Your choices are predicted on various mobile applications in the form of recommendations all thanks to machine learning. Ever since the coining of the term ‘machine learning’ by Arthur Samuel in 1959, this subfield of artificial intelligence has made significant progress. It is based on the idea that machines can learn from data, identify patterns, and become capable of making decisions with minimal human intervention. It is a fast-emerging field. You may or may not realize but ML is already being used in a lot of our daily life gadgets and applications.
When Mattel’s Hello Barbie doll talks with kids, the microphone on the doll’s necklace captures the sound and sends it to the ToyTalk servers, where using natural language processing, machine learning, and big data, a proper response is generated from the 8,000 lines of dialogue. Servers at ToyTalk take less than a second to transmit the response to the kids. All the major technology giants like Alphabet, NVidia, Microsoft, Facebook, Salesforce, Google, and Tencent, etc. and various educational and governmental institutions around the world are investing a good deal in advancing this technology. As the hype and fear mongering around AI and ML subsides, we go through some of the developments that have happened in the field of machine learning.
The brainchild of Elon Musk, it is a neurotechnology company that aims to combine technology with implants that, instead of activating movement, can interface at broadband speed with other types of external software and gadgets. According to Elon, the long term, the goal of Neuralink is achieving ‘symbiosis’ with AI. Musk perceives artificial intelligence as an existential threat to humanity if it is unchecked. In a recent live streaming event, Elon presented the team that has made a robot capable of working as a sewing machine to implant threads or ‘neural lace’ which are incredibly thin (4 to 6 nanometer) deep in the brain of a person.
‘Neural lace’ will integrate with the brain to create a seamless machine and neural network circuitry. The concept of ‘neural lace’ was first mentioned in The Culture novel by Iain M. Banks. This technology has the potential to be helpful in the world of medicine. Amputees can regain mobility by the use of prosthetics and diseases like Parkinson’s can possibly be treated. The machine they have been able to make since the inception of Neuralink in 2017 involves drilling holes into the subject’s skull and inserting the threads. But, according to one of the co-founders Max Hodak, they will shift to laser drilling in the future that will be much less invasive and would not be even felt by the human being operated upon.
The notable thing about Neuralink is that the technology they have come up with is about 10x faster than what the existing sensors offer. Although commercialization of neural lace will take longer than it seems, Elon has displayed his seriousness about AI and ML with Neuralink.
AlphaZero and AlphaStar
DeepMind is a London-based 2010 company that Alphabet (Google) took over in 2014. DeepMind developed AlphaZero in late 2017 which is a system that taught itself from scratch how to master the game of shogi (Japanese chess), go, and chess. The system has been described by the chess community to have ground-breaking playing style. Chess master Gary Kasparov exclaimed that he can’t disguise his satisfaction with how AlphaZero plays with a very dynamic style, much like his own. AlphaZero’s style is like no other chess playing programs that were developed before it. It swiftly learns games to become the strongest in history for each, in spite of its training from random play, with zero in-built domain knowledge but basic game rules.
Stockfish and IBM’s Deep Blue, the world computer champions of chess, depend upon heuristics and thousands of rules handcrafted by great human players that try to make a response for every possibility in a game. AlphaZero is unconventional as it replaces these hand-crafted responses with a deep neural network and general-purpose algorithms. It took it just 9 hours of self-training to decisively defeat Stockfish. And currently, it is the strongest player of go in the world, human or machine, after defeating AlphaGo Zero, which beat Lee Sedol in 2016, the then reigning champion of the game!
In the same line of game-playing programs, DeepMind next developed AlphaStar which became the first artificial intelligence to win over a top professional player. It decisively beat GrzegorzKomincz, one of the world’s best professional StartCraft players by 5-0. StarCraft is considered to be one of the most challenging RTS or Real-Time Strategy games and one of the longest-played e-sports ever. It had been a big challenge for AI until now. AlphaStar played the full game utilizing a deep neural network trained by reinforcement and supervised learning.
