NASSCOM Center of Excellence (COE)- IoT & AI announces the launch of the third edition of its Healthcare Innovation Challenge (HIC). Hospitals, diagnostic chains, insurance firms, pharma companies, government representatives, technology enterprises, and deep tech startups will participate in the newly launched NASSCOM CoE program.
The use cases in this third series of HIC are automated credit business settlement, inpatient volume prediction based on outpatient volume, prescription digitalization using voice recognition, early detection of microbes, comprehensive patient care, OPD automation, preventive health checkup tracking, artificial intelligence-based surgical video, recording cum reporting, and cashless OPD expense management.
Use Case sponsors Apollo Hospitals, HealthCare Global Enterprises, Max Healthcare, Hinduja Hospital, Cygnus Hospital, Aditya Birla Health Insurance, and others participated in panel discussions at the launch event.
The previous editions of this program were highly appreciated and received participation from the healthcare industry. However, the newly announced third series of HIC has piqued the interest of insurance and technology companies as well.
Vice President of Imaging System Software at Wipro GE Healthcare India said, “Healthcare is going through a rapid transformation and providing comprehensive patient care connecting hospital systems, laboratory diagnostics, physicians, etc. is the need of the hour. To provide timely diagnostics and treatment, connecting these functions remotely is crucial.”
He further added that using specialized technologies like artificial intelligence, deep learning, safe data management, telemedicine, telemonitoring, and digital solutions can deliver an all-around patient experience.
Over the last few years, NASSCOM has taken many initiatives to promote research and deployment of artificial intelligence technologies in India that have helped not only students but also numerous tech startups to scale their businesses.
HIC will accelerate broad-based deployment participation, boosting the influence of digital technology in the healthcare sector toward the Ministry of Electronics and Information Technology’s (MeitY) aim of a $1 trillion digital economy.
Catering to the increasing needs, Microsoft is extending the native capabilities of Microsoft Defender for Cloud to the Google Cloud Platform to provide better security to its customers.
Customers will benefit from the new capabilities in improving visibility and control across multiple cloud providers, workloads, devices, and digital identities from a centralized management perspective. With this new addition, Microsoft Azure, Amazon Web Services (AWS), and Google Cloud Platform (GCP) will all have native multi-cloud protection.
It’s vital for businesses to properly embrace multi-cloud strategies as their security solutions decrease complexity and provide comprehensive protection.
Vasu Jakkal, Corporate Vice President, Security, Compliance, Identity, and Management at Microsoft, said, “Support for GCP comes with out-of-box recommendations that allow you to configure GCP environments in line with key security standards like the Center for Internet Security (CIS) benchmark, protection for critical workloads running on GCP, including servers, containers and more.”
She further added that in this multi-cloud, multi-platform world, security operations must assess emerging cyber dangers and detect potential blind spots across a wide range of users, devices, and destinations.
Microsoft also announced the release of a public preview of CloudKnox Permissions Management. It is a unique platform that provides total visibility into users’ identities and workloads across clouds.
The platform includes automated features that enforce least privilege access consistently and use ML-powered continuous monitoring to identify and remediate suspicious activity.
“IT teams lack visibility into identities and their permissions and struggle with ever-increasing permission creep. These challenges require a comprehensive, unified solution for full visibility and risk remediation,” said Alex Simons from Azure.
In today’s digital era, artificial intelligence is advancing at a greater pace by having deep learning as its primary contributor. Since deep learning is one of the subfields of artificial intelligence, most of the AI tasks and applications involve deep learning models. Deep learning works similar to the human brain, which perceives and transmits information through countless neuron interactions. The applications of deep learning include image processing, text classification, object segmentation, natural language processing, and much more. To build such high-end applications and use cases, you have to employ appropriate deep learning libraries at different phases of an end-to-end deep learning model development lifecycle. There are a vast collection of libraries available for implementing deep learning tasks, from which you can select the most suitable and efficient library based on your use cases and business models.
This article mainly focuses on the top 8 deep learning libraries that are primarily used by developers at different phases of the deep learning lifecycle.
