range of process (e.g. working memory), and sub-processes (e.g. visuospatial sketchpad) which deemed to be elicited from such tasks. This would then serve as a more appropriate target of inference, when observing such neural signatures as discussed above.

Conclusion
<!-- EDIT BELOW THIS LINE

Reverse inference is employed in neuroimaging (Poldrack, 2006).

Reverse inference

Psp8e5

500455561

PPP4005 - Methods in Brain and Cog Research

Bangor University

Reverse inference, is a commonly employed practice within neurocognitive science (Poldrack, 2011; Poldrack, 2006). It involves inferring the presence of a particular cognitive process from a given neural signature, typically employed in order to support a given hypothesis (e.g. Greene, 2001). However, in recent years, various issues have arisen regarding the legitimacy of such inferences (Poldrack, 2011; Thorpe, 2014). Specifically, many have noted poor selectivity between a neural signature (e.g. activation in a given neural region) and the cognitive process deemed to be putatively engaged as a result. The following provides a brief overview of the positions and attempts made to resolve this issue.

Reverse Inference, and it’s use

Cognitive neuroscience commonly involves the manipulation of a particular psychological process (in light of the cognitive task at hand), and, upon observing the resultant neural activity       based on the concurrent neural activity (e.g. measured via fMRI) makes a ‘forward inference’ regarding what neural regions correlate with the given cognitive process (Henson, 2005). Reverse inference, in contrast - involves inferring the presence of a cognitive process from a given neural signature (Poldrack, 2006; Poldrack, 2011). Specifically, researchers may draw upon the results of other studies which derive a forward inference between cognitive process X and brain area Z. Upon observing activation in brain area Z within the study at hand, they may then infer the presence of cognitive process X accordingly (Poldrack, 2006). These inferences have been employed across a variety of studies (Greene 2001; DeQuervain, 2004; Richeson et al., 2003). For example, Greene (2001) scanned subjects while they made decisions on a moral dilemma task. Specifically, conditions were classed as either moral-personal (pushing and subsequently killing the person from a bridge, halting the impending death of five individuals), or moral-impersonal (simply pulling a lever, though to produce identical effects). The moral-personal condition elicited greater activity within region BA 9 and 10 of the medial frontal gyrus and BA 31 in the posterior cingulate gyrus, with lesser activity in the BA 46 in the middle frontal gyrus (right) and BA 7/40 in the parietal lobe (bilateral). In light of external experiments which implicated emotional processing within the same former set of neural regions (see Greene, 2001), they employed a reverse inference, suggesting that those in the moral-personal condition were more emotionally engaged on account of the particular neural regions activated. Similiar inferences have been made inferring the presence of reward in the provision of punishment (Quervain et al., 2004), as well as pup suckling (relative to cocaine) (Ferris et al., 2005) from activity in regions such as the dorsal striatum. Indeed one neuroscientist even self diagnosed themselves as psychopathic, in light of links between hypo activity in the frontal and temporal lobes, and empathy scores (Fallon, 2013).

