Google Research: Themes from 2021 and Beyond
2886
Of your peers have already read this article.
4:00 Minutes
The most insightful time you'll spend today!
Posted by Jeff Dean, Senior Fellow and SVP of Google Research, on behalf of the entire Google Research community
Over the last several decades, I’ve witnessed a lot of change in the fields of machine learning (ML) and computer science. Early approaches, which often fell short, eventually gave rise to modern approaches that have been very successful. Following that long-arc pattern of progress, I think we’ll see a number of exciting advances over the next several years, advances that will ultimately benefit the lives of billions of people with greater impact than ever before. In this post, I’ll highlight five areas where ML is poised to have such impact. For each, I’ll discuss related research (mostly from 2021) and the directions and progress we’ll likely see in the next few years.
· Trend 1: More Capable, General-Purpose ML Models
· Trend 2: Continued Efficiency Improvements for ML
· Trend 3: ML Is Becoming More Personally and Communally Beneficial
· Trend 4: Growing Benefits of ML in Science, Health and Sustainability
· Trend 5: Deeper and Broader Understanding of ML
Trend 1: More Capable, General-Purpose ML Models
Researchers are training larger, more capable machine learning models than ever before. For example, just in the last couple of years models in the language domain have grown from billions of parameters trained on tens of billions of tokens of data (e.g., the 11B parameter T5 model), to hundreds of billions or trillions of parameters trained on trillions of tokens of data (e.g., dense models such as OpenAI’s 175B parameter GPT-3 model and DeepMind’s 280B parameter Gopher model, and sparse models such as Google’s 600B parameter GShard model and 1.2T parameter GLaM model). These increases in dataset and model size have led to significant increases in accuracy for a wide variety of language tasks, as shown by across-the-board improvements on standard natural language processing (NLP) benchmark tasks (as predicted by work on neural scaling laws for language models and machine translation models).
Many of these advanced models are focused on the single but important modality of written language and have shown state-of-the-art results in language understanding benchmarks and open-ended conversational abilities, even across multiple tasks in a domain. They have also shown exciting capabilities to generalize to new language tasks with relatively little training data, in some cases, with few to no training examples for a new task. A couple of examples include improved long-form question answering, zero-label learning in NLP, and our LaMDA model, which demonstrates a sophisticated ability to carry on open-ended conversations that maintain significant context across multiple turns of dialog.
(Weddell Seal image cropped from Wikimedia CC licensed image.)
Transformer models are also having a major impact in image, video, and speech models, all of which also benefit significantly from scale, as predicted by work on scaling laws for visual transformer models. Transformers for image recognition and for video classification are achieving state-of-the-art results on many benchmarks, and we’ve also demonstrated that co-training models on both image data and video data can improve performance on video tasks compared with video data alone. We’ve developed sparse, axial attention mechanisms for image and video transformers that use computation more efficiently, found better ways of tokenizing images for visual transformer models, and improved our understanding of visual transformer methods by examining how they operate compared with convolutional neural networks. Combining transformer models with convolutional operations has shown significant benefits in visual as well as speech recognition tasks.
The outputs of generative models are also substantially improving. This is most apparent in generative models for images, which have made significant strides over the last few years. For example, recent models have demonstrated the ability to create realistic images given just a category (e.g., “irish setter” or “streetcar”, if you desire), can “fill in” a low-resolution image to create a natural-looking high-resolution counterpart (“computer, enhance!”), and can even create natural-looking aerial nature scenes of arbitrary length. As another example, images can be converted to a sequence of discrete tokens that can then be synthesized at high fidelity with an autoregressive generative model.
Because these are powerful capabilities that come with great responsibility, we carefully vet potential applications of these sorts of models against our AI Principles.
