Sundar Pichai, CEO of Alphabet, demonstrates a breakthrough Conversational AI technology LaMDA (Language Model for Dialogue Applications. Like other language models, LaMDA is a Transformer-based neural network architecture model but are trained on dialogue. Since the release of Transformer by Google Research in 2017, several large-scale language models like GPT-3, DeBERT, and ROBERTA were released that have revolutionized the artificial intelligence industry. Today, language models can generate code, summarise articles, and more.
However, Transformer-based models can be heavily limited to specific tasks or require training the pre-build models with new information to effectively on a wide range of functions.
To make models topic/task agnostic, Google blazed a trail and trained LaMDA on dialogue, especially with chatbots. “During its training, it picked up on several of the nuances that distinguish open-ended conversation from other forms of language. One of those nuances is sensibleness. Basically: Does the response to a given conversational context make sense?;” mentions Google.
Google has witnessed a superior performance of LaMDA while asking a few questions. The video below demonstrates LaMDA’s capability with open-ended conversation.
LaMDA resulted from Google’s earlier research published in 2020 that showcased that Transformer-based language models can be trained on dialogue to improve use cases on numerous tasks. Since it is still early in the research, Google is committed to revolutionizing Conversational AI technologies.
For now, Google will be focusing on the sensibleness and satisfyingness of response from LaMDA to further enhance its responses.
Humans have the ability to forget unnecessary information to make space for new patterns that can be significant while making decisions. Facebook AI is moving the needle with Expire-Span to make this a reality for machine learning models. As machine learning models learn new patterns, it keeps on collecting new information, making them computation intensive. As a workaround, researchers embraced the compression technique, where less relevant data is compressed. But, this resulted in blurry visions of memory for tasks that require models to look a long way back to enhance accuracy.
To eliminate this challenge, Facebook AI introduced a novel approach — Expire-Span — to set the expiration time of data. According to the researchers, Expire-Span is a first-of-its-kind operation that enables neural networks to forget at scale. It allows machine learning-based systems to make space for more information while reducing the computational requirements.
Facebook AI’s Expire-Span
For instance, if a machine learning model is tasked to find a yellow door, it stores all the patterns collected while iterating to find the right path. Even after finding the correct patterns, it remembers other unnecessary details that might not help it achieve its goal in the future. This is where Facebook AI’s Expire-Span approach is making the grounds in achieving human-like abilities by deleting nonessential data.
“Expire-Span calculates the information’s expiration value for each hidden state each time a new piece of information is presented, and determines how long that information is preserved as a memory,” mentions the researchers. Facebook AI’s Expire-Span determines the span based on context learned from data and influenced by its surrounding memories. Besides, the span size can be adjusted when needed at a later stage to retain information for a longer period.
Image Credit: Facebook AI blog
Facebook AI researchers evaluated the performance of models that are equipped with Expire-Span against the state-of-the-art models. The Expire-Span models required less computation while delivering comparable performance. “The impressive scalability and efficiency of Expire-Span has exciting implications for one day achieving a wide-range of difficult, human-like AI capabilities that otherwise would not be possible,” wrote the researchers.
The research is in its early stages, and Facebook AI is committed to further enhancing the capabilities of the approach. Nevertheless, the researchers believe that Expire-Span can go beyond research and help in real-world applications.
DeepLearning.AI launches Machine Learning Engineering or Production (MLOps) Specialization to help learners become industry-ready. Today, MLOps has become an essential part of organizations to build robust machine learning-based solutions. However, there is a dearth of courses that can enable learners to build end-to-end AI solutions.
Taught by instructors from Google, Pulsar, and Andrew Ng, there are four courses in the MLOps specialization by DeepLearning.AI — introduction to machine learning in production, machine learning data lifecycle in production, machine learning modeling pipelines in production, and deploying machine learning models in production.
Andrew Ng announced on LinkedIn the release of the MLOps specialization by DeepLearning.AI on Coursera. The specialization covers the designing of an ML production system, modeling strategies, development requirements, establishing a model baseline, building data pipelines, and more.
“Being able to train ML models is essential. And, to build an effective AI career, you need production engineering skills as well, to build and deploy ML systems. With this specialization, you can grow your knowledge of ML into production-ready skills.,” wrote Andrew Ng on LinkedIn.
