How Large Language Models are Trained and Tuned using Reinforcement Learning with Human Feedback (RLHF).

Why is ChatGPT so good?

In order to facilitate the automation and speed of maintaining surface to air missiles (SAM) systems chain of custody, Scale developed an AI powered system to track movements at these sites using EO satellite imagery. This blog post explores some of the implementation and novel concepts around the problem.
SAM systems are guided missiles that are typically launched from a mobile launch vehicle or a stationary launcher and used to defend against air attacks. Two common stationary SAM systems are the SA-2 and SA-3, often used in former Soviet satellite states.&nbsp;
From overhead imagery, these systems are typically identified by hexagonal patterns in the ground with launchers placed around the hexagon (with or without missiles), a radar placed near the center (typically accompanied by a communication vehicle), and missile transloader vehicles dispersed throughout the site.
<img src="https://drive.tiny.cloud/1/0lbhu31lrngl0y5eo8e6p0jckqgeikc94mnkaig43yuahupv/1efbb734-34e8-484f-8d41-f0858d34f69a" alt="" width="1200" height="720">
<h6 dir="ltr" style="text-align: right;"><code class="c-mrkdwn__code" data-stringify-type="code">Image Source: Skyfi</code>&nbsp;</h6>
A SAM system's perceived threat level is defined by the activity over time at a given site and the objects of interest at the site. For example, a site with ground vehicles and launchers present may be considered occupied but not operationally functional and therefore less of a threat. However, if a radar were to show up at that site it could then become operational, which is a potential cause for alert.&nbsp;
<h2 dir="ltr">Machine Learning Approach</h2>
<h3 dir="ltr">Narrowing the problem set</h3>
Tracking the status of SAM systems requires maintaining custody of 4 object classes: radars, ground vehicles, launchers, and missiles. However, there exists nuance about how these objects are detected: SA-2/SA-3 sites can be located in both dense urban and rural settings. Training an object detection model on a general region with a SAM system will lead to perceived annotation errors and false positives.&nbsp;
If, for example, there is a parking lot nearby (as there is in the image above) that does not actually contain vehicles relevant to the status of the SAM site, feeding this parking lot into the neural network without any annotations will likely confuse training and lead to declined model performance. Similarly, during inference time, the network is likely to output a number of false positive predictions around the vehicles in this parking lot when we are only interested in military vehicles relevant and directly within the SAM site.&nbsp;&nbsp;
To counteract this, Scale built a general SAM site object detector which places a focused bounding box around the site itself (see algorithm A in the figure below). This allows preprocessing for cropping out the specific site itself to feed into the final object detection network (see algorithm B in the figure below). Additionally, it provides a stand alone algorithm that can be used for broad area searches of new SA-2/SA-3 sites around the world.
<img src="https://drive.tiny.cloud/1/0lbhu31lrngl0y5eo8e6p0jckqgeikc94mnkaig43yuahupv/0c0b67e2-15fb-4f41-a1c0-c6b0fdb0d5f4" alt="" width="1200" height="430">
<h3 dir="ltr">Including Apriori Knowledge</h3>
Custody tracking in satellite images provides a unique training advantage over a static, stand alone object detection network that intakes a raw image and outputs detection predictions. In the case of custody tracking, there is almost always a prior timepoint of that Area of Interest that has been reviewed by an intel analyst and &ldquo;labeled&rdquo; with points. The work of the intel analyst can provide incredibly valuable information to the neural network learning to detect these objects on its own. To take advantage of this, instead of just feeding in an image of a SAM site to the network, we feed in both the image of the SAM site and a heatmap of the location of the objects of interest from the prior timepoint.&nbsp;
<img src="https://drive.tiny.cloud/1/0lbhu31lrngl0y5eo8e6p0jckqgeikc94mnkaig43yuahupv/46203af2-8fdd-4bbd-b10a-7a3c07c10497" alt="" width="1200" height="888">
The addition of these prior heatmaps improved algorithm detection performance across all 4 classes of interest. Object detection performance on these classes is measured using mean average precision at 50% IoU (mAP50).
<div dir="ltr" align="center">
<table><colgroup><col width="218"><col width="277"></colgroup>
<tbody>
<tr>
<td>
Class of interest
</td>
<td>
mAP50% improvement w/ Prior Heatmaps
</td>
</tr>
<tr>
<td>
SAM specific ground vehicles
</td>
<td>
4.2%
</td>
</tr>
<tr>
<td>
Launchers
</td>
<td>
17%
</td>
</tr>
<tr>
<td>
Radars
</td>
<td>
20.9%
</td>
</tr>
<tr>
<td>
Missiles
</td>
<td>
13%
</td>
</tr>
</tbody>
</table>
</div>
<h2 dir="ltr">Conclusion</h2>
The simultaneous transition to great power competition and explosion of imagery data creates an urgent need for imagery analysts to exploit imagery faster than ever before. High performance SAM detection algorithms allow analysts to receive automated disposition of forces insights at machine-speed, freeing analysts to spend their time on more complex analyses. Because the algorithm&rsquo;s outputs are structured data, analysts can further automate downstream processes such as setting up alerts for anomalous behavior or maintaining chain of custody.&nbsp;
At Scale, we believe technology should be used to empower (not replace) the analyst in crafting their narratives to deliver decision advantage.&nbsp;
NOTE:
This material is based upon work supported by the USAF AFRL/SBRK under Contract No. FA864921P0899. Any opinions, findings and conclusions or recommendations expressed in this material are those of&nbsp; the author(s) and do not necessarily reflect the views of the USAF AFRL/SBRK.&nbsp;
&nbsp;