Beyond advanced single-modality models, we are also starting to see large-scale multi-modal models. These are some of the most advanced models to date because they can accept multiple different input modalities (e.g., language, images, speech, video) and, in some cases, produce different output modalities, for example, generating images from descriptive sentences or paragraphs, or describing the visual content of images in human languages. This is an exciting direction because like the real world, some things are easier to learn in data that is multimodal (e.g., reading about something and seeing a demonstration is more useful than just reading about it). As such, pairing images and text can help with multi-lingual retrieval tasks, and better understanding of how to pair text and image inputs can yield improved results for image captioning tasks. Similarly, jointly training on visual and textual data can also help improve accuracy and robustness on visual classification tasks, while co-training on image, video, and audio tasks improves generalization performance for all modalities. There are also tantalizing hints that natural language can be used as an input for image manipulation, telling robots how to interact with the world and controlling other software systems, portending potential changes to how user interfaces are developed. Modalities handled by these models will include speech, sounds, images, video, and languages, and may even extend to structured data, knowledge graphs, and time series data.
Often these models are trained using self-supervised learning approaches, where the model learns from observations of “raw” data that has not been curated or labeled, e.g., language models used in GPT-3 and GLaM, the self-supervised speech model BigSSL, the visual contrastive learning model SimCLR, and the multimodal contrastive model VATT. Self-supervised learning allows a large speech recognition model to match the previous Voice Search automatic speech recognition (ASR) benchmark accuracy while using only 3% of the annotated training data. These trends are exciting because they can substantially reduce the effort required to enable ML for a particular task, and because they make it easier (though by no means trivial) to train models on more representative data that better reflects different subpopulations, regions, languages, or other important dimensions of representation.
All of these trends are pointing in the direction of training highly capable general-purpose models that can handle multiple modalities of data and solve thousands or millions of tasks. By building in sparsity, so that the only parts of a model that are activated for a given task are those that have been optimized for it, these multimodal models can be made highly efficient. Over the next few years, we are pursuing this vision in a next-generation architecture and umbrella effort called Pathways. We expect to see substantial progress in this area, as we combine together many ideas that to date have been pursued relatively independently.
4915
Of your peers have already watched this video.
5:00 Minutes
The most insightful time you'll spend today!
Video: How AI is Helping Biologists Protect Wildlife
According to the World Wildlife Fund, vertebrate populations have shrunk an average of 60 percent since the 1970s. And a recent UN global assessment found that we’re at risk of losing one million species to extinction, many of which may become extinct within the next decade.
To better protect wildlife, seven organizations, led by Conservation International, and Google have mapped more than 4.5 million animals in the wild using photos taken from motion-activated cameras known as camera traps. The photos are all part of Wildlife Insights, an AI-enabled, Google Cloud-based platform that streamlines conservation monitoring by speeding up camera trap photo analysis.
With photos and aggregated data available for the world to see, people can change the way protected areas are managed, empower local communities in conservation, and bring the best data closer to conservationists and decision-makers.
Camera traps help researchers assess the health of wildlife species, especially those that are reclusive and rare. Worldwide, biologists and land managers place motion-triggered cameras in forests and wilderness areas to monitor species, snapping millions of photos a year.
But what do you do when you have millions of wildlife selfies to sort through? On top of that, how do you quickly process photos where animals are difficult to find, like when an animal is in the dark or hiding behind a bush? And how do you quickly sort through up to 80 percent of photos that have no wildlife at all because the camera trap was triggered by the elements, like grass blowing in the wind?
Watch this video to find out.
Make Meaningful Analysis with Geo Boundary Public Datasets on BigQuery
6360
Of your peers have already read this article.
4:00 Minutes
The most insightful time you'll spend today!
Geospatial data is a critical component for a comprehensive analytics strategy. Whether you are trying to visualize data using geospatial parameters or do deeper analysis or modeling on customer distribution or proximity, most organizations have some type of geospatial data they would like to use – whether it be customer zipcodes, store locations, or shipping addresses. However, converting geographic data into the correct format for analysis and aggregation at different levels can be difficult. In this post, we’ll walk through some examples of how you can leverage the Google Cloud platform alongside Google Cloud Public Datasets to perform robust analytics on geographic data. The full queries can be accessed from this notebook here.
