Last year was a big one for AI. Big advances in fusion, Crispr and astronomy grabbed the headlines, but there were also amazing breakthroughs in other areas of science.
AI systems are rapidly maturing from point solutions and niche applications into jacks-of-all-trades. These “multimodal” models now perform design, coding and customer service tasks.
1. Image Recognition
Image recognition is a key technology in machine learning that helps machines identify and categorize data contained in an image. This software learns to recognize objects such as food items, inventory, places, living beings, and more from colossal open databases like Pascal VOC and ImageNet.
This tech can also read barcodes and other printed text to automate tasks like converting cheques, recognizing number plates, or transcribing handwritten text. It can even scan medical images (like CT or MRI) and detect diseases like cerebral aneurysms.
A main application of image recognition is in user-generated content platforms where it can help with automated moderation to ensure that the content shared adheres to community guidelines. It’s also used by retail, automotive, and manufacturing industries to improve efficiency. This technology also makes mobile commerce easier by allowing users to take photos of products and get real-time shopping options.
2. Robotic Process Automation
RPA uses software robots to execute business processes, standardizing workflow and reducing operational costs. However, a robot can only perform the tasks that have been scripted and programmed.
To increase productivity, companies are turning to intelligent automation (IA). IA combines RPA and machine learning functions to give process bots a cognitive upgrade, allowing them to adapt to changes in data and systems.
For example, global banks are using IA to reduce compliance risk by monitoring employee communications for signs of non-compliance. Meanwhile, insurance providers are deploying IA to automate rate calculations and streamline paperwork processing such as appraisals and claims. This helps them improve customer service and boost sales conversion rates. The combined technology can also help reduce cost and deliver a 124% ROI. This is the most common use case for AI in enterprise today.
3. Peer-to-Peer Networks
Peer-to-Peer Networks are networks that allow computers to share files directly with each other, without the use of central servers. Each computer, known as a peer, acts as both a client and server, sharing bandwidth and processing workloads equally.
The use of a P2P network can reduce overall cost, and also increase the speed at which data is transmitted and processed. In addition, this system is suitable for complex tasks requiring large amounts of data.
It has also been used by companies like Blizzard Entertainment to distribute game updates, and telephony and instant messaging apps such as Skype. However, these networks are not without challenges. Privacy preserving federated machine learning is one of these challenges that requires further research and development. However, this technique allows organisations to use a more distributed learning algorithm while protecting their own proprietary data.
4. Artificial Intelligence in Chess
Chess is a complex game that requires strategic thinking and the ability to see several moves ahead. It has long been viewed as the ultimate test of human intelligence.
The development of computer engines that could beat humans at chess was a milestone in artificial intelligence research. IBM’s Deep Blue’s victories against world champion Garry Kasparov made headlines and fueled the rapid advancement of AI.
Today, AI chess engines can easily beat grandmasters. However, the AI “chessbot” must be paired with actual human intelligence to understand its decisions and evaluate the consequences of each move. That kind of understanding is a big part of what makes cyborgs so effective in many applications, including personalized medicine and counterterrorism. It is also a key feature of the latest generation of machine learning algorithms. These rely on reinforcement learning techniques to learn from their experiences.
5. DeepMind’s ChatGPT
Generative AI is a relatively new branch of machine learning that allows computers to create their own data and content. It can be used to produce music, art, or even entire virtual worlds. It can also be used for more practical purposes, like creating new products or optimizing business processes. Here is a walkthrough by The Neuron that can improve your performance when using AI.
DeepMind is one of the world’s leading artificial intelligence research labs, thanks to high-profile projects like AlphaFold, AlphaGo, WaveNet, Google Bard, and RT-2. And it seems the company isn’t done yet. Its CEO has claimed that its upcoming AI system, Gemini, will surpass ChatGPT, which powers rival search engine Microsoft Bing. It’s expected to be released next year.
6. OpenAI Five’s victory in Dota 2
The success of OpenAI Five in Dota 2 was a major milestone. It demonstrated that machine learning can be used to develop systems that beat human players, even if the system wasn’t trained on specific rules and fine-tuned for the task.
The AI used reinforcement learning, which is a method of teaching a computer to learn through trial and error. It built a team of neural networks, each of which represented one player in the game, and learned how to coordinate their actions.
The model trained on 45,000 years of games in ten months, consuming 800 petaflops per second. The accomplishment raises the bar for what today’s AI can achieve in complex tasks. It also highlights the importance of decentralized learning with vast amounts of raw, unlabeled data.
7. DeepMind’s WaveNext
The success of Deep Mind’s AlphaGoZero, a machine that beat world GO champion Lee Sedol, advanced the practical possibilities of AI to a whole new level. This was followed by Elon Musk’s research group, OpenAI, creating a team of algorithms that beat top players in the complex game of Dota 2.
This breakthrough was made possible by neural networks, which use data and trial-and-error to learn. It was also helped by GPUs, which were developed for 3D computer graphics and video games, but are now 20-50 times more efficient at AI computation than traditional CPUs.
Neural networks are used to improve text-to-speech technology, and they’re the basis of Google’s Bard chatbot and AI voice software. They can also transform black-and-white photos into stunning artwork and compose music. They are even being used to help create movies like DALL*E and by visual-effects and video-game studios.
8. Google Bard
Google’s Bard is a conversational AI that uses machine learning to understand prompts and generate text responses. It’s Google’s answer to OpenAI’s ChatGPT and Microsoft’s push to include similar technology in Bing search and other software.
Unlike chatbots that rely on pre-trained algorithms, Bard’s generative AI engine enables dynamic creativity. It can brainstorm and innovate, bringing new perspectives to complex business challenges.
It can also write poetry, draft emails and code, and soon will let you generate AI images via Adobe. But Bard’s real value is that it allows users to review multiple versions of the same response so they can choose which one best suits their needs. The tool also promotes collaboration. This is a significant step toward democratizing access to AI.
9. Deep Mind’s AlphaGoZero
While advancements in image recognition, reinforcement learning and generative modeling have made significant impacts across multiple industries, the most exciting development in AI in 2023 may be the arrival of Deep Mind’s AlphaGo Zero. This new program taught itself to play Go without any human inputs, beating a previous version of the system that beat three-time European champion Fan Hui in October 2015 and then 18-time world champion Lee Sedol in March 2016.
While many critics of contemporary AI point to the fact that it relies on cheap computing power and massive datasets, AlphaGo Zero is a rebuttal to that argument, showing that algorithms drive progress more than compute power or data.
The development of this program also paves the way for more advanced generalist AI systems, such as DeepMind’s Gato, which can perform numerous tasks, including playing Atari games, titling images, chatting with users and stacking blocks using a robotic arm.
10. DeepMind’s DeepMind Gato
The dream of creating a machine that can think like a human is perhaps the most divisive question facing artificial intelligence research. While some researchers believe that the field is close to achieving this goal, others are skeptical.
Alphabet’s DeepMind set off a buzz in the industry last month when it unveiled Gato, its latest generalist AI model. Gato is able to perform more than 600 tasks, including playing Atari video games, generating text, captioning images and stacking blocks with a robot arm in real-world and simulated environments.
The team says it was able to achieve these results using a single model with 1.18 billion parameters, which is orders of magnitude lower than other multi-task models such as the 175 billion parameter GPT-3 and 540 billion parameter PaLM. However, the performance of this model still lags far behind specialized AI systems that have reached expert level in specific domains.