Since the specialization is categorized as advanced level, you would require prerequisites like knowledge of Python and familiarity with deep learning frameworks like PyTorch, Keras, or TensorFlow.
As India is navigating through an unprecedented second wave of the covid-19 spread, various organizations, including Coursera, have come together to assist the country. Coursera has devised a special collection of courses to offer for free with certifications in India.
The curated course includes not only artificial intelligence, cloud, and application development course but also personal development and public health. The last day to avail of the offer is on June 30.
The discount will be automatically applied during the checkout. According to Coursera, you can only enroll in one course during the offer period. However, we were able to enroll in multiple courses.
Microsoft Counterfit release will allow the artificial intelligence community to quickly determine security flaws in their machine learning-based applications used for businesses. As the adoption of AI applications is proliferating in both business and consumer markets, the need to protect personal information from sneaking out from the ML models.
According to a survey by Microsoft, 25 out of 28 businesses do not have the right tools to secure their AI systems. Unlike other applications, AI-based software are prone to a wide range of security attacks, including adversarial attacks and data leaks. Such attacks not only hamper the brand of organizations but also lead to monetary loss due to stringent data privacy laws in place.
Since machine learning applications vary widely based on the algorithms and architecture used, companies specifically address every application’s shortcoming in security. However, to assist organizations, Microsoft releases Counterfit, which can be leveraged with most machine learning systems.
Counterfit was born out of the internal needs of Microsoft AI systems for pinpointing vulnerabilities. Over the years, the company enhanced Counterfit to make it a generic automation tool that can evaluate multiple AI systems at scale. Today, Counterfit is environment, model, and data agnostic, making it an ideal tool to leverage in numerous use cases.
“Under the hood, Counterfit is a command-line tool that provides a generic automation layer for adversarial AI frameworks such as Adversarial Robustness Toolbox and TextAttack,” mentions Microsoft.
Users can leverage Counterfit for penetration testing and red teaming AI systems, vulnerability scanning for AI systems, and logging for AI systems.
Microsoft heavily relies on Counterfit to make their artificial intelligence applications robots before shipping them to the market. Currently, it can not be used before the models and applications hit production. But, it is being piloted to find AI-specific vulnerabilities before taking the efforts into production.
In a move to influence Google to make necessary changes in its ethical AI practices, three groups — Black In AI, Queer in AI, and Widening NLP — release a joint statement. All three groups work toward elevating underrepresented voices in artificial intelligence and work in collaboration with Google, Apple, IBM, Microsoft, NVIDIA, DeepMind, and more.
Google and DeepMind are some of the gold and diamond sponsors of these groups, making significant contributions to fund the initiatives of the community. But, due to the recent frenzy by Google with its top Ethics AI researchers Timnit Gebru and Dr. Margaret Mitchell.
While both the researchers claim that Google fired them, the search engine giant says that Timnit left the company, and Magaret moved files out of the organization. However, according to Timnit, Google fired her because she voiced her concerns about the companies pressure to pull her name out from the paper that pinpointed flaws in large language models, especially toward the community of color.
“We strongly condemn Google’s actions to dismiss Dr. Timnit Gebru and Dr. Margaret Mitchell, disrupting the lives and work of both researchers and stymying efforts of the Ethical AI team they managed,” reads the joint statement.
In the joint statement, the three groups also extend their support to all others who have more or less witnessed a similar resentment but did not go mainstream. The groups believe that Google’s approach to handling the situation has harmed the Black and Queer community by undermining the importance of inclusion and critical research.
“They not only have caused damage but set a dangerous precedent for what type of research, advocacy, and retaliation is permissible in our community,” believes the three groups.
The joint statement by the Black In AI, Queer in AI, and Widening NLP also asked Google to implement steps demanded by Google Walkout to bring back the trust among researchers and the Black and Queer community.
As a part of Google’s Contact Center AI (CCI) initiative to integrate artificial intelligence in customer support, Google announces Agent Assist for Chat. The solution will help human support agents during live calls and chats by providing relevant answers and FAQs to quickly resolve the issues in real time.