Public US Geo Boundaries dataset
BigQuery hosts a slew of public datasets for you to access and integrate into your analytics. Google pays for the storage of these datasets and provides public access to the data via the bigquery-public-data project. You only pay for queries against the data. Plus, the first 1 TB per month is free! These public datasets are valuable on their own, but when joined against your own data they can unlock new analytics use cases and save the team a lot of time.
Within the Google Cloud Public Datasets Program there are several geographic datasets. Here, we’ll work with the geo_us_boundaries dataset, which contains a set of tables that have the boundaries of different geospatial areas as polygons and coordinates based on the center point (GEOGRAPHY column type in BigQuery), published by the US Census Bureau.
Mapping geospatial points to hierarchical areas
Many times you will find yourself in situations where you have a string representing an address. However, most tools require lat/long coordinates to actually plot points. Using the Google Maps Geocoding API we can convert an address into a lat/long and then store the results in the BigQuery table.
With a lat/long representation of our point, we can join our initial dataset back onto any of the tables here using the ST_WITHIN function. This allows us to check and see if a point is within the specified polygon.
ST_WITHIN(geography_1, geography_2)
This can be helpful for ensuring standard nomenclature; for example, metropolitan areas that might be named differently. The query below maps each customers’ address to a given metropolitan area name.
SELECTcust.id as customer_id,metro.name as metro_nameFROM `looker-private-demo.retail.customers` as cust,`bigquery-public-data.geo_us_boundaries.metropolitan_divisions` as metroWHERE ST_WITHIN(ST_GEOGPOINT(cust.longitude, cust.latitude),metro.metdiv_geom)
It can also be useful for converting to designated market area (DMA), which is often used in creating targeted digital marketing campaigns.
SELECTcust.id as customer_id,dma.dma_nameFROM `looker-private-demo.retail.customers` as cust,`bigquery-public-data.geo_us_boundaries.designated_market_area` as dmaWHERE ST_WITHIN(ST_GEOGPOINT(cust.longitude, cust.latitude),dma.dma_geom)
Or for filling in missing information; for example, some addresses may be missing zip code which results in incorrect calculations when aggregating up to the zipcode level. By joining onto the zip_codes table we can ensure all coordinates are mapped appropriately and aggregate up from there.
SELECTzip.zip_code,count(distinct cust.id) as unique_customersFROM `looker-private-demo.retail.customers` as cust,`bigquery-public-data.geo_us_boundaries.zip_codes` as zipWHERE ST_WITHIN(ST_GEOGPOINT(cust.longitude, cust.latitude),zip.zip_code_geom)GROUP BY 1
Note that the zip code table isn’t a comprehensive list of all US zip codes, they are zip code tabulation areas (ZCTAs). Details about the differences can be found here. Additionally, the zip code table gives us hierarchical information, which allows us to perform more meaningful analytics. One example is leveraging hierarchical drilling in Looker. I can aggregate my total sales up to the country level, and then drill down to state, city and zipcode to identify where sales are highest. You can also use the BigQuery GeoViz tool to visualize geospatial data!
Aside from simply checking if a point is within an area, we can also use ST_DISTANCE to do something like find the closest city using the centerpoint for the metropolitan area table.
SELECTcust.id as customer_id,ARRAY_AGG(metro.name order by ST_DISTANCE(ST_GEOGPOINT(cust.longitude, cust.latitude),metro.internal_point_geom) asc limit 1)[offset(0)] as metro_nameFROM`looker-private-demo.retail.customers` as cust,`bigquery-public-data.geo_us_boundaries.metropolitan_divisions` as metroGROUP BY cust.id
This concept doesn’t just hold true for points, we can also leverage other GIS functions to see if a geospatial area is contained within areas that are listed in the boundaries datasets. If your data comes into BigQuery as a GeoJSON string, we can convert it to a GEOGRAPHY type using the ST_GEOGFROMGEOJSON function. Once our data is in a GEOGRAPHY type we can do things like check to see what urban area the geo is within – using either ST_WITHIN or ST_INTERSECTS to account for partial coverage. Here, I am using the customer’s zip code to find all metropolitan divisions where the zip code polygon and the metropolitan polygon intersect. I am then selecting the metropolitan area that has the most overlap (or the intersection has the largest area) to be the customer’s metro that we use for reporting.
