The power of ‘cooling gas entropy’ for the development of deep learning
An Indian research team has developed a new algorithm for learning deep neural networks using cooling gas.
The team, led by Prof. Anurag Tiwari, a professor at the Tata Institute of Fundamental Research (TIFR), has also made a novel method for finding the optimal solutions to problem solving problems in deep learning.
Tiwar said that he used cooling gas to create an optimised neural network algorithm that had the power to learn a problem in only three months.
The deep learning algorithm that the team used was based on the concept of “dynamical entropy” or the difference between the “cooling” and “envelope” state of the neural network.
In this model, a neuron has a high degree of entropy and a lower degree of thermal entropy.
The difference between these states is called “coolness”.
“In the original research, we used the same kind of cooling gas, but we chose to use it in a novel way.
We were looking for the optimum solution of a problem and this was a good opportunity to explore a novel approach,” he told the India Today news channel.
The research team, which included Tiwary and Prof. Sudhir Prakash, an expert in artificial intelligence from Tata Institute for Fundamental Research, applied a novel algorithm to the design of an optimiser to find the best solution to a deep neural network problem.
They found that the optimiser algorithm could be used to solve problems in 10 to 15 days.
“When solving a problem, we need to keep an eye on a lot of parameters that affect the performance of the system.
In particular, we have to keep in mind that the number of parameters is not constant.
If you have a large number of inputs, the system can take longer to learn.
However, in the case of deep neural nets, if the number is not large, the learning rate will be very low.
We wanted to create a machine that could learn in a very short time,” said Tiwaria.
The cooling gas used in the algorithm is called the H 2 gas.
“The H 2 is a gas which absorbs and reflects thermal energy.
As it absorbs energy, it has a temperature and a phase transition between two states, called ‘halo states’,” explained Tiwario.
“It has an energy density of only 1.8 keV per mole of H 2 .
When a neuron is placed inside the H gas, the temperature and phase transitions become very small and the cooling gas acts as a cooling membrane,” he added.
“In other words, when a neuron was placed inside this cooling gas the temperature of the H molecule would be low, but the phase transition would become very high.
The H gas acts like a cooling device that absorbs energy from the incoming energy and returns it to the neuron,” said the professor.
The new algorithm has been validated by an experiment.
“We have seen a 100 per cent learning rate with a small amount of training data,” said Prof. Prakasu.
The researchers believe that the cooling mechanism can be used in many other applications.
“A cooling gas has been proven to be efficient for thermal evaporation of CO 2 into water, for example, in a water treatment plant or for the cooling of electrical grids,” said Dr. Senthil Ravi, an assistant professor at TIFR.
TIFRC is a leading research institute in the field of artificial intelligence and machine learning.
He added that there are many different types of cooling gases available.
“There are gas solutions that can be made from natural gas, such as kerosene, propane and even ethane.
We are also looking at ways of reducing the energy loss in these solutions,” said Ravi.
“For example, we are working on reducing the thermal loss of the gas when the cooling solution is applied to a gas turbine, to reduce the energy required for generating electricity.
We also want to create cooling solutions for other kinds of energy-saving applications such as electricity generation, heating and cooling systems.”
The research group has now developed the cooling agent to be used for cooling neural networks.
“By using this agent, we can create cooling gases which are both stable and transparent.
They will not become trapped in a membrane when they are used.
The agent will act as a passive radiator, keeping the temperature constant in the network,” he said.
In order to reduce heat dissipation, the cooling gases are placed in a container.
“Cooling gas solutions are also available in a wide variety of materials, including glass, ceramics, metal and plastics.
We want to focus on developing cooling gas solutions for the applications where they are needed most,” said Prakas.