1. Keras
Keras is one of the most prominent open-source libraries used mainly for implementing deep learning-related tasks. It initially started its journey as a Google project named ONEIROS (Open-Ended Neuro Electronic Intelligent Robot Operating System) for enabling faster experimentation with neural networks. In 2017, Keras was added to Google’s TensorFlow machine framework, making it a high-level API for building and training deep learning models.
Since Keras runs on top of the TensorFlow framework, Keras APIs can be used to effectively run both machine learning and deep learning-related tasks. Keras is highly scalable to run on both high-level GPUs and CPUs for developing complex neural network models with less computation time. Because of such enhanced features and functionalities, Keras empowers researchers and engineers to fully exploit scalability and cross-platform capabilities, thereby enabling them to achieve high accuracy and performance while building deep learning models. In addition, Keras is being used in most popular companies like YouTube, NASA, and Waymo because of its industry-strength performance and scalability.
Keras is highly compatible with Python 3.6 to 3.9 versions, Windows, Ubuntu, and Mac Operating systems. Since Keras is an open-source project, it offers greater community support by means of forums, Google groups, and Slack channels. Keras also provides users with straightforward and well-structured documentation, allowing beginners to easily learn and implement deep learning tasks.
2. TensorFlow
Developed by the Google Brain team, TensorFlow is an open-source Python library for implementing high-level numerical computations and large-scale deep learning tasks. Ever since its development, TensorFlow was only used for inter-organizational purposes in Google. However, it was made open-source under the Apache License 2.0 in 2015. Although the TensorFlow framework is primarily used to build and develop deep learning, it also has flexible tools and libraries for building end-to-end machine learning models.
With TensorFlow, you not only build and develop machine learning and deep learning models but also can perform probabilistic reasoning, predictive modeling, and statistical analytics. Since TensorFlow consists of high-level APIs like Keras and Theano, it can be effectively used at any phase of the model development life cycle. In addition, since TensorFlow supports cross-platform deployment, you can easily build and deploy deep learning models in any production platform, such as cloud and on-premises systems,
TensorFlow is made to be compatible with macOS, Windows, 64 bit-Linux, and mobile computing platforms, including Android. To make developers and researchers effectively work with TensorFlow, its official documentation clearly explains the features, functionalities, and implementation methodologies of the respective library.
3. PyTorch
PyTorch is one of the most popular and open-source deep learning libraries developed by the AI research team of Facebook in 2016. The name of the respective library is based on the popular deep learning framework called Torch, a scientific computation and scripting tool written in the Lua programming language. However, Lua is a complex language to learn and get hands-on and does not offer enough modularity to interface with other libraries. To eradicate such complications, researchers of FaceBook developed and implemented the Torch framework using Python, thereby naming it as PyTorch.
PyTorch not only allows you to implement deep learning-related tasks but also enables you to build computer vision and NLP (Natural Language Processing) applications. In addition, the primary features of PyTorch include tensor computation, automatic differentiation, and GPU acceleration, which makes it stand apart from other top deep learning libraries.
You can flexibly run PyTorch on Linux, Windows, macOS, and any of your preferred cloud computing platforms. PyTorch also offers you standard documentation that specifies its features, functionalities, and algorithms, allowing any user to learn and try implementing deep learning models on their own.
4. MXNet
Developed by Apache Software Foundation, MXNet is one of the open-source deep learning libraries in Python that allows you to define, train, build, and deploy deep neural networks. With MXNet, you can develop and deploy deep learning models in any platform like cloud infrastructure, on-premises, and mobile devices. Since MXNet has ultra-scalability and distributive features, it can be seamlessly scaled across multiple GPUs and machines, leveraging them to achieve fast-model training and high performance.
MXNet supports a wide range of programming languages like Python, C++, R, Julia, Scala, JavaScript, and MatLab, eliminating the need to learn new languages for working with specific frameworks. Since MXNet is language independent, you can build portable and lightweight neutral network representations that can seamlessly run on low-powered devices with limited memory like Raspberry Pi and other single-board computers. Because of such efficient features, MXNet is being used and supported by the most prominent organizations like Amazon, Baidu, Intel, and Microsoft.