Problems with Reverse Inference However, one central problem has surfaced within the literature with regard to the legitimacy of such practices ( Poldrack, 2006; Thorpe, 2009). Specifically, it has been pointed out that such inferences are not deductively valid (Poldrack, 2006). That is, any given neural region is often tied to many different cognitive processes (either given it’s multiple functionality in nature (Chang et al., 2016), or the weaknesses of methods so far produced to generate adequate constraints (Poldrack, 2006). That is, if neural region X correlates with mental process A, it does not follow that process A is active in light of the present activity of region X (Poldrack, 2006). This tends to be the rule rather than the exception for neural signatures . For example, the left fusiform gyrus has been impliacted in different cognitive processes related to both face-processing and word recognition (Hutzler, 2015). However, as Poldrack (2006) points out, it is also noteworthy that given that scientific progress does not rely on deductive reasoning alone (Hustzler, 2014), reverse inference may be harnessed once necessary precaution is taken.  Poldrack (2006) therefore prescribes that one may frame any such inference more formally in Bayesian terms (for box diagrams, see Poldrack, 2006), in order to avoid the pitfalls mentioned above. From this it is clear that the ‘posterior probability’, namely, the likelyhood of the inference being correct, is directly related to A) the likelyhood of the neural region being active given the mental process,(i.e. consistency of forward inferences (e.g. Henson, 2005) and B) the ‘prior’, encompassing the likelyhood of the given process, based on data external to the neural signature. Conversely, the ‘base rate’ of the given neural region (neural activity, whether or not process is present) will be inversely related to the confidence of the inference, as many such cases may involve alternative cognitive processes (Poldrack, 2006; Poldrack, 2011). Drawing upon data, either within large scale databases, (e.g. Neurosynth (Yarkoni et al. 2011) or meta-analysis’ (Hutzler, 2014), a particular value can then be computed (with >10 considered ‘strong’, (see Hutzler, 2014). As discussed below, the greater amount of data one can employ, the greater precision once can generate (and thus strength in probability), in estimating any given reverse inference (Hutzler, 2014; Poldrack & Yarkoni, 2016). Task conditions and the ‘prior’. Some attempts to increase the confidence of any given reverse inference have targeted the potential utility of the ‘prior’ (Poldrack, 2006), that is drawing upon information external to the neural signature at hand, in order to increase the strength of the inference (Machery, 2014, Klein 2012). Hutzler (2014) suggests, in light of the fact that not all mental processes have equal likelyhood of being active during the operation of a given cognitive task, taking the nature of the task into consideration would increase the precision of the inference. For example, while the left fusiform gyrus is tied to both, process related to the mental lexicon (such as visual word recognition), and, face perception the experimenter can also consider the nature of the task (e.g. the presentation of visual words), and conclude that word recognition processes are more likely to be at play.

Neuroselectivity

Big Data One means of resolving this issue may lie within relying upon large scale neuro databases (Yarkoni, 2011). Briefly, researchers are invited to insert their findings into this database, upon which both reported activations and the cognitive terms tied to such will be uploaded onto the database. Neurosynth can then synthesise all such data across thousands of studies, generating a cumulative result, and thereby produce a more accurate estimate (via probabilistic mappings) of the relation between any given neural region and a cognitive process (see Neurosynth  ; Yarkoni, 2011; Poldrack & Yarkoni, 2016).

Decoding Mental States An alternative approach aims to make similiar progress with use of machine learning techniques (Haynes & Rees 2006, Norman et al. 2006). More specifically, this involves the employment of particular algorithms which, after processing an initial piece of data, can then ‘learn’ and detect patterns from this data, in order to make novel predictions. One key challenge has been to produce algorithms which would successfully generalise across participants (training on one and making predictions on the other), rather than simply within, in addition to a range of additional challenges (Poldrack et al., 2011). Although, results appear to be promising (see Poldrack & Yarkoni, 2016 for overview).

Cognitive Atlas Programme Poldrack has also argued that the reason for poor selectivity (and thus inconclusive reverse inferences) may be due to the possibility that cognitive scientists have not carved the mind into the appropriate ‘pieces’. In this case Poldrack (2006; see also Poldrack & Yarkoni, 2016) prescribe that cognitive psychologists co-operate to form something akin to the Gene Consortium (Ashburner et al., 2000), that is, by each inserting annotated information regarding the particular nature of the task, the particular cognitive process, and it’s respective constituents deemed to be involved, forming a cognitive ontology (i.e. the Cognitive Atlas Programme  )  The ambition is to amalgamate all such entries in order to generate a nodal network, specifying the full range of cognitive tasks (e.g. the sternberg task, Poldrack & Yarkoni, 2016) previously employed as well as the full range of process (e.g. working memory), and sub-processes (visuospatial sketchpad) which deemed to be elicited from such tasks (Poldrack & Yarkoni, 2016). This would then serve as a more appropriate target of inference, when observing such neural signatures as discussed above (Poldrack & Yarkoni, 2016; Poldrack, 2006).

CONCLUSION  -  100 --> In conclusion, reverse inference is a practice within cognitive neruscience which holds tremendous potential in the midst of forming a greater understanding of the relation between brain and cognition. However, with such promises, come numerous risks. In light of such inferential problems, a host of exciting developments have followed, draw upon large scale synthesis of neuroscientific data, as well as sophisticated pattern recognition techniques, serving as an ever improving accomplice to the scientist in their effort to draw a precise reverse inference.