-
StyleFusion: A Generative Model for Disentangling Spatial Segments
Authors:
Omer Kafri,
Or Patashnik,
Yuval Alaluf,
Daniel Cohen-Or
Abstract:
We present StyleFusion, a new mapping architecture for StyleGAN, which takes as input a number of latent codes and fuses them into a single style code. Inserting the resulting style code into a pre-trained StyleGAN generator results in a single harmonized image in which each semantic region is controlled by one of the input latent codes. Effectively, StyleFusion yields a disentangled representatio…
▽ More
We present StyleFusion, a new mapping architecture for StyleGAN, which takes as input a number of latent codes and fuses them into a single style code. Inserting the resulting style code into a pre-trained StyleGAN generator results in a single harmonized image in which each semantic region is controlled by one of the input latent codes. Effectively, StyleFusion yields a disentangled representation of the image, providing fine-grained control over each region of the generated image. Moreover, to help facilitate global control over the generated image, a special input latent code is incorporated into the fused representation. StyleFusion operates in a hierarchical manner, where each level is tasked with learning to disentangle a pair of image regions (e.g., the car body and wheels). The resulting learned disentanglement allows one to modify both local, fine-grained semantics (e.g., facial features) as well as more global features (e.g., pose and background), providing improved flexibility in the synthesis process. As a natural extension, StyleFusion enables one to perform semantically-aware cross-image mixing of regions that are not necessarily aligned. Finally, we demonstrate how StyleFusion can be paired with existing editing techniques to more faithfully constrain the edit to the user's region of interest.
△ Less
Submitted 15 July, 2021;
originally announced July 2021.
-
Information Theoretic Approach to Social Networks
Authors:
Oded Kafri
Abstract:
We propose an information theoretic model for sociological networks. The model is a microcanonical ensemble of states and particles. The states are the possible pairs of nodes (i.e. people, sites and alike) which exchange information. The particles are the energetic information bits. With analogy to bosons gas, we define for these networks model: entropy, volume, pressure and temperature. We show…
▽ More
We propose an information theoretic model for sociological networks. The model is a microcanonical ensemble of states and particles. The states are the possible pairs of nodes (i.e. people, sites and alike) which exchange information. The particles are the energetic information bits. With analogy to bosons gas, we define for these networks model: entropy, volume, pressure and temperature. We show that these definitions are consistent with Carnot efficiency (the second law) and ideal gas law. Therefore, if we have two large networks: hot and cold having temperatures TH and TC and we remove Q energetic bits from the hot network to the cold network we can save W profit bits. The profit will be calculated from W equal or smaller than Q (1-TH/TC), namely, Carnot formula. In addition it is shown that when two of these networks are merged the entropy increases. This explains the tendency of economic and social networks to merge.
△ Less
Submitted 30 July, 2014;
originally announced July 2014.
-
The Distributions in Nature and Entropy Principle
Authors:
Oded Kafri
Abstract:
The derivation of the maximum entropy distribution of particles in boxes yields two kinds of distributions: a "bell-like" distribution and a long-tail distribution. The first one is obtained when the ratio between particles and boxes is low, and the second one - when the ratio is high. The obtained long tail distribution yields correctly the empirical Zipf law, Pareto's 20:80 rule and Benford's…
▽ More
The derivation of the maximum entropy distribution of particles in boxes yields two kinds of distributions: a "bell-like" distribution and a long-tail distribution. The first one is obtained when the ratio between particles and boxes is low, and the second one - when the ratio is high. The obtained long tail distribution yields correctly the empirical Zipf law, Pareto's 20:80 rule and Benford's law. Therefore, it is concluded that the long tail and the "bell-like" distributions are outcomes of the tendency of statistical systems to maximize entropy.
△ Less
Submitted 28 July, 2009;
originally announced July 2009.
-
A Comment on Nonextensive Statistical Mechanics
Authors:
Oded Kafri
Abstract:
There is a conception that Boltzmann-Gibbs statistics cannot yield the long tail distribution. This is the justification for the intensive research of nonextensive entropies (i.e. Tsallis entropy and others). Here the error that caused this misconception is explained and it is shown that a long tail distribution exists in equilibrium thermodynamics for more than a century.
There is a conception that Boltzmann-Gibbs statistics cannot yield the long tail distribution. This is the justification for the intensive research of nonextensive entropies (i.e. Tsallis entropy and others). Here the error that caused this misconception is explained and it is shown that a long tail distribution exists in equilibrium thermodynamics for more than a century.
△ Less
Submitted 30 June, 2009; v1 submitted 28 June, 2009;
originally announced June 2009.
-
Entropy Principle in Direct Derivation of Benford's Law
Authors:
Oded Kafri
Abstract:
The uneven distribution of digits in numerical data, known as Benford's law, was discovered in 1881. Since then, this law has been shown to be correct in copious numerical data relating to economics, physics and even prime numbers. Although it attracts considerable attention, there is no a priori probabilistic criterion when a data set should or should not obey the law. Here a general criterion…
▽ More
The uneven distribution of digits in numerical data, known as Benford's law, was discovered in 1881. Since then, this law has been shown to be correct in copious numerical data relating to economics, physics and even prime numbers. Although it attracts considerable attention, there is no a priori probabilistic criterion when a data set should or should not obey the law. Here a general criterion is suggested, namely that any file of digits in the Shannon limit (namely, having maximum entropy) has a Benford's law distribution of digits.