Experiments with Google
It is an open-source platform for displaying simple experiments making it easier for anyone to discover machine learning. It is an ongoing experiment where people contribute data in real-time by launching the experiments and interacting with machine learning and AI applications. The games are designed to be cool and fun. Experiments with Google is a very serious exercise than it seems. The seemingly silly experiments fortify the brand, grows its share in the market, and make people think about AI differently.Built with a library code deeplearn.js, ‘Teachable Machine’ works live in one’s browser, and without any need for coding you get to teach a simple AI app by interacting with it. In ‘Quick, Draw’ users try their hand at doodling and machine learning has to find out what it is. Millions of people have launched this experiment, adding a large amount of doodling data. ‘AI Duet’ is a piano that responds to your input. ‘Semi-Conductor’ uses camera live feed from a user to detect arm movement and play instruments simulating a live orchestra wherein musicians follow the movements of the conductor.
Google has ventured further into the field of AI/ML with Gnome and TensorFlow. Google Gnome is an Alexa-like voice-activated device that is meant for one’s backyard. It answers weather-related queries, can tell whether a particular plant is edible, if the material is compostable, and do things like controlling hose and other garden equipment such as a grass mover using voice. TensorFlow is Google’s open-source library for building machine learning applications like neural networks. It makes it easier to refine future results, acquire data, train models, and service predictions. Python is used to provide an easy front-end API for making apps, while C++ is used to execute those apps.
Accurate Estimates from Seemingly Useless Information
Apart from the ‘groundbreaking’ ‘breakthroughs’ in the field of ML and AI, many subtle yet significant things that make on say “wow” have happened in the last years.
MIT CSAIL (Computer Science and Artificial Intelligence Laboratory) has developed RF-Pose, an AI that can sense people through walls. It senses the disruptions in the Wi-Fi signals caused by the person on the other side of the room. RF-Pose uses deep learning to teach wireless devices to sense a person’s pose. It uses only wireless signals. A confidence map is generated by a neural network using wireless data. A skeleton-like figure of a person is then generated depicting the pose through the wall. It is undeterred by poor lighting or multiplicity of individuals. The researchers are now working upon creating a technology that captures even micro-movements by depicting a 3D skeleton of the human behind a wall.
Visual Vibrometry refers to the estimation of properties of materials from small motions in videos. MIT researchers published a paper talking about an algorithm that is capable of deciphering intelligible speech by analyzing minute vibrations of objects captured in a video behind soundproof glass. The same technology can be used to infer the material properties of objects like their weight and stiffness from the video. This technique could be applied to figuring out the physical properties of materials without sample extraction from them. For example, structural defects in an airplane’s wings can be analyzed from its vibration video during flight.
Audio captured by a single regular smartphone placed adjacent a keyboard can be used to estimate the keystrokes with 94% accuracy. Previously supervised deep learning was employed with several microphones placed around the keyboard, but this paper used a very simple machine learning method (K-means clustering) and unsupervised learning to arrive at the result.
Deepfake combines and superimposes videos and images onto source videos and images using the generative adversarial network, a technique in machine learning. Deepfake is a sensitive and controversial technique, it has been used to create sensational celebrity fake videos and it can be termed as one of the reasons for people’s aversion to AI.
In the field of art, AI and ML have been utilized to teach machines how to draw. RNN or recurrent neural network is used to make drawings that are graphics based on vectors. You can think of it as an automated Adobe Illustrator. In another technique, researchers were able to make a fake video of any real person dancing amazingly by combining two videos- short videos of a good dancer dancing and another of the target person dancing.
Self-driving cars, humanoids, and advanced weapons are other important subjects on the topic. The enthusiasm with which research is being carried out in AI and ML, something cool gets discovered very often. We hope to see many more exciting AI and ML technology this year. This sure is an adventurous time!