advancements of neural networks and how it can impact the automotive industry

advancements of neural networks and how it can impact the automotive industry

 Literature Review

“Investigating the advantages of using machine learning to develop a GAN model which constructs an image of a car and exploring the impact on design in the future.”

Literature review topic

“Explore the advancements of neural networks and how it can impact the automotive industry”

Introduction

Generative adversarial networks are an area of machine learning that uses input data to create examples of whatever dataset you use to teach it. It consists of two algorithms, discriminative and generative, which work towards producing a photorealistic product without supervision in an actor critic method. The discriminative network is used to correct the data that the generative network synthesises until the network synthesises an output more like the expected outputs based on the training set data that the network is fed. The forefather of generative adversarial networks, (Ian Goodfellow 2014) uses examples from the human face database to generate profile images of humans that don’t exist. For my dissertation I will use a similar framework to create a model which will enable the synthesis of an automotive design from a noise vector. However due to the nature of the concept being in its early stages i have decided to specifically focus on the impact that neural networks is having in industry and more explicitly the automotive industry. Neural networks as a concept has been around since the 19^th century, however with the advancement of technology and artificial intelligence developers have been experimenting with ways that they can introduce these methods into different fields of work. The purpose of my dissertation is to investigate the uses of neural networks, from this I intend on creating software that can produce an image of a car. The goal is to see how the GANs manifest their own perception of car into a unique design, drawing from the vast number of images that it has learned. From this we will be able to analyse the results and conclude whether this type of technology could be at the frontiers of design for vehicle manufacturers.  Exploring the technology of neural networks will allow me to understand how I should create my artefact and will also help me with which methods I should use to teach my machine in order to get accurate results. Furthermore my findings will help me answer some of my aims and objectives that I identified early on. My literature review will allow me to develop and reinforce many key concepts and ideas in the field of artificial intelligence by providing myself with relevant resources that I find in articles and books.

What makes neural networks so significant to machine learning is the plethora of divergent algorithms that have been developed, each one being applied in different specified situations. For example, the Boltzmann machine neural network architecture is used to develop algorithms for deep belief networks; these networks enable the accurate analysis and recognition of videos and motion capture data. Whereas perceptrons are single layered neural networks that can…

Understanding which algorithm would best suit my artefact design is vital in the production stage as there are many different algorithm designs that may offer the same results in theory, but the accuracy of the outcome may differ.

The concept of neural networks has unlimited potential in industry and its influence can be seen in many important fields such as healthcare or the stock market. Neural networks can succeed the deliberation phase that humans go through, this allows for an increase in accuracy and speed as the technology doesn’t have bias or opinions, it works on the data it is given. A recent study by (Robert Geirhos 2018) sees how accurately object recognition with a weakened signal can be executed with deep neural networks. This is then compared to a human’s competency in image recognition. The images are subject to different types of degredations, eidolons, contrast and noise. In most of the experiments carried out, the human observers seem to outperform the three deep neural networks that are tested (VGG-16, GoogLeNet’s, AlexNet’s)e.g the noise-experiment. “by increasing the noise width from 0.0 (no noise) to 0.1, VGG-16’s performance drops from an accuracy of 89.91% to 44.02%; GoogLeNet’s drops from 81.70% to 34.02% and AlexNet’s from 70.00% to 19.29%. Human observers, on the other hand, only drop from 80.50% to 75.13%”. However, in the contrast experiment for accuracy, VGG-16 reflects results that are like the human participants at 91% – 94%. It also manages to drop accuracy slower than the other two neural networks. My project is based around generative adversarial networks which rely on a framework of noise to develop an image out of it, however, theoretically my artefact focuses on design and developing the design of cars. Using neural networks in this format would be considerably quicker than having a team of designers work on a frame from scratch. Furthermore, the accuracy of the final image generated by the GAN is not entirely dependent on the input image, the training data set is what the discriminator network uses to deliver a realistic product.

Images source – https://arxiv.org/pdf/1706.06969.pdf

Evolution of neural networks

Tavanaei, Amirhossein et al. (2019) Deep learning in spiking neural networks. Neural Networks.

The premise of this article written in 2019 is to analyse machine learning of artificial neural networks by using backpropagation. To get the best out of neural networks the scientists get inspiration from the brain and the way information is learnt at a faster rate through spikes. “In a biological neuron, a spike is generated when the running sum of changes in the membrane potential, which can result from presynaptic stimulation, crosses a threshold. The rate of spike generation and the temporal pattern of spike trains carry information about external stimuli (Gerstner & Kistler, 2002; Rieke, 1999) and ongoing calculations’. From this we can learn that rather then it working from continuous values, the spikes that are generated from when a neuron gets to a specific time or level can help when it comes to dealing with the processing of specific real world data. The publication thoroughly goes through the results of supervised and unsupervised machine learning whilst also drawing comparisons between the tests. They tested the models against each other with different learning methods but the same datasets to observe the varying classification abilities. Spiking neural networks are useful because of how closely it is linked to computational neuroscience. Furthermore because of how similar it is to the human brain SNN’s should be able to execute jobs which require faster methods of learning, although when compared with ANN’s the output is relatively alike. Artificial neural networks are used predominantly for the discrimination of data whereas generative adversarial networks are used to generate data, because of this it has more benefits for aspects of design which is what it’s purpose will be in my artefact.

Valle, Rafael (2019) Hands-on generative adversarial networks with Keras : your guide to implementing next-generation generative adversarial networks . Birmingham;: Packt Publishing Ltd.

Geirhos, R., Janssen, D.H., Schütt, H.H., Rauber, J., Bethge, M. and Wichmann, F.A., 2017. Comparing deep neural networks against humans: object recognition when the signal gets weaker. arXiv preprint arXiv:1706.06969.

https://www.sciencedirect.com/science/article/pii/S1474667017425146