MXNet provides you with a greater community that enables you to participate in discussion forums, collaborate with other researchers, and learn the features and functionalities of the respective library via tutorials and documentation.
5. Microsoft CNTK
Released by Microsoft in 2016, CNTK (Cognitive Toolkit), previously known as Computational Network ToolKit, is an open-source deep learning library used to implement distributed deep learning and machine learning tasks. With the CNTK framework, you can easily combine the most popular predictive models like CNN (Convolutional Neural Network), feed-forward DNN (Deep Neural Network), and RNN (Recurrent Neural Network) to effectively implement end-to-end deep learning tasks.
Although CNTK is primarily used to build deep learning models, it can also be used for implementing machine learning tasks and cognitive computing. Though CNTK’s framework functions are written in C++, it also supports a wide range of programming languages like Python, C#, and Java. Furthermore, you can use CNTK for developing efficient deep learning models by either importing it as a library into your preferred development frameworks or using it as a standalone deep learning tool, or launching it in cloud platforms. Due to its platform compatibility and performance, CNTK is being used by the most prominent companies like Cyient and Raytheon.
CNTK provides you with standard documentation and is also available as an open-source repository in GitHub, making it easier for developers and researchers to learn and implement high-level deep learning methodologies.
6. Fastai
Developed by Jeremy Howard and Rachel Thomas in 2016, Fastai is an open-source library primarily used for building deep learning and artificial intelligence models. Since Fastai is built on top of PyTorch, users can leverage the advanced features of both frameworks, thereby achieving high accuracy models with remarkable speed and performance. Apart from its other prominent features, Fastai is the first deep learning library to offer a standalone interface for building various end-to-end deep learning applications, including computer vision, text classification, neural network, and time series models.
As its name implies, Fastai helps developers build efficient and high-level models using minimal amounts of code with faster experimentation capability. Fastai achieves high-speed experimentation since it can automatically figure out suitable pre-processing techniques and training parameters for the specific dataset, making it more accurate than other deep learning libraries.
Fastai offers basic to advanced practical courses for beginners and developers. It also provides users with clear documentation, incorporating features and algorithms of the Fastai library along with their use cases.
7. Theano
Developed in 2010, Theano is an open-source deep learning library for implementing and evaluating more complex mathematical and scientific computations that involve multi-dimensional arrays. With Theano, you can achieve transparent GPU usage by manually setting the GPU usage limit and frequency. You can develop a highly scalable and reliable training framework by utilizing multiple GPUs across the cluster, which results in accelerating the training speed of deep learning models.
With Theano, you can express and define your model in terms of mathematical expressions and computational graphs, making it easy to evaluate and assess the training capability of the respective model. Since Theano offers developers a general-purpose computing framework for implementing complex neural network models with remarkable speed and accuracy, it is extensively utilized in the Python community, especially for deep learning research.
Theano provides you with the comprehensive documentation that incorporates all its functions, methodologies, and algorithms, making it easy for beginners to understand and implement deep learning techniques.
8. Caffe
Developed by BAIR (Berkeley AI Research), Caffe is one of the most popular deep learning libraries for Python, which is used to implement machine vision and forecasting applications. Cafe\fe serves as a one-stop framework for training, building, evaluating, and deploying deep learning models. With Caffe, you can build and evaluate your deep neural networks with a sophisticated set of layer configurations options. However, you can also access pre-made neural networks from the Caffe community website based on your use cases and model preferences.
Since Caffe can be scaled across multiple GPUs and CPUs, you can achieve greater training and processing speed, allowing you to train deep learning models in less time. Because of its features, enhanced training speed, and performance, Caffe is being used by popular organizations like Adobe, Yahoo, and Intel.
Caffe offers you a well-documented user guide, incorporating its philosophy, architecture, methodologies, and use cases. It’s also accessible as an open-source repository on GitHub, letting users experiment with Caffe’s functions and algorithms.
