Chainge Technology Salon — Chain / AI Blockchain + Artificial Intelligence
Chainge Technology Salon #2 — Bo Peng On Deep Convolutional Neural Networks
Bo Peng has more than 20 years of experience in R&D and is a technical specialist in AI, quantitative trading and blockchain. He is the founder of the company Blink and the author of “Deep Convolutional Neural Network: Principle and Practice.”
At the 11th Technology Salon “Chain” — Chain/ AI, Bo gave a talk named “The Opportunity and Challenges of Integrating Deep Learning and Blockchain”.
The event was held in Shanghai Jiaoda Huigu Training Centre on 16 September 2018. It was organized by Blockchain EXE × 8BTC together with TomatoWallet, Bytom and Cortex. “Chainge” is an offline event started by 8 BTC — the most influential independent platform for bitcoin, blockchain, and cryptocurrency news in China — to focus on blockchain’s hotspot discussion and communication between industries and companies. By hosting offline meetups, hackathons, blockchain summits, the organisers hope to help industry experts create high quality content, collaboration and promote the dissemination of knowledge and innovation in the field of blockchain.
From AI to Deep LearningBlockchain
Peng focused part of his presentation on Google’s Alpha Go’s impact on deep learning and AI. Alpha Go uses a Deep Convolutional Neural Network (DCNN) and is able to simulate how humans play Go, and through playing millions of games against itself is able to improve itself. Its success was cemented through beating a string of the best human Go players on the planet, including the world champion, illustrating how powerful DCNNs can become in a very short period of time. .
Peng Bo also showed an example of a face Generator Adversarial Network (GAN) where a model is given huge amount of photos of celebrities. The computer then automatically finds a suitable interpolation approach without the need for any human input. (The computer will, for example, automatically recognise that a human has two eyes, one nose, etc.)
In current posture recognition, there are still imperfections that may limit its use. For example tied to the limited amount of input most GANs work with (few images) and the inability to use the process on continuous messaging of a video, as it will take up a lot of calculation.
The Human Switch
DCNNS and GANs lead to a completely new generation of AI systems based on self-developing statistical models whererules can be differentiated into logical rules and statistical rules. An example of a logical rule is that sum of all three angles of a triangle is always 180 degrees. A statistical rule is like when you see dark clouds in the sky above, you know that is very likely going to rain soon. Human thinking contains both logical and statistical thought. Throughout the history of AI, there have been arguments for and against logical and statistical approaches. Currently, the statistical approach seems to have the upper hand. Deep learning is a classical statistic calculation, which is able to simulate human’s intuition just like playing Go.
In traditional programming, input and rules were fed to a computer system, which then calculated and provided the answer. Deep learning systems like DCNNs and GANs are able to use machine learning to autonomously find rules if provided with input of data and desired answers. Even though they currently still tend to need humans to validate their findings, Auto ML is quickly coming closer to being possible, which means fully auto learning is on the event horizon.
The Math Challenges
From the view of math, deep learning is like matrix multiplication. With every calculation of each layer of linear algebra (matrix multiplication) + non linear algebra (function activation), the data will change accordingly. Currently, AI systems that give good result will have 100 and above layers of calculation.
This relies on data, which is now more readily available than ever before. Increasing calculation speeds of GPUs has also helped a lot. As most of model are published publicly and many of them have existing code, the entry barrier to working with specific AI models is very low.
Peng identified three areas where the main challenges for DCNNs and GANs exist today: data, calculation, and algorithm.
The challenges are:
- Data : Similar to the trading market, the common direction of blockchain entrepreneurship. The challenge is how to provide additional value to data, like that found in a centralized data market.
- Calculation: It is challenging to do the calculations in a distributed way
- Today’s deep convolutional neural networks are difficult to decentralized due to technical shortages.
- Security of the data have to be ensured.
- Cost efficiency is a challenge on the business side.
- The ideal situation is using real task as PoW
- Model (Algorithm): just like task trading
Peng Bo is now considering another integration which is called CloudMind. The idea is to train the network to be able to learn from multiple tasks at once rather than complete one task at one time. This is a much hyped topic in the industry, referred to as multimodel. It would, for example, make the AI systems able to analyze a picture while simultaneously doing some translation.