△ Less
Submitted 8 March, 2009; v1 submitted 20 January, 2009;
originally announced January 2009.
-
Sociological Inequality and the Second Law
Authors:
Oded Kafri
Abstract:
There are two fair ways to distribute particles in boxes. The first way is to divide the particles equally between the boxes. The second way, which is calculated here, is to score fairly the particles between the boxes. The obtained power law distribution function yields an uneven distribution of particles in boxes. It is shown that the obtained distribution fits well to sociological phenomena,…
▽ More
There are two fair ways to distribute particles in boxes. The first way is to divide the particles equally between the boxes. The second way, which is calculated here, is to score fairly the particles between the boxes. The obtained power law distribution function yields an uneven distribution of particles in boxes. It is shown that the obtained distribution fits well to sociological phenomena, such as the distribution of votes in polls and the distribution of wealth and Benford's law.
△ Less
Submitted 9 February, 2009; v1 submitted 21 May, 2008;
originally announced May 2008.
-
The Second Law as a Cause of the Evolution
Authors:
Oded Kafri
Abstract:
It is a common belief that in any environment where life is possible, life will be generated. Here it is suggested that the cause for a spontaneous generation of complex systems is probability driven processes. Based on equilibrium thermodynamics, it is argued that in low occupation number statistical systems, the second law of thermodynamics yields an increase of thermal entropy and a canonic e…
▽ More
It is a common belief that in any environment where life is possible, life will be generated. Here it is suggested that the cause for a spontaneous generation of complex systems is probability driven processes. Based on equilibrium thermodynamics, it is argued that in low occupation number statistical systems, the second law of thermodynamics yields an increase of thermal entropy and a canonic energy distribution. However, in high occupation number statistical systems, the same law for the same reasons yields an increase of information and a Benford's law/power-law energy distribution. It is therefore, plausible, that eventually the heat death is not necessarily the end of the universe.
△ Less
Submitted 28 November, 2007;
originally announced November 2007.
-
Informatics Carnot Machine
Authors:
Oded Kafri
Abstract:
Based on Planck's blackbody equation it is argued that a single mode light pulse, with a large number of photons, carries one entropy unit. Similarly, an empty radiation mode carries no entropy. In this case, the calculated entropy that a coded sequence of light pulses is carrying is simply the Gibbs mixing entropy, which is identical to the logical Shannon information. This approach is supporte…
▽ More
Based on Planck's blackbody equation it is argued that a single mode light pulse, with a large number of photons, carries one entropy unit. Similarly, an empty radiation mode carries no entropy. In this case, the calculated entropy that a coded sequence of light pulses is carrying is simply the Gibbs mixing entropy, which is identical to the logical Shannon information. This approach is supported by a demonstration that information transmission and amplification, by a sequence of light pulses in an optical fiber, is a classic Carnot machine comprising of two isothermals and two adiabatic. Therefore it is concluded that entropy under certain conditions is information.
△ Less
Submitted 17 May, 2007;
originally announced May 2007.
-
The Second Law and Informatics
Authors:
Oded Kafri
Abstract:
A unification of thermodynamics and information theory is proposed. It is argued that similarly to the randomness due to collisions in thermal systems, the quenched randomness that exists in data files in informatics systems contributes to entropy. Therefore, it is possible to define equilibrium and to calculate temperature for informatics systems. The obtained temperature yields correctly the S…
▽ More
A unification of thermodynamics and information theory is proposed. It is argued that similarly to the randomness due to collisions in thermal systems, the quenched randomness that exists in data files in informatics systems contributes to entropy. Therefore, it is possible to define equilibrium and to calculate temperature for informatics systems. The obtained temperature yields correctly the Shannon information balance in informatics systems and is consistent with the Clausius inequality and the Carnot cycle.
△ Less
Submitted 18 April, 2007; v1 submitted 3 January, 2007;
originally announced January 2007.
-
Information theory and Thermodynamics
Authors:
Oded Kafri
Abstract:
A communication theory for a transmitter broadcasting to many receivers is presented. In this case energetic considerations cannot be neglected as in Shannon theory. It is shown that, when energy is assigned to the information bit, information theory complies with classical thermodynamic and is part of it. To provide a thermodynamic theory of communication it is necessary to define equilibrium f…
▽ More
A communication theory for a transmitter broadcasting to many receivers is presented. In this case energetic considerations cannot be neglected as in Shannon theory. It is shown that, when energy is assigned to the information bit, information theory complies with classical thermodynamic and is part of it. To provide a thermodynamic theory of communication it is necessary to define equilibrium for informatics systems that are not in thermal equilibrium and to calculate temperature, heat, and entropy with accordance to Clausius inequality. It is shown that for a binary file the temperature is proportional to the bit energy and that information is thermodynamic entropy. Equilibrium exists in random files that cannot be compressed. Thermodynamic bounds on the computing power of a physical device, and the maximum information that an antenna can broadcast are calculated.
△ Less
Submitted 7 February, 2006;
originally announced February 2006.