SELECTcust.id as customer_id,ARRAY_AGG(metro.name order by ST_AREA(ST_INTERSECTION(zip.zip_code_geom,metro.metdiv_geom)) desc limit 1)[offset(0)] as metro_nameFROM`looker-private-demo.retail.customers` as custJOIN `bigquery-public-data.geo_us_boundaries.zip_codes` as zip on cust.zip=zip.zip_code,`bigquery-public-data.geo_us_boundaries.metropolitan_divisions` as metroWHERE ST_INTERSECTS(zip.zip_code_geom,metro.metdiv_geom)GROUP BY cust.id
The same ideas can be applied to the other tables in the dataset including the county, urban areas and National Weather Service forecast regions (which can also be useful if you want to join your datasets onto weather data).
Correcting for data discrepancy
One problem that we may run into when working with geospatial data is that different data sources may have different representations of the same information. For example, you might have one system that records state as a two letter abbreviation and another using the full name. Here, we can use the state table to join the different datasets.
SELECTst.state_name,sum(ab.sales+fn.sales) as total_salesFROM `bigquery-public-data.geo_us_boundaries.states` as stLEFT JOIN abbreviated_table as ab on ab.state = st.stateLEFT JOIN fullname_table as fn on fn.state = st.state_nameWHERE COALESCE(ab.state, fn.state) IS NOT NULLGROUP BY 1
Another example might be using the tables as a source of truth for fuzzy matching. If the address is a manually entered field somewhere in your application, there is a good chance that things will be misspelled. Different representations of the same name may prevent tables from joining with each other or lead to duplicate entries when performing aggregations. Here, I use a simple Soundex algorithm to generate a code for each county name, using helper functions from this blog post. We can see that even though some are misspelled they have the same Soundex code.
Next, we can join back onto our counties table so we make sure to use the correct spelling of the county name. Then, we can simply aggregate our data for more accurate reporting.
SELECTc.county_name,sum(sales) as total_salesFROMtableJOIN `bigquery-public-data.geo_us_boundaries.counties` as con testing.dq_fm_Soundex(table.county) = testing.dq_fm_Soundex(c.county_name)WHERE c.state_fips_code = cast(36 as string)GROUP BY 1
Note that fuzzy matching definitely isn’t perfect and you might need to try different methods or apply certain filters for it to work best depending on the specifics of your data.
The US Geo Boundary datasets allow you to perform meaningful geographic analysis without needing to worry about extracting, transforming or loading additional datasets into BigQuery. These datasets, along with all the other Google Cloud Public Datasets, will be available in the Analytics Hub. Please sign up for the Analytics Hub preview, which is scheduled to be available in the third quarter of 2021, by going to g.co/cloud/analytics-hub.
Latest Features and Updates to Globally Bolster Translation Services
5305
Of your peers have already read this article.
2:00 Minutes
The most insightful time you'll spend today!
Let’s face it: in the globalized world, which is now more than ever a digital demand world, you need to scale and reach your customers right where they’re at. Translation is a critical piece of that, whether you’re translating a website in multiple languages or releasing a document, a piece of software, or training materials.
Manual translation does not scale, which is why machine translation, powered by machine learning (ML), is becoming more important to our customers. Machine translation has historically been challenging because of the sheer volume and breadth of content that can add value when translated into multiple languages. Companies acquire and share content in many languages and formats, and scaling translation to meet needs is a tall order due to multiple document formats, integrations with optical character recognition (OCR), and the need to correct for domain-specific terminology.
Our goal is to simplify translation services, while enabling flexibility and control for our customers’ unique needs across industries. Read on to learn more about recent features and updates.