QuantrolOx has secured £1.4 million in seed investment led by Nielsen Ventures and Hoxton Ventures to expand quantum computing. The round was also led by Voima Ventures, Remus Capital, Dr. Hermann Hauser, and Laurent Caraffa. Founded by Oxford professor Andrew Briggs, tech entrepreneur Vishal Chatrath, the company’s chief scientist Natalia Ares, and head of quantum technologies Dominic Lennon co-founded the company, the company aims to manage qubits within quantum computers using machine learning.
Instead of the straightforward manipulation of ones and zeros in traditional binary-based computers, quantum computers employ quantum bits or qubits. In addition, these qubits feature a third state known as “superposition,” which may represent either a one or a zero at the same time. Instead of having the value of either a one or a zero, superposition allows two qubits to represent four situations at the same time. This characteristic can allow a computing revolution in which future computers will be capable of more than just mathematical computations and algorithms.
Quantum computers also use the entanglement principle, which Albert Einstein described as “spooky action at a distance.” The fact that the state of particles from the same quantum system cannot be represented independently of each other is known as entanglement. They are still part of the same system, even though they are separated by huge distances.
QuantrolOx is developing automated machine learning-based control software for quantum technologies that allows them to tune, stabilize, and optimize qubits. Quantum computers need thousands of qubits, yet qubits vary somewhat owing to errors in control instruments, manufacture, and design, necessitating distinct sets of control parameters to make each one useful. To create a functional quantum computer, a complex procedure is necessary. The issues of turning and characterizing qubits become increasingly difficult and significant as the number of qubits grows.
QuantrolOx’s software is technology-neutral, meaning it may be used with any quantum technology. However, for the time being, the company is concentrating on solid-state qubits. That’s primarily because they are systems to which the company has access, including through a tight collaboration with a Finnish lab that the company wasn’t ready to reveal. QuantrolOx, like any other machine learning challenge, requires a large amount of data in order to create successful machine learning models.
QuantrolOx is now focusing on forming new agreements with quantum computer manufacturers. These are significant collaborations since the team not only requires physical access to the equipment but also the source code that controls them in order to interact with these systems.
The Organization for Economic Co-operation and Development (OECD) has created a user-friendly tool or framework to analyze AI systems from a policymaking viewpoint. This framework assists lawmakers, regulators, organization policymakers, and others in characterizing the policies of AI systems deployed in specific sectors. The OECD framework also assists professionals in assessing the potential and risks present in various types of AI systems, as well as informing national AI plans and strategies.
In addition, the OECD framework helps policymakers distinguish between various types of AI systems and all the possible influences that AI has on people’s lives, whether positive or negative. This not only includes what AI systems are capable of but also where and how they implement it. For example, image recognition can be highly useful for smartphone security, but when used in other situations, it might be violating human rights.
The OECD policy classification framework differentiates AI systems based on the dimensions, such as People & Planet, Economic Context, AI Model, Data & Input, and Task & Output. Each dimension has its own set of characteristics and traits that helps in evaluating the policy implications of specific AI systems.
The respective framework works by referring to the AI system’s lifecycle as a supplemental structure for understanding the primary technical properties of the respective system. Furthermore, by defining the qualities and characteristics of AI systems that matter the most, the OECD framework promotes a common understanding of AI and its effective usage across the organization.
According to the OECD’s announcement, the current framework is meant to provide the basis for developing a future risk-assessment framework to help with diminishing and mitigating risks. It will also provide a baseline for the OECD, Members and partner organizations to develop a common framework for reporting about AI incidents.
After the successful deployment of the OECD framework across different organizations, it is expected that the respective classification tool assists in generating more data about the various types of AI systems currently in use around the world. This provides policymakers with the information they need to map the most impactful AI domain and trace interventions that make AI more favorable across the globe.
Global technology and services company Bosch announces that it has acquired high-definition maps providing firm Atlatec.
The new acquisition will allow Bosch to considerably increase its autonomous vehicle offerings to its customers. Both the companies have signed an agreement to complete this deal.
However, no information has been provided by the companies regarding the valuation of this acquisition. Atlatec will join the Bosch Cross-Domain Computing Solutions division as an independent company.