Agent Assist will understand the intent of the customer and pull out the most likely resources that would help human assistants in addressing the problems quickly. According to Google, clients leveraging Agent Assist for Chat have managed conversations concurrently, up by 28%. In addition, it has improved the response time by 15%, thereby reducing the wait time on calls or chats for other customers.
Agent Assist for Chat consists of two crucial components — Smart Reply and Knowledge Assist. While Smart Reply assists in suggesting the right messages extracted from the top performing agents, Knowledge Assist provides articles and FAQs to instantaneously find the perfect answer to customers’ problems.
One of the most critical aspects of Agent Assist is that the machine learning model is trained on your data to enhance the accuracy of recommendations. The is a crucial component as support can vary widely based on the way businesses conduct their operations.
With the public API, organizations can also integrate Agent Assist into their agent desktop to control the agent experience from end-to-end. Superior customer support can become a differentiating factor for organizations in the competitive market to boost their businesses. Using artificial intelligence can empower customer support representatives to provide exceptional customer experience as well as quickly resolve queries in real-time.
As the rate at which data is being collected by organizations is increasing, traditional relational databases have failed to offer the required scale, compute, and flow of data. Consequently, data warehouses became prominent across data-driven organizations for accelerating insights delivery. Today, almost every organization is switching to cloud data warehouses from on-premise infrastructures for streamlining the data workloads across the departments. This is enabling companies to democratize data and boost decision-making for business growth. In the competitive landscape, a proper data warehouse house selection can be the differentiating factor for companies to augment their business processes. However, due to the availability of numerous options for cloud data warehouses, enterprises have struggled to evaluate and find the best fit according to their needs. To simplify the evaluation for companies, this post will focus on comparing famous cloud warehouses — Redshift vs BigQuery vs Snowflake.
While there could be numerous ways in which a data warehouse can be evaluated by organizations, some of the most prominent ways to determine the best data warehouse can be as follows:-
Maintenance In Redshift vs BigQuery vs Snowflake
Organizations embrace managed data warehouses to eliminate maintenance overheads, making it one of the most vital factors while assessing different data warehouse service providers. While Snowflake and BigQuery require little to no maintenance, Redshift requires experts to perform manual maintenance occasionally.
Since storage and compute are not separated in AWS Redshift, you need to set up suitable clusters and optimize the workflow for better performance. But, with BigQuery and Snowflake, initial configuration can be performed without considering different requirements; the flexibility at the later stage by BigQuery and Snowflake removes the necessity of doing due diligence at the very beginning.
AWS also requires you to clear the vacuums — unoccupied spaces — created by the data over a period of time. BigQuery and Snowflake, in contrast, automate the removal of voids to optimize the storage capacity for better performance. Overall, to manage Redshift, an expert familiar with AWS would be vital for you to purge any hindrance during operations. With BigQuery and Snowflake, you do not necessarily need an expert to manage the workflows.
The ability of vertical and horizontal scalability was crucial to the proliferation of cloud data warehouses for organizations. While vertical scaling helps in increasing the load, horizontal scaling enables vast computation. Unlike BigQuery and Snowflake where the storage and compute is different, Redshift has grouped the two called cluster.
Each cluster is a collection of computing resources called nodes, which contains databases. Configuring these clusters is not immediate, disrupting the workflow while amending clusters for vertical or horizontal scaling. But, with Google BigQuery and Snowflake, scaling is performed in a flash to allow users to have continuous access to data warehouses while scaling.
Pricing In Redshift vs BigQuery vs Snowflake
As configurations of clusters are mostly fixed, the pricing with AWS is predictable. You can start with $0.25 per hour and scale according to your needs. However, to optimize costs when you are using it less than usual, you would be required to adjust the clusters on a daily or weekly basis. Therefore, AWS Redshift is popular among companies that have a steady usage of data warehouses. For companies that witness idle time or a surge in usage, it is recommended to fancy Snowflake or BigQuery.
Since Snowflake and BigQuery have different storage and computer pricing, predicting costs is not straightforward. For storage, BigQuery has two pricing models — active storage and long-term storage. While active storage is any table that has been modified in the last 90 days, long-term storage refers to tables that have witnessed an amendment for 90 days.