Formatting matters: Document Translation is now GA
In many cases, the layout of a document dictates how it should be interpreted—e.g., readers navigate text and discern meaning based on formatting, like bold or italicized text, or markups for headers, paragraphs, and columns. Previously, to automate translation of documents, text needed to be separated from these layout attributes, meaning the document’s structure was either lost or needed to be recreated later in the developer pipeline, after the text had been translated. This required translation teams to do a lot of extra work and maintain a lot of additional code. But now, those steps are unnecessary. Formatting can be retained throughout the translation process, handled directly by the Translation API Advanced.
This feature lets customers translate documents in 100+ languages and supports document types such as Docx, PPTx, XLSx, and PDF while preserving document formatting.
And if your needs go beyond Document Translation, we can help you translate audio as well. For real-time streaming translation, check out the Media Translation API, and for offline transcription translation, combine the Translation API with the Video Intelligence API.
Real-Time translation when you need it, Batch when you don’t
One of the biggest differentiators for Translation API Advanced’s document translation capabilities is the ability to do real-time, synchronous processing for a single file.
For example, if you are translating a business document such as HR documentation, online translation provides flexibility for smaller files and provides faster results. You can easily integrate with our APIs via REST or gRPC with mobile or browser applications, with instant access to 100+ language pairs so that content can be understandable in any supported language.
Meanwhile, batch translation allows customers to translate multiple files into multiple languages in a single request. For each request, customers can send up to 100 files with a total content size of up to 1 GB or 100 million Unicode codepoints, whichever limit is hit first.
State of the Art (SOTA) accuracy, with flexibility for customization
In order to achieve the highest level of accuracy for your translation, we now support multiple options:
- Use Google’s SOTA translation models: Each year, Google heavily invests to improve the quality of our translations across Apps, Cloud APIs, and Chrome, as well to enable multilanguage answers in Search. A popular metric for automatic quality evaluation of Machine translation systems is the BLEU score, which is based on the similarity between machine translation and the reference translations that were generated by people. While we push out incremental improvements for individual models on a monthly cadence, there are also times where we make significant leaps. In the releases since 2019, we have improved our average BLEU score by 5pts on average across 100+ languages and 7pts on low resource languages.
- Leverage glossaries for specific terms and phrases: Glossary is our terminology control feature. It allows you to import source content to define preferred translations, such as product names or department names. Then, when calling the glossary in the API request, your preferred translations will be enforced. This will work for words as well as phrase translation.
- Pick a pre-trained model with model selection: If you create custom models for machine translation, we don’t think you should have multiple client libraries and multiple APIs to maintain in order for you to use the best model for your needs. Translation API Advanced now supports Model Selection. Pick your pretrained model or pick your custom ML model built on AutoML for any language pair you’ve created and use the same API and the same client library.
- Build custom translation models with AutoML: AutoML Translation is a suite of ML products that enable you to build high quality models for your own use case or data, with limited-to-no ML expertise or coding required. Bring your past human-validated translations to improve translation specificity for your domain.
Keep localization local with Regional Endpoints
If you are a customer operating in the EU, we recently launched an endpoint specifically for EU regionalization. This is a configurable endpoint for customers to store and perform machine translation processing of customer data only in the EU multi region. For now, this only supports our pretrained translation models and glossary, but batch translations will be coming soon.
How Eli Lilly uses Cloud Translation to translate content globally
Historically, translations at Eli Lilly have been complicated: numerous translation vendors have been needed for different languages and organizations, all with their own processes and expectations. On top of that, translations have been costly and slow.
To solve this, Eli Lilly took a codified approach to enable users and systems to spend less time and resources to safely generate quality translations.
Learn more, and even catch a demo, from Thomas Griffin, Translation Tech Lead & Global Regulatory Architect for Eli Lilly.
Learn more
- To get started using Cloud Translation – Advanced, complete the setup and then try the Translate text (Advanced edition) quickstart.
- Document Translation is priced per page. For more information, see pricing.
5655
Of your peers have already downloaded this article
5:30 Minutes
The most insightful time you'll spend today!
