Artificial Intelligence

2212 Submissions

[5] viXra:2212.0212 [pdf] submitted on 2022-12-29 04:53:14

Beyond Rewards and Values: a Non-Dualistic Approach to Universal Intelligence

Authors: Akira Pyinya
Comments: 14 Pages.

Building an AI system that aligns with human values is believed to be a two-step process: first design a value function or learn human value using value learning methods, then maximize those values using rational agents such as AIXI agents. In order to integrate this into one step, we analyze the dualistic assumptions of AIXI, and define a new universal intelligence model that can align with human preferences or specific environments, called Algorithmic Common Intelligence (ACI), which can behave the same way as examples. ACI does not have to employ rewards or value functions, but directly learns and updates hypothetical policies from experience using Solomonoff induction, while making actions according to the probability of every hypothesis. We argue that the rational agency model is a subset of ACI, and the coevolution of ACI and humans provides a pathway to AI alignment.
Category: Artificial Intelligence

[4] viXra:2212.0208 [pdf] submitted on 2022-12-30 03:47:42

Design Autoencoder using BSnet (BSautonet)

Authors: Sing Kuang Tan
Comments: 3 Pages.

In this paper, I am going to propose a design for an Autoencoder using BSnet. To take advantage of the BSnet design, the autoencoder will be easy to train with more convex training optimization function. The idea is to develop a simple and standard unsupervised machine learning model that can easily be used on most of the data without label. In the experiment result, the output is subjectively evaluated by a human and it has shown to achieve human level accuracy on denoising the MNIST human handwriting digits dataset.
Category: Artificial Intelligence

[3] viXra:2212.0193 [pdf] submitted on 2022-12-27 00:22:31

Boolean Structured Deep Learning Network (BSnet)

Authors: Sing Kuang Tan
Comments: 5 Pages.

In this paper, I am going to propose a new Boolean Structured Deep Learning Network (BSnet) based on the concept of monotone multi-layer Boolean algebra. I have shown that this network has achieved significant improvement in accuracy over an ordinary Relu Deep Learning Network.
Category: Artificial Intelligence

[2] viXra:2212.0176 [pdf] replaced on 2023-02-14 09:34:24

Efficient Integration of Perceptual VAE into Dynamic Latent Scale GAN

Authors: Jeongik Cho
Comments: 10 Pages.

Dynamic latent scale GAN is a method to train an encoder that inverts the generator of GAN with maximum likelihood estimation. In this paper, we propose a method to improve the performance of dynamic latent scale GAN by integrating perceptual VAE loss into dynamic latent scale GAN efficiently. When training dynamic latent scale GAN with normal i.i.d. latent random variable, and latent encoder is integrated into discriminator, a sum of a predicted latent random variable of real data and a scaled normal noise follows normal i.i.d. random variable. This random variable can be used for both VAE and GAN training. Considering the intermediate layer output of the discriminator as a feature encoder output, the generator can be trained to minimize perceptual VAE loss. Also, inference & backpropagation for perceptual VAE loss can be integrated into those for GAN training. Therefore, perceptual VAE training does not require additional computation. Also, the proposed method does not require prior loss or variance estimation like VAE.
Category: Artificial Intelligence

[1] viXra:2212.0163 [pdf] submitted on 2022-12-22 03:23:02

The SP-multiple-alignment Concept as a Generalisation of Six Other Variants of "Information Compression via the Matching and Unification of Patterns"

Authors: J. G. Wolff
Comments: 23 Pages.

This paper focusses on the powerful concept of SP-multiple-alignment, a key part of the SP System (SPS), meaning the SP Theory of Intelligenceand its realisation in the SP Computer Model. The SPS is outlined in an appendix. More specifically, the paper shows with examples how the SP-multiplealignment construct may function as a generalisation of six other variants of ‘Information Compression via the Matching and Unification of Patterns’ (ICMUP). Each of those six variants is described in a separate section, and in each case there is a demonstration of how that variant may be modeled via the SP-multiple-alignment construct.
Category: Artificial Intelligence