Using An ANN (Artificial Neural Networks) In Neuroscience

The development of biologically inspired artificial neural networks (ANNs) provides new ways to gain a better understanding of the human brain. ANNs are computing systems that learn by examples to carry out tasks without being explicitly programmed. ANNs can consist a single or multiple layers and can be trained in several ways. Deep neural networks involves an input, an output and one or more hidden layers which makes the network more similar to the hierarchically organized brain. A frequently applied deep network architecture in analyzing visual data is the convolutional networks (CNN) due to its sparsely connected property (Yamins & DiCarlo, 2016, Shai & Larkum, 2017).

Using an ANN to access more insight of step by step sensory processing is appear to be the recent approach of neuroscience. However, such a network has to meet with some criteria to serve as a model of sensory cortex. First, the model should be able to operate with arbitrary stimuli. Second, the model should provide a correct prediction of the neural response. Third, the model must consist identifiable components corresponding to the particular cortical areas in the brain.

There are different approaches to map artificial neural networks to biological networks. Task information consistency, population representational similarity are both population – level metrics and the single – unit response predictivity which builds upon linear neural response predictivity of single unit. This approach uses each empirically measured neuron, which must be an approximately linear combinations of units of the brain, as the target of a linear regression from units in the model. A synthetic neuron, which is linear combinations of model units, should have the same response as the real neuron (Yamins, & DiCarlo, 2016).

Object recognition is a hierarchical process via the ventral visual stream, which extends from the visual cortex to the inferotemporal cortex (IT). The brain process sensory inputs through complex linear and nonlinear transformations. Therefore, appropriate artificial model needs to involve nonlinear neural motifs as well. A class of models called hierarchical convolutional network (CHNN) seem to be the potential option. In the multi - layer HCNNs the deep networks are composed in a way that can compute a complex transformation of the input data which process occur in the ventral stream.

In convolutional layers only a subset of units are connected to the next layer. The weights of some units are shared allowing the detection of features across the entire input layer which leads to decreased number of trainable weights. Layers in the HCNNs works with linear-nonlinear (NL) operations namely filtering, thresholding, pooling and normalization. To represent the synaptic strength of real neurons each layer contains a set of weights, filters. Filtering is a linear operation by which the dot product of the filter itself and a patch of input data are computed and stored, then the same takes place with another patch and continues until the entire image is covered.

Pooling is a non-linear aggregation operation and normalization refers to the process of adjusting the output values corresponding to a standard range. In HCNNs a complex nonlinear transformation of the original input is possible due to the structure of multiple layers. Additionally, HCNNs are mappable, able to deal with arbitrary stimuli and predictive, thus they have high potential as models of the ventral visual stream. However, the challenge is to choose the correct parameters for HCNNs that match best with the biological system (Yamins & DiCarlo, 2016).

A recent improvement is a performance – based approach called the goal driven HCNNs. These models are trained for one specific task such as category recognition with the use of millions of real world images in a large number of categories from the internet. Then they are tested on quite distinct visual tasks like synthetically created pictures including objects from other categories than those were used during the training. Thus, the model has to generalize which can be reached by supervised learning. It means that the strength of connections has to be adjusted in a way that results in the desired output. Additionally, model parameters first have to be optimized on a particular performance, then once these have been fixed, can be compared with neural data (Yamins & DiCarlo, 2016).

HCNN performed successfully multiple tasks such as image categorization, face identification or action recognition. Furthermore, HCNNs are the first models which were able to predict the neural response in V4 cortical areas. Several factors are contributing to these achievements. Supervised learning of a task means reducing the error between the output by the network and the correct response. Then parameter settings of a network are optimized by learning algorithms. In multi layer networks it requires backpropagation. Via backpropagation the contribution of each weight to the final error can be computed therefore it can change its value and update the network parameters by gradient descent.

Computation of error starts at the top layer then propagate backward and adjust the weights through the network down to the first layer. After it can continue with the subsequent training. A possible difficulty can occur when the network cannot generalize due to its capacities compare to the massive data set, overfitting. A recent development of graphical processing unit (GPU) can treat this issue better. Therefore can be used on goal driven HCNN.

Another technical advance which has led to the development of recent deep networks is the automated learning procedures for artificial parameters. While previously these were selected by hand, currently it has been replaced by Gaussian process optimization and genetic algorithms. Besides, the of half – rectified thresholds instead of sigmoid activation since it is less sensitive to the vanishing gradient problem, when error gradient become too small to optimize effectively (Yamins & DiCarlo, 2016).

Overall, the improvements of ANN can provide a considerable contribution to explain a complex neural system such as ventral the visual stream. Although, our understanding is far from complete regarding the learning process of deep networks. Nonetheless, the fact that HCNNs are capable of performing on a high level on complex nonlinear tasks shows a positive direction of discovering the human brain in more detail.