Public cloud, big data, and AI technologies offer competitive advantages and cost savings for capital markets firms ready to make the transition. This paper discusses the three phases capital markets firms go through in transitioning to public cloud, and the workloads, benefits, and cultural changes that characterize the three phases:
Infrastructure Optimizers: The first step on the public cloud journey, where firms focus on migrating specific workloads to save costs.
Cautious Strategists: Firms build on the success of their first public cloud migrations, and begin to change the way they develop technology to increase cost savings and start taking advantage of capabilities only available on public cloud.
Transformative Innovators: Firms shift to a fully public cloud-enabled mentality, and fully leverage the flexibility and agility of the public cloud to build industry-changing solutions and attract top IT talent.
Additionally, we reveal the five things that capital markets innovators who have advanced to the transformation phase do well in their adoption of cloud, big data, and AI technologies across the front, middle, and back office functions.
4676
Of your peers have already listened to this podcast
30:30 Minutes
The most insightful time you'll spend today!
Why it’s Easier Than Ever for Developers to Break Into Machine Learning and Data Science
Posted by Jeff Dean, Senior Fellow and SVP of Google Research, on behalf of the entire Google Research community
Over the last several decades, I’ve witnessed a lot of change in the fields of machine learning (ML) and computer science. Early approaches, which often fell short, eventually gave rise to modern approaches that have been very successful. Following that long-arc pattern of progress, I think we’ll see a number of exciting advances over the next several years, advances that will ultimately benefit the lives of billions of people with greater impact than ever before. In this post, I’ll highlight five areas where ML is poised to have such impact. For each, I’ll discuss related research (mostly from 2021) and the directions and progress we’ll likely see in the next few years.
· Trend 1: More Capable, General-Purpose ML Models
· Trend 2: Continued Efficiency Improvements for ML
· Trend 3: ML Is Becoming More Personally and Communally Beneficial
· Trend 4: Growing Benefits of ML in Science, Health and Sustainability
· Trend 5: Deeper and Broader Understanding of ML
Trend 1: More Capable, General-Purpose ML Models
Researchers are training larger, more capable machine learning models than ever before. For example, just in the last couple of years models in the language domain have grown from billions of parameters trained on tens of billions of tokens of data (e.g., the 11B parameter T5 model), to hundreds of billions or trillions of parameters trained on trillions of tokens of data (e.g., dense models such as OpenAI’s 175B parameter GPT-3 model and DeepMind’s 280B parameter Gopher model, and sparse models such as Google’s 600B parameter GShard model and 1.2T parameter GLaM model). These increases in dataset and model size have led to significant increases in accuracy for a wide variety of language tasks, as shown by across-the-board improvements on standard natural language processing (NLP) benchmark tasks (as predicted by work on neural scaling laws for language models and machine translation models).
Many of these advanced models are focused on the single but important modality of written language and have shown state-of-the-art results in language understanding benchmarks and open-ended conversational abilities, even across multiple tasks in a domain. They have also shown exciting capabilities to generalize to new language tasks with relatively little training data, in some cases, with few to no training examples for a new task. A couple of examples include improved long-form question answering, zero-label learning in NLP, and our LaMDA model, which demonstrates a sophisticated ability to carry on open-ended conversations that maintain significant context across multiple turns of dialog.
(Weddell Seal image cropped from Wikimedia CC licensed image.)
Transformer models are also having a major impact in image, video, and speech models, all of which also benefit significantly from scale, as predicted by work on scaling laws for visual transformer models. Transformers for image recognition and for video classification are achieving state-of-the-art results on many benchmarks, and we’ve also demonstrated that co-training models on both image data and video data can improve performance on video tasks compared with video data alone. We’ve developed sparse, axial attention mechanisms for image and video transformers that use computation more efficiently, found better ways of tokenizing images for visual transformer models, and improved our understanding of visual transformer methods by examining how they operate compared with convolutional neural networks. Combining transformer models with convolutional operations has shown significant benefits in visual as well as speech recognition tasks.
