• News
  • Spirituality
    • Dream Interpretation
    • Angel Numbers
    • Tarot
    • Prayers
    • Spells
  • Medicine
  • Science
  • Celebs
  • Betting

AI Build AI - Developing AI Models To Build AI

19.6KShares
261.4KViews

AI Build AI - Machine Learning is a subset of Artificial Intelligence in which a machine learns from previous experience, i.e. data. Unlike traditional programming, which requires the developer to foresee and write every possible scenario, a machine learning solution changes the output based on the data. A Machine Learning algorithm does not write code directly but constructs a computer model of the environment, which it then updates based on how it is educated.

We have worked hard to create machine intelligence. Maybe we should have let them get on with it. The majority of artificial intelligence is a numbers game. Deep neural networks, a type of AI that learns to recognize patterns in data, began outperforming standard algorithms 10 years ago because we finally had enough data and processing capacity to utilize them fully. Today's neural networks are even more data and power-hungry. Training them necessitates fine-tuning the values of millions, if not billions, of parameters that define these networks and indicate the strength of connections between artificial neurons. The objective is to obtain near-ideal settings for them, a process called optimization, but training the networks to get there is difficult. "Training might take days, weeks, or even months," said Petar Velickovic, a DeepMind staff research scientist in London. That might soon change.

AutoML By Google

The invention of AutoML, an artificial intelligence (AI) capable of producing its artificial intelligence, was disclosed in May 2017 by Google Brain researchers. Google scientists used a method known as reinforcement learning to automate the building of machine learning models for use in artificial intelligence. AutoML functions as a neural controller network that generates a child artificial intelligence network tailored to a given purpose. The aim for this specific kid artificial intelligence, which the researchers dubbed NASNet, was to recognise items in a video in real-time, such as people, automobiles, traffic lights, purses, backpacks, etc. According to the researchers, researchers found that NASNet was 82.7% accurate in predicting pictures on ImageNet's validation set.

Can An AI Create An AI?

COPYRIGHT_OAPL: Published on https://www.oapublishinglondon.com/hh/ai-build-ai/ by Alexander McCaslin on 2022-05-13T07:08:12.253Z

Boris Knyazev, a professor at the University of Guelph, and his colleagues created and trained a "hypernetwork," a form of overlord for other neural networks and can speed up the training process. When given a new, untrained deep neural network built for a specific purpose, the hypernetwork predicts its parameters in fractions of a second, potentially eliminating the requirement for entirely training the deep neural network. This is because the hyper network has been trained to recognise highly subtle patterns in deep neural network design, and this study may also have more enormous theoretical implications. For the time being, the hypernetwork performs well in some situations, but there is still room for improvement – which is only natural given the scale of the problem. If they succeed in addressing it, "this will have a tremendous impact across the board for machine learning," according to Velikovi.

Hypernetwork GHN-2

Ren, his former University of Toronto colleague Chris Zhang, and their mentor Raquel Urtasun took a new strategy in 2018. Given a series of potential architectures, they created a graph hypernetwork (GHN) to discover the optimum deep neural network architecture to accomplish a job. When Knyazev and his colleagues discovered the graph hypernetwork concept, they recognized they could expand on it. The team demonstrates how to utilize GHNs to determine the best architecture from a set of samples and forecast the parameters for the best network. It performs well in an absolute sense in their new study. In cases where the best is insufficient, the network can be trained further via gradient descent.

Knyazev and his colleagues dubbed their hyper network GHN-2, improving on two critical characteristics of Ren and colleagues' graph hyper network. First, they relied on Ren's representation of a neural network's architecture as a graph. Each node in the graph represents a subset of neurons that do a particular computation. The graph's edges show how information moves from node to node, from input to output.

The process of training the hyper network to produce predictions for new candidate designs was the second notion they drew on. This necessitates the use of two more neural networks. The first allows computations on the original candidate graph, resulting in changes to information associated with each node, while the second takes the updated nodes as input and predicts the parameters for the candidate neural network's corresponding computational units. These two networks each have their own set of parameters that must be improved before the hyper network can forecast parameter values appropriately.

However, Velikovi raises a potentially significant issue if hypernetworks like GHN-2 become the standard way for improving neural networks. "You have a neural network — effectively a black box — anticipating the parameters of another neural network," he explained of graph hypernetworks. As a result, if it makes a mistake, you have no method of explaining it. Deep neural networks typically detect patterns in photos, text, or audio signals, all of which are very organized sources of information. GHN-2 searches for patterns in the graphs of entirely random neural network topologies. That's a lot of information. GHN-2, on the other hand, can generalize, which means it can generate good predictions of parameters for unknown and even out-of-distribution network structures. This work demonstrates that many patterns are similar in many architectures and that a model may learn how to transfer information from one architecture to another.

Share: Twitter | Facebook | Linkedin

About The Authors

Alexander McCaslin

Alexander McCaslin

Recent Articles

No articles found.