The active storage plan costs $0.020 per GB, and the long-term storage plan costs $0.010 per GB. Google also offers two pricing models for computing — on-demand pricing and flat-rate pricing. With on-demand, you are charged for each query, which is $5 per TB (the first 1TB per month is free). However, you are not charged for queries that return an error or provide results from cache. And for flat-rate pricing, you will shell out $2,000 for 100 slots — a dedicated query processing capacity.
Snowflake has set the storage pricing of $23 per TB, which is almost similar to BigQuery’s storage cost. However, you will be charged $40 per TB if you opt for on-demand usage. And for computing resources, Snowflake charges $0.00056 per second per credit for Standard Edition.
The pricing of Snowflake is more complicated, but it makes up with its cluster management, which stops clusters when not in use. As a result, you save significantly on processing costs. As per a benchmark, Snowflake is slightly cheaper than BigQuery on regular usage.
Performance In Redshift vs BigQuery vs Snowflake
Evaluating a data warehouse’s performance can be subjective and can differ based on the metric you want to consider. According to a benchmark, separating compute with storage has a unique advantage in processing speed. For instance, Snowflake processes 6 to 60 million rows between 2 to 10 seconds.
But, as per another benchmark that assessed Redshift vs BigQuery vs Snowflake on 24 tables, with the largest one containing 4 million rows of data, the average runtime of 99 TPC-DS queries for BigQuery was 11.18 seconds, Redshift was 8.24, and Snowflake was 8.21 seconds. If your usage is intensive, leverage Redshift or Snowflake is an ideal choice.
Conclusion
Accessing Redshift vs BigQuery vs Snowflake can be a lot easier when your requirements are well-defined. If you are looking for heavy and steady usage without maintenance overhead, Snowflake would be the best choice. On the other hand, if you want flexibility with performance, Redshift should be the go-to service for data warehousing. And in case of varied workloads, BigQuery shall cater to your needs with minimal cost since it charges you for the queries you request.
Apple hires Samy Bengio, former scientists of Google, to lead a new AI research unit. Bengio has joined as a research director at Apple after leaving Google on 28 April 2021.
According to reports, Bengio will report to his former colleague John Giannandrea, who is now a senior vice president of machine learning and AI strategy. Giannandrea had earlier worked at Google for a little less than 8 years and was a senior vice president of engineering.
It is believed that Bengio left Google due to the recent restructuring of the ethical AI team. In his email to the Google research team, obtained by CNBC, he wrote “This is one of the most difficult emails I can think of sending to all of you: I have decided to leave Google in order to pursue other exciting opportunities. There’s no doubt that leaving this wonderful team is really difficult.”
Google AI, especially its ethical AI team, has witness criticism from every corner of the AI community after the ouster of Timmit Gebru and Margaret Mitchell, who were the leads of the ethical AI team. As per Timmit, Google fired her for her work that exposed biases in large language models. Following Timmit’s exit, Bengio shared his disbelief on social media and extend his support to the team and Timmit.
Bengio has moved on to a new challenge after working for more than 14 years at Google. While working for Google, Bengio worked on a wide range of projects, including search and speech. Bengio is also the brother of Yoshua Bengio, a joint recipient of the Turing award 2018 winner.
Julia Computing releases DataFrames.jl 1.0 to allow data scientists to work effectively with tabular data. DataFrames is an equivalent to pandas library that allows data scientists to manipulate large datasets for gaining insights. The latest release of DataFrames brings new capabilities for users to effectively handle and analyze data.
Julia lang is gaining popularity in the data science landscape due to its ability to quickly process a colossal amount of datasets. According to various reports, Julia is faster than Python, giving data scientists an edge while analyzing a plethora of information at once.
However, DataFrames is not the only tool for working with tabular data in Julia. Depending on the use cases, one can also leverage TypedTables and JuliDB. While TypedTables is used to obtain optimized performance when the table does not contains thousands of columns, JulidDB is ideal for when you are handling large datasets that cannot fit in the available memory.
One of the crucial features of Julia is that it allows you to switch between the libraries effectively. For instance, you can use Query.jl code to manipulate data in a DataFrame, JuliaDB, and more. Julia DataFrames is available through Julia packages and can be installed using the command Pkg.add(“DataFrames”).
A wide range of libraries for statistics, machine learning, plotting, data wrangling, and more are integrated with Julia DataFrames to streamline the data science workflows.
