What is Deep Learning?
Deep learning is one of the modern computer science fields thriving with research and application such as never imagined. A subset of artificial intelligence based on neural networks and backtracking algorithms, Deep Learning can be a proper interest of choice. Deep learning generally works on multiple layers of neural networks to get maximum accuracy and data entities. Single-layer of a neural network cannot be said bad as there is always past machine learning experience used in it. Millions of studies and historical data can be used to get a computer-based judgment. Mathematically it requires knowledge of Propositional logic, probability distributions, and latent variables organized with layer-wise, deep generative models, for example, a deep Boltzmann machine (a deep learning machine model with more hidden layers with directionless connections between the nodes).
How does it work?
In deep learning, each level learns to transform its input data into a slightly varied, more abstract, and composite representation. One of the applications is computerized vision. If taken as an example for it, raw input is taken in the form of a matrix of pixels. The first layer abstracts pixel and encoded edges, the second layer may represent the arrangement of edges, and finally third level (assumed three neural network layers) may identify unique shapes such as the nose, eyes, etc. The word deep identifies how deep a system is; this system is a deep: 3 system. Scientifically can be called a CAP number (credit assignment path).
Every layer works on greedy algorithms. Deep System tries to interpret based on performance improvement factors. It uses probabilistic inference rules and expectations based on historical data.
Future: Self-Supervised Deep Learning
Deep learning relies on data and computational power heavily today. We need to replace supervised learning- the training method of most current DL systems. It is an idea to learn about the world before learning a task. This is exactly what babies and animals do; after having good impressions of the world, learning a task requires a few trials and improvisations.
Instead of training a system with labeled data, the system will learn by itself (by making that data).
Humans learn a lot faster than deep systems because we go for observation more rather than needing thousands of samples. If a DL system needs to identify a tree, it needs to see thousands of images of trees, but humans do not need such extensive data learning. Humans strive on observation and intuition. This is why they learn faster but are not purely accurate.
CHESS and Deep Learning!
Chess is a game that originated in India, the national game of Russia was at its peak in 1996 when Vishwanathan Anand was preparing against the world champion Garry Kasparov (Russia) after a defeat in the 1995 world championship match. A Carnegie Mellon University researcher “Feng-Hsu” made a super-computer under the name “Chip test.” It won the North Americal Chess championship in 1987; later, after completing the doctorate, he and his team joined IBM and made a system with deep learning named “Deep-Blue.”
Testing of this system was crucial, and who could give it the most resistance from the human world, of course, a world champion. Garry accepted IBM’s invitation and played with the system in 1996, defeating the machine by a 4-2 score; later, in 1997 machine defeated Garry by three wins and one draw. The first win of Deep Blue over Garry is still considered the overpowering of computers over humans in chess or deep learning.
Chess engines Later, Vishwanathan Anand himself prepared with the computer(with deep systems) and became a five-time world champion. At today’s date, Stockfish and alpha-zero (developed by Google) are the most advanced chess Engines.
Game of Shadows:
Chess engines require high computational power. The Best deep learning mechanism will need super-computers to deduce the best combination of moves to win a specific position in chess. Nowadays, this game is so much influenced by these systems that whoever holds the best machine is considered a threat to the opponent.
It is said that in chess, after only four moves, there are more than 300 billion options to consider. Games with 40+ moves are more in numbers than the stars in all the galaxies. It takes to the very edge of the human mind, so much memorization of best games, openings, and practicing well-defined endgames. It is true for computers as well. The Stockfish learns the knowledge of positions after analyzing millions of data-structures nodes at some depth. It can look to the end of the game after one move only. Some computers take a whole night to identify which move is best; some can tell in 1 minute only, depending on the type of system used. These deep-learning systems have a lot of theories developed and used in this game and changed this game entirely. Playing this game without the aid of deep-learning systems is not possible anymore at the Grand-Master level.
Stockfish Engine and Difference with Humans:
Look at the following position held between current world champion Magnus Carlsen and Vishy Anand at Norway Chess (5th June 2022). It is white to move and is in a completely dominant position, but the next move suggested by Stockfish (11.0) is the Qg4. It is a very unhuman move to consider. If Qg4 is not playing right now, any other move goes towards drawing position almost. So, the difference between a chess engine and a human is based on intuition and calculation. Human relies on calculation and bit touch of intuition. Intuition sometimes may go wrong, but computers never go wrong because there is nothing like intuition in them; systems are purely computational and learned.
Leading software technologies to build deep learning systems:
As discussed above, Stockfish (11.0) is the most dominant deep learning system in chess. However, there are several more like Sesse, Leela, etc. We can build our own systems with the following technologies as well.
1. Neural Designer:
It is an application used for data mining with the help of neural networks on which deep learning systems are built. Tons of data can be analyzed and can help to examine patterns.
Open-source, fast, scalable machine learning API. It was written in JAVA from scratch and easy to integrate with Apache Hadoop and Spark to give customers the flexibility to resolve the most challenging data problems. It is useful in a random forest, Generalized linear modeling, K-means, etc.
A deep learning library for Theano and TensorFlow. It is written in Python with a focus on fast experimentation. Being able to see the results from ideas in a quick time is the best motivation for good research. It allows for fast prototyping of the system. It supports convolutional networks, recurrent networks, and a hybrid combination.
4. Caffe and Torch:
Both are built on the C/CUDA framework and are used for neural networks and energy-based models. Torch is widely accepted, as it provides algebra routines also. Caffe is a Ph.D. project by Yangqing Jia at UC Berkeley. 5. Microsoft Cognitive ToolKit & DeepLearningKit: DeepLearningKit is built by Apple to build deep learning systems for Mac OS, TV OS, IOS, etc. Microsoft Cognitive Toolkit is used for training and hosting with azure.
Applications of Deep Learning.
1. Virtual Assistants: These are cloud-based applications that run on human voice commands. Amazon Alexa, Cortana, Siri, and Google Assistant are famous examples.
2. Health Care: These deep systems are very useful nowadays for computer-aided disease detection, drug discovery, and diagnosis of life-threatening diseases such as cancer. In India, they used IISC situated Super-Computer to find the cure for COVID-19 to find the best combination of drugs to cure it.
3. Entertainment: Based on a person’s browsing history, interest, and behavior over the Internet, online streaming companies such as Netflix, Amazon Prime, etc., provide their filtered suggestions. Deep learning systems are also used to generate sounds for silent movies and generate subtitles.
4. Robotics: Deep learning-powered robotic systems sense any obstacles in the path and preplan their journey instantly. These can be useful to carry goods in factories, hospitals, warehouses, inventories, etc.
5. Chatbots: These are used to satisfy customers’ general needs or at least gather data to solve the problems. These deep learning systems are useful in Customer interaction, marketing on social networks, and instant messaging to clients after identifying their needs.
Last year at work, someone asked Siri to tell them who is their best friend. Siri gave the exact answer by analyzing which person that specified person was calling most. If it is not possible to analyze, it will ask for an answer and record it for future reference. Deep-learning systems are two to three-fold stronger than they were three years ago. Robotics and health care will take a lot of advantage in the coming years. Metaverse is also going to use the deep-learning mechanism. The Prospect of this branch is wide in AI, and it’s very promising to deliver products that are smart and never imagined before.
Here chess is taken as an example to see how evolved it is now before DL systems. It changed the entire game now. DL has so much potential and will help humans improve in many modern fields.