Cable theory for dendritic neurons
Methods in neuronal modeling
Compartmental models of complex neurons
Methods in neuronal modeling
The logic of activation functions
Parallel distributed processing: explorations in the microstructure of cognition, vol. 1
NMDA-based pattern discrimination in a modeled cortical neuron
Neural Computation
The handbook of brain theory and neural networks
Neural Networks for Pattern Recognition
Neural Networks for Pattern Recognition
Learning in networks of structured hypercubes
Learning in networks of structured hypercubes
| Hi-index | 0.00 |
In this paper and its companion, we address the question as to whether there are any general principles underlying information processing in the dendritic trees of biological neurons. In order to address this question, we make two assumptions. First, the key architectural feature of dendrites responsible for many of their information processing abilities is the existence of independent sub-units performing local non-linear processing. Second, any general functional principles operate at a level of abstraction in which neurons are modelled by Boolean functions. To accommodate these assumptions, we therefore define a Boolean model neuron--the multi-cube unit (MCU)--which instantiates the notion of the discrete functional sub-unit. We then use this model unit to explore two aspects of neural functionality: generalisation (in this paper) and processing complexity (in its companion). Generalisation is dealt with from a geometric viewpoint and is quantified using a new metric--the set of order parameters. These parameters are computed for threshold logic units (TLUs), a class of random Boolean functions, and MCUs. Our interpretation of the order parameters is consistent with our knowledge of generalisation in TLUs and with the lack of generalisation in randomly chosen functions. Crucially, the order parameters for MCUs imply that these functions possess a range of generalisation behaviour. We argue that this supports the general thesis that dendrites facilitate input pattern generalisation despite any local non-linear processing within functionally isolated sub-units.