The Early History of Neural Representations

Kant defined representation as ‘inner determination of our mind in this or that relation of time’ and ‘perceptions’ are held to be conscious representations. Helmholtz held to this idea in the nineteenth century when he developed experimental psychology whereas the British clung to the ancient concepts of ideas and impressions. The expression ‘neural representation’ rapidly grew in use in the latter half of the twentieth century. What does it mean and how did this come about? Its origins can be traced to Charles Wheatstone, who at the beginning of the nineteenth century introduced the term representations while investigated the phenomenon of stereopsis. He stated that “the most complex figures of three dimensions may be accurately represented to the mind by representing their—two perspective projections to the two retinas”. It is taken for-granted that it is the mind that perceives objects in a human being’s field of vision. Helmholtz later says “our purpose is to identify the matter of sensation which occasions the formation of representations”—so adopting the Kantian view, which we have previously criticised. Fritz and Hitzig (On the electrical excitability of the cerebrum. In Some Papers on the Cerebral Cortex. Illinois: Charles C. Thomas, 1870) stimulated the cortex now known as ‘motor’ and obtained contractions of muscles on the opposite side to that of the stimulated cortex. Beevor and Hurley (A minute analysis (experimental) of the various movements produced by stimulating in the monkey different regions of the cortical centre for the upper limb, as defined by Professor Ferrier. Philos. Trans. R. Soc. Lond. B 153–167, 1887) showed an ‘extensive representation of movement’ in motor cortex. Subsequently Grunbaum and Sherrington (Observations on the Physiology of the Cerebral Cortex of Some of the Higher Apes. Proceedings of the Royal Society of London, Vol. 69, pp. 206–209, 1901–1902) showed ‘the arrangement of the representation of the various regions of the musculature follow the segmental sequelae of the cranio-spinal nerve to a remarkable extent. These various uses of representation amount to little more than ‘causal correlates’. The concept of ‘deep learning’ using neural networks with what they called the ‘back-propagation algorithm’, was due to Rumelhart and colleagues. It was first applied to the association parietal cortex by Zipser and Andersen who suggested that neurons functioned in this region of cortex that represented the transformation from retinal coordinates to those of eye-centered coordinates. The leader in ‘deep learning’ comments recently that “Deep-learning methods are representation-learning methods with multiple levels of representation, obtained by composing simple but non-linear modules that each transform the representation at one level (starting with the raw input) into a representation at a higher, slightly more abstract level”. We show that as far as humans are concerned this is conceptually flawed, both from the point of view of the Representational Fallacy enumerated above and from that of the Mereological Fallacy that we have spelt out in previous books.