Digital maps play a critical part in the development of self-driving technology from the outset, and the higher the level of automation, the more tightly map production and driving plan programming must be integrated.
Bosch says that high-resolution digital maps are vital in making automated driving functions safe and easy to use. Therefore Bosch has acquired Atlatec as it provides all of the necessary solutions for data recording, processing, map generation, and quality control.
With its own sensor box and associated software, Atlatec has created a scalable solution. This new deal will allow Bosch to reach its goal of developing level 4 autonomous driving vehicles.
President of the Cross-Domain Computing Solutions division of Bosch, Dr. Mathias Pillin, said, “The planned acquisition of Atlatec further expands our expertise in the field of high-resolution digital maps and makes us even more diversified. It makes Bosch the only company that can offer its customers all the necessary building blocks of automated driving.”
Digital maps, in addition to onboard sensors with radar, video, and ultrasonic technology, are essential sensors in automated driving. The company’s solution collects raw data and analyzes it using artificial intelligence algorithms to generate accurate maps. An autonomously driving car, for example, can adjust its speed in advance of a tight bend based on the map information.
The Gujarat Forest Department announces its plans to use an artificial intelligence-powered solution to identify Asiatic lions in the region.
The department is using a unique AI software named SIMBA (Software with Intelligence Marking Based Identification of Asiatic lions) to expand its capabilities of spotting lions based on their physical features.
According to officials, the tool will be used to better analyze population demographics and to expand conservation and management efforts to help Asian lions survive. Individual identification helps to improve conservation and management operations along with ecological elements of the species.
Teliolabs, a Hyderabad-based business that develops enterprise software solutions, has built the AI-powered software SIMBA.
The software can also be used to pinpoint a single lion based on gender, name, microchip number, life status, and lactating status information available in the database.
Deputy Conservator of Forest, Sasan, Mohan Ram, said, “The SIMBA works with a deep machine learning technique that matches a point pattern for pair-wise, and that automates the individual identification, based on the variability in the individual’s whisker spot patterns, scars on the face, notches on the ears, and other metadata of the photograph.”
The software can isolate the region of interest (ROI) and segregate them for identification after being trained with enough training data. In addition, SIMBA can be used to create a database with a unique identification number or name. The software is capable of extracting a photograph’s originality and aggregating comparable patterns or marks within the machine learning’s embedding space.
An advantage of SIMBA is that it comes with a very user-friendly interface, making the software easy to operate. Experts believe that SIMBA can play a significant role in boosting the government’s efforts to conserve the species in the region.
The Scientists of Max Planck Institute for Intelligent Systems developed “Insight,” a robust and soft haptic sensor, which leverages the capabilities of computer vision and deep neural networks to estimate from where and how it is being touched or contacted. The sensor also determines the strength and magnitude of the applied forces.
The research work by MPI-IS is regarded as a massive step for making robots feel and sense their surroundings as precisely as people and animals. Imitating its natural predecessor, the haptic sensor is made to be extremely sensitive and durable.
By sensing the contact and friction of external objects, the sensor can easily identify the touch. The sensor is made of a soft shell that surrounds a lightweight stiff skeleton, which makes the respective sensor look like a thumb. The stiff skeleton supports the sensor structure in the same way that bones support the tissues of the finger.
To provide sensitivity, robustness, and soft contact, the sensor shell consists of a single sheet of elastomer over-molded over a strong frame. Furthermore, the elastomer is combined with dark yet reflective metal flakes to create an opaque grayish finish that keeps any outside light out. A small 160-degree fish-eye camera placed within this finger-sized cap helps in recording beautiful images lighted by a ring of LEDs.
The sensor also has a nail-shaped zone that is comparatively thinner than the other parts, making the sensor more accurate in sensing external contacts. The scientists made an elastomer with a thickness of 1.2 mm for this super-sensitive zone rather than the 4 mm, which is utilized on the remainder of the finger sensor. In addition, the haptic fovea is built to sense even the smallest forces and intricate object forms.
