“In the past 2,000 years, the hardware in our brains has not improved… In the next 30 years, AI will overtake human intelligence,” says Softbank CEO Masayoshi Son.
If you’ve read Ray Kurzweil’s “The Singularity is Near: When Humans Transcend Biology,” you’d expect that AI is going to exhibit human-level intelligence in a decade or two. The startlingly thought-provoking work by the futurist gives you a fair picture of the road ahead, a time when humans, with the aid of advanced technologies, will “transcend their biological limitations.”
And you know what? This plausible scenario is at our doorstep. With superintelligence on the brink of becoming a reality, his words ring true, although they are downright scary. Computers and their growing abilities are likely to outpace our skills sooner than we think.
Terms like artificial intelligence and machine learning have been bandied about for a while now. Despite the groundbreaking strides, in terms of intuition, vision, common sense, and language, there are miles to cover. Machines can’t still beat us at everything we do, but they’ve surely have outsmarted us in some ways.
This post talks about some amazing artificial intelligence software that are just so smart.
Latest Artificial Intelligence Software
1. Deep Mind’s AlphaGo
In 2016, AlphaGo was in the news for beating the 9-Dan top player Lee Sedol at Go. According to Wikipedia, the ancient Chinese game of Go is “an abstract strategy board game for two players, in which the aim is to surround more territory than the opponent.”
The AI software from Google beat the South Korean grandmaster in a five-game match, winning 4–1. Brute-force calculations will not work with this complex game. It needed much more.
AlphaGo used deep neural networks and advanced tree search to win. “AlphaGo learned to discover new strategies for itself, by playing millions of games between its neural networks, against themselves, and gradually improving,” said David Silver, Go team’s main programmer. Of the two artificial networks used, the policy network predicted the next move and the value network evaluated the winner of every position on the board.
The team used the Google Cloud Platform for the massive computing power it needed. With advanced machine learning techniques, such as reinforcement learning, and fantastic engineering skills, DeepMind did much better than expected. The cyborg had to figure out how to win, and not just know how to mimic human moves.
This highly publicized event marked the beginning of a new era. Considering the magic of Moves 37 and 78, it was more a case of a human and machine than human against machine. This outcome has immense possibilities. Like computer scientist Andy Salerno says, “AlphaGo isn’t a mysterious beast from some distant unknown planet. AlphaGo is us. AlphaGo is our incessant curiosity. AlphaGo is our drive to push ourselves beyond what we thought possible.” You can read more here.
Quite like Go, Poker fell to the magic of AI as well. In a hands-on no-limit Texas hold’em game, DeepStack beat pro poker players. The algorithm had a staggering 450 milli big blinds per game when a professional player typically has a win rate of 50 milli big blinds per game. This is quite an achievement considering this version of poker has 10160 paths that are possible for each hand!
DeepStack is based more on “intuition” than on working out the moves ahead of time. The algorithm makes real-time decisions by computing fewer possibilities in a matter of seconds.
In their paper, a team of researchers from the Czech Technical University and Charles University in the Czech Republic and the University of Alberta in Canada, talks about the winning AI algorithm DeepStack, which “combines recursive reasoning to handle information asymmetry, decomposition to focus computation on the relevant decision, and a form of intuition that is automatically learned from self-play using deep learning.” A team from Carnegie Mellon has also developed another winning AI software called Libratus. However, game theory won’t hold for multi-player games.
This approach has important implications in other fields that have imperfect information such as medicine, finance, cybersecurity, and defense.
3. AI Duet
An artificial “pianist” from Google’s Creative Lab, AI Duet was built in collaboration with Yotam Mann, developer/musician. Watch this short video and see it working:
In this video, he tells you how this AI software works using the concept of neural networks. This interactive experiment is part of Magenta, an open-source project from Google’s Google Brain unit. You can access the code here. AI Duet is built with Tone.js, TensorFlow, and other Magenta tools.
Who needs a partner when this virtual piano player will accompany you in a lilting duet!
Even if you are no Chopin, this intelligent software will respond to you and create a rhythm. It could even inspire you. It is not going to get you ready for a concert in Boston Symphony Hall, you could have some real fun hitting random notes and waiting for the computer to come back with something improvisational based on melodies it has been trained on.