The outputs of generative models are also substantially improving. This is most apparent in generative models for images, which have made significant strides over the last few years. For example, recent models have demonstrated the ability to create realistic images given just a category (e.g., “irish setter” or “streetcar”, if you desire), can “fill in” a low-resolution image to create a natural-looking high-resolution counterpart (“computer, enhance!”), and can even create natural-looking aerial nature scenes of arbitrary length. As another example, images can be converted to a sequence of discrete tokens that can then be synthesized at high fidelity with an autoregressive generative model.
Because these are powerful capabilities that come with great responsibility, we carefully vet potential applications of these sorts of models against our AI Principles.
Beyond advanced single-modality models, we are also starting to see large-scale multi-modal models. These are some of the most advanced models to date because they can accept multiple different input modalities (e.g., language, images, speech, video) and, in some cases, produce different output modalities, for example, generating images from descriptive sentences or paragraphs, or describing the visual content of images in human languages. This is an exciting direction because like the real world, some things are easier to learn in data that is multimodal (e.g., reading about something and seeing a demonstration is more useful than just reading about it). As such, pairing images and text can help with multi-lingual retrieval tasks, and better understanding of how to pair text and image inputs can yield improved results for image captioning tasks. Similarly, jointly training on visual and textual data can also help improve accuracy and robustness on visual classification tasks, while co-training on image, video, and audio tasks improves generalization performance for all modalities. There are also tantalizing hints that natural language can be used as an input for image manipulation, telling robots how to interact with the world and controlling other software systems, portending potential changes to how user interfaces are developed. Modalities handled by these models will include speech, sounds, images, video, and languages, and may even extend to structured data, knowledge graphs, and time series data.
Often these models are trained using self-supervised learning approaches, where the model learns from observations of “raw” data that has not been curated or labeled, e.g., language models used in GPT-3 and GLaM, the self-supervised speech model BigSSL, the visual contrastive learning model SimCLR, and the multimodal contrastive model VATT. Self-supervised learning allows a large speech recognition model to match the previous Voice Search automatic speech recognition (ASR) benchmark accuracy while using only 3% of the annotated training data. These trends are exciting because they can substantially reduce the effort required to enable ML for a particular task, and because they make it easier (though by no means trivial) to train models on more representative data that better reflects different subpopulations, regions, languages, or other important dimensions of representation.
All of these trends are pointing in the direction of training highly capable general-purpose models that can handle multiple modalities of data and solve thousands or millions of tasks. By building in sparsity, so that the only parts of a model that are activated for a given task are those that have been optimized for it, these multimodal models can be made highly efficient. Over the next few years, we are pursuing this vision in a next-generation architecture and umbrella effort called Pathways. We expect to see substantial progress in this area, as we combine together many ideas that to date have been pursued relatively independently.
More Relevant Stories for Your Company
How Digital Simulation of Physical Stores Helped e-Commerce Companies Replenish Stocks and Fulfil Online Orders at Scale in 2020
Editor’s note: We’re inviting partners from across our retail ecosystem to share stories, best practices, and tips and tricks on how they are helping retailers transform during a time that has seen tremendous change. The original version of this blog was published by Trax Retail in October 2021. Please enjoy this updated
Prototyping Language Applications Made Easy with Generative AI
Did you know generative AI allows developers to prototype applications quickly? With Generative AI Studio on Google Cloud, developers can quickly explore and customize AI models that can be leveraged in Google Cloud applications. Watch along and see how developers, with the right tools, can experiment with new ideas in
Hospitals Can Offer Interconnected Patient Experiences Using Google’s Natural Language Services
Machine Learning (ML) in healthcare helps extract data from conversations, medical records, forms, research reports, insurance claims and other documents across the care value-chain to help care providers have a holistic view of their patients to draw insights for diagnoses and treatments. With Natural Language Processing(NLP), healthcare organizations can program
Wayfair: Carving the path towards MLOps excellence with Vertex AI
Editor’s note: In part one of this blog, Wayfair shared how it supports each of its 30 million active customers using machine learning (ML). Wayfair’s Vinay Narayana, Head of ML Engineering, Bas Geerdink, Lead ML Engineer, and Christian Rehm, Senior Machine Learning Engineer, take us on a deeper dive into