The respective sensor works by sensing and predicting the external contact or touch. When any object comes in contact with the sensor’s shell (outer layer), the color pattern inside the sensor changes. The sensor’s camera continuously captures the pictures at a high rate and feeds the image data to a deep neural network. Even the tiniest change of light in each pixel is detected by the algorithm. The well-trained machine learning algorithm can map out precisely where the finger is striking or touching the shell and calculate the strength and magnitude of the applied force. The model also generates a force map, which gives a force vector for each point in the three-dimensional fingertip.
The Indian MOS&T (Ministry of Science and Technology) has recently announced the plan of introducing nine more supercomputers at well-known institutions across the nation. As per the ministry, the nine supercomputers will be installed at Indian institutions such as IIT Bombay, IIT Madras, IIT Delhi, IUAC Delhi, IIT Patna, CDAC-Pune, SNBNCBS, NCRA Pune, and NIC Delhi.
The supercomputers have been installed and commissioned under the scheme of the National Supercomputing Mission (NSM). According to the MOS&T, the National Supercomputing Mission was established to strengthen the country’s research skills and capabilities by linking them to form a supercomputing grid, also with National Knowledge Network (NKN) serving as the backbone.
Previously, several indigenous supercomputers were launched at Indian academic institutes as part of the National Supercomputing Mission. PARAM Shivay was the first supercomputer built in India, which was installed at IIT (BHU). After the successful installation of Param Shivay, several supercomputers like PARAM Shakti, PARAM Brahma, PARAM Yukti, and PARAM Sanganak were installed at IIT Kharagpur, IISER Pune, JNCASR, and IIT Kanpur, IIT Hyderabad, NABI Mohali, and CDAC Bengaluru, respectively.
In addition, in the first week of February, a supercomputer named Param Pravega was installed at the Indian Institute of Science (IISc) Bangalore under the NSM mission. Param Pravega is said to be one of the most powerful supercomputers of the country, as well as the largest supercomputer deployed in all Indian academic institutions.
As a part of the NSM scheme, the mission has also trained over 11,000 HPC-aware personnel and faculty, resulting in the next generation of supercomputer specialists. Furthermore, four NSM nodal centers HPC (High Performance Computing) and AI training have been created at IIT Kharagpur, IIT Madras, IIT Goa, and IIT Palakkad to extend the activities of HPC training.
According to the MOS&T’s press release, “Powered by the NSM, India’s network of research institutions, in collaboration with the industry, is scaling up the technology and manufacturing capability to make more and more parts in India, taking indigenous manufacturing to 85%.”
After a big rebranding from Facebook, Meta hosted its inaugural conference to unveil its latest Metaverse projects. The virtual Meta event ‘Inside the lab: Building for the metaverse with AI’ was hosted on February 23 at 10:30 p.m. (IST) and featured speakers from across the AI spectrum, with CEO Mark Zuckerberg delivering the opening and closing addresses.
The Metaverse is now a headlining topic, particularly in terms of its potential to revolutionize how we work, live, and play. The premise that the Metaverse will be a well-defined ideal of a technological future that will accomplish for society what the Internet did for the twenty-first century has already been pursued by tech pundits.
Even Zuckerberg is betting that the Metaverse will be the successor to the mobile Internet. While highlighting the difficulties in creating the Metaverse, Zuckerburg stated that AI is the key to unlocking many breakthroughs. Furthermore, because these worlds will be dynamic and constantly changing, AI needs to grasp the context and learn the same way humans do. Zuckerberg said the Metaverse would consist of immersive worlds the user can create and interact with, including the position in 3D space, body language, facial gestures, etc. “So, you experience it and move through it as if you are really there,” he added.
No Language Left Behind
At present we have numerous tools and apps that help us communicate in common languages like English, Spanish, and Mandarin with online tools and software. However, there are billions of people who are not able to access the Internet and its offerings in their native languages. While machine translation systems are catching up, the key problem is building such tools when there are fewer or almost no textual resources.