It looks like artificial intelligence is revolutionizing investment banking. JPMorgan’s software COIN, which is an acronym for contract intelligence, has worked magic by “interpreting commercial loan agreements” in seconds, a task that previously cost 360,000 man hours.
COIN is based on machine learning concepts. The software is naturally less error-prone while checking loan-servicing agreements. A Bloomberg report said that JPMorgan is keen on “deploying the technology which learns by ingesting data to identify patterns and relationships. The bank plans to use it for other types of complex legal filings like credit-default swaps and custody agreements. Someday, the firm may use it to help interpret regulations and analyze corporate communications.”
The company believes that it is only the start of smart automation of processes in the financial industry. JP Morgan is committed to new initiatives. “We’re willing to invest to stay ahead of the curve, even if in the final analysis some of that money will go to product or a service that wasn’t needed,” said Marianne Lake, the finance chief.
Lip reading has become so easy with University of Oxford’s Department of Computer Science’s AI software, LipNet. The team of researchers have detailed it in the paper titled Lipnet: End-to-end sentence-level lipreading.
The paper says, LipNet “maps a variable-length sequence of video frames to text, making use of spatiotemporal convolutions, a recurrent network, and the connectionist temporal classification loss, trained entirely end-to-end.”
Watch this short interesting video:
When you compare this neural network-based software to human lip readers where the accuracy is 12.3%, it has an accuracy of 46.8% while annotating video footage. “All existing [lip-reading approaches] perform only word classification, not sentence-level sequence prediction…. To the best of our knowledge, LipNet is the first lip-reading model to operate at sentence-level,” say the researchers. AI will soon be able to transcribe footage that has a low frame rate and poor image quality sooner than we think.
Apart from the immense help it will be to people who suffer from disabling hearing loss, the team is also interested in its practical possibilities such as “silent dictation in public spaces, covert conversations, speech recognition in noisy environments, biometric identification, and silent-movie processing.”
For those who fear the dark side of AI, this new “killer” program is just another factor reinforcing their misgivings. MIT’s Computer Science and Artificial Intelligence Laboratory has come up with “Philip,” who is out for blood in the popular Super Smash Bros Melee multiplayer video game.
It is based on neural networks and is an “in-game computer player that learned everything from scratch.” The team led by Vlad Firou fed the vicious AI coordinates of the gameplay objects. In their deep reinforcement learning technique, the computer played itself repeatedly in Nintendo’s popular console game.
The team used algorithms such as Actor-Critic and Q Learning to beat 10 top-ranked human players. Philip bested the players with a reaction time of 33 milliseconds and being 6 times faster than humans.
You can read the research paper here.
Cambridge University and Microsoft have come up with deep learning-based software, called DeepCoder, that can write code on its own. “The approach is to train a neural network to predict properties of the program that generated the outputs from the inputs. We use the neural network’s predictions to augment search techniques from the programming languages community, including enumerative search and an SMT-based solver,” says the team in its research paper.
They used a domain-specific language to teach the system to solve online programming challenges involving 3 to 6 lines of code. The system practices and figures out what code combinations work best. Using program synthesis, DeepCoder puts together pieces of code from software that already exists just like a programmer would.
One of the researchers Marc Brockschmidt says, “We’re targeting the people who can’t or don’t want to code, but can specify what their problem is.”
A deep learning AI system from Google can detect cancer with better accuracy and speed than pathologists. Identifying tumors scanning images can be error-prone and laborious.
Here’s a video tutorial on learning about googlenet in detail:
Google says, “After additional customization, including training networks to examine the image at different magnifications (much like what a pathologist does), we showed that it was possible to train a model that either matched or exceeded the performance of a pathologist who had unlimited time to examine the slides.”
“We present a framework to automatically detect and localise tumours as small as 100 × 100 pixels in gigapixel microscopy images sized 100,000×100,000 pixels. Our method leverages a convolutional neural network (CNN) architecture and obtains state-of-the-art results on the Camelyon16 dataset in the challenging lesion-level tumour detection task,” writes Google’s team in its white paper.
Google will continue its research, working on larger datasets, to improve patient outcomes.
New possibilities and advances in artificial intelligence are pushing the boundaries of the human brain like never before. The brilliant artificial intelligence programs outlined in this post is only a glimpse into a terrifying future. If these trends continue, scientists believe that machines could surpass human capabilities sooner than later. But there really is no reason for mass hysteria as of now argues the other camp. Only time will tell, right?