According to Zuckerberg, Meta AI is invested in developing language and machine translation technologies that will cover the majority of the world’s languages. And now, two more projects have been added to the mix. The first is No Language Left Behind, a new advanced AI model that can learn from languages with fewer samples to train from, which will enable expert-quality translations in hundreds of languages, from Asturian to Luganda to Urdu. This will be a huge milestone in NLP-based AI tools that focus on eliminating the language barrier in accessing the Internet.
Universal Speech Translation
He also stated that Meta’s research division is working on a universal speech translation system that will let users engage with AI more efficiently within the company’s digital realm. According to Zuckerberg, the main objective is to create a universal model that includes knowledge from all modalities and data collected by rich sensors. Zuckerberg said, “This will enable a vast scale of predictions, decisions, and generation as well as whole new architectures training methods and algorithms that can learn from a vast and diverse range of different inputs.”
Project CAIRaoke
Meta is focusing on AI research to allow users to have more natural interactions with voice assistants, revealed Zuckerberg, shedding light on how humans will connect with AI in the Metaverse. During the Meta AI: Inside the Lab event, meta introduced Project CAIRaoke, a new approach to conversational AI for chatting with chatbots and helpers. The project aims to use neural models to help chatbots better comprehend people when they talk in a conversational manner, enabling more fluid interaction between people and their gadgets using natural language processing.
People will be able to communicate naturally with their conversational assistants using models created with Project CAIRaoke, allowing them to go back to anything from previously in the discussion, shift topics entirely, or say things that require complicated, nuanced context. They’ll also be able to communicate with them in new ways, such as through gestures.
Traditional AI assistants rely on four distinct components: natural language understanding (NLU), dialog state tracking (DST), dialog policy (DP) management, and natural language generation (NLG). These disparate AI systems must then be integrated for hassle-free conversation, making them challenging to optimize, slow to adapt to new or unfamiliar jobs, and reliant on time-consuming annotated data sets. In addition, changes in one component might also break the others, requiring all following modules to be retrained, slowing them down. On the other hand, building a model using Project CAIRaoke will eliminate the reliance on upstream modules, thereby speeding up development and training, and allowing users to fine-tune other models with less time and data.
A demonstration of Project CAIRaoke technology showed a family using it to help make a stew, with the voice assistant warning that salt had already been added to the pot. The assistant also observed that they were low on salt and placed an order for more.
Meta also wants to incorporate Project CAIRaoke into devices with AR and VR, and it will be paired with Meta’s video-calling Portal device. Jérôme Pesenti, Meta’s VP of AI, said the company is currently limiting the answers of its new CAIRaoke-based assistant until it is assured that the assistant didn’t create foul words.
BuilderBot
At the Inside the Lab event, Meta also revealed it is also developing a virtual world AI called BuilderBot, which is based on the same conversational technology and allows individuals to talk their own worlds into existence literally. When BuilderBot is turned on, a user may walk into an empty 3D area, which is only populated with a horizon-spanning grid, and tell the bot what they want to appear in the world.
During the demonstration, Zuckerberg started engaging with a low-resolution virtual environment and began populating it with voice commands. With only verbal commands, Zuckerberg’s avatar transformed the entire 3D landscape into a park and eventually a beach. The BuilderBot then generates a picnic table, boom box, beverages, and other minor items in response to spoken commands.
However, it’s unclear if Builder Bot uses a library of predefined items to fulfill these tasks or if the AI creates them from scratch. If the latter is true, it will open new avenues for generative AI models.
Next, to showcase Meta’s commitment to AI transparency and open science, Meta unveiled TorchRec, a library for creating SOTA recommendation systems for the open-source PyTorch machine learning framework, at the Inside the Lab event. It’s available as PyTorch memory and includes primitives for common sparsity and parallelism, allowing researchers to create the same cutting-edge personalization that Facebook Newsfeed and Instagram Reels employ.
AI Learning Alliance
Meta also stated that it would extend free education initiatives focused on attracting more racial minorities to the field of technology, which experts believe is necessary to develop AI systems that are devoid of bias. Through its online learning platform Blueprint, Meta’s AI Learning Alliance aims to make machine learning curriculum available to everybody. The company is also building a consortium of teachers who will teach this curriculum at colleges with large populations of students from underrepresented groups.