High performance SAM detection algorithms allow analysts to receive automated disposition of forces insights at machine-speed

Maintaining (SAM) Systems

SAM Introduction

ChatGPT has captured LLM headlines and amazed the AI community with its extensive natural language processing capabilities. A new LLM from Anthropic called Claude is competitive with ChatGPT and offers great promise. We evaluate both models head to head and give you our thoughts on how they compare.

Claude vs. ChatGPT: Let’s Think Step by Step

Meet Claude: Anthropic’s Rival to ChatGPT

Scale AI, the data infrastructure for AI, has been named a Cool Vendor in the 2022 <a href="https://www.gartner.com/doc/reprints?id=1-2C7YMC4Z&amp;ct=230109&amp;st=sb">Gartner Cool Vendors&trade; in Data-Centric AI</a> by Svetlana Sicular, Arun Chandrasekaran, Sumit Agarwal, Oleksandr Matvitskyy and Anthony Mullen.
The biggest barriers for organizations looking to implement data-centric AI strategies are having the necessary resources, talent, experience and high-quality data needed to feed algorithms and leverage the power of AI. Scale provides AI solutions tailored to business use cases and machine learning tools to unlock breakthroughs like Generative AI and operationalize AI for all organizations spanning the U.S. government, researchers, startups and Fortune 500 companies across e-commerce, logistics, technology, fintech, and more industries. Organizations such as Microsoft, Open AI, SAP, Flexport and the U.S. Army partner with Scale to solve problems with data labeling and annotation, scenario-based model testing and validation, ​​content understanding and contextualization, AI catalog for asset reusability and more.
According to Gartner, &ldquo;Data challenges plague AI implementations. They are often unexpected because AI teams focus on algorithms, and data concerns come as an afterthought.&rdquo; It recommends that when evaluating data-centric AI solutions, organizations should &ldquo;work with stakeholders to determine what data characteristics are high value for AI,&rdquo; and &ldquo;add AI capabilities such as AI-specific data management, synthetic data and data labeling &ndash; to your data management toolbox.&rdquo;
&ldquo;It&rsquo;s an honor to be named a Data-Centric AI Cool Vendor by Gartner. We believe that it is a testament to Scale&rsquo;s mission to accelerate the development of AI applications and showcases the market impact of our vision,&rdquo; said Alexandr Wang, CEO and Founder of Scale AI. &ldquo;The solutions that we provide enable organizations to reap the benefits of AI, whether they are building or applying AI. This means better business outcomes, a real return on investment, and a faster time to production.&rdquo;
According to the Gartner report, &ldquo;By 2025, 90% of new data and analytics deployments will be through an established data ecosystem, causing consolidation across the data and analytics market.&rdquo; Therefore, introducing a data management approach, such as data quality, data diversity, data and analytics governance, and data literacy to AI teams now, is imperative to any organization's AI strategy.
Download the full report <a href="https://www.gartner.com/doc/reprints?id=1-2C7YMC4Z&amp;ct=230109&amp;st=sb">HERE</a>.
Gartner Disclaimer
GARTNER and COOL VENDORS are the registered trademarks of Gartner Inc., and/or its affiliates and has been used herein with permission. All rights reserved. Gartner does not endorse any vendor, product or service depicted in its research publications, and does not advise technology users to select only those vendors with the highest ratings or other designation. Gartner research publications consist of the opinions of Gartner&rsquo;s research organization and should not be construed as statements of fact. Gartner disclaims all warranties, expressed or implied, with respect to this research, including any warranties of merchantability or fitness for a particular purpose.
About Scale AI
Scale accelerates AI development by providing the data infrastructure for AI. Its data-centric platform manages the entire ML lifecycle, from data annotation and curation to model testing and evaluation, enabling any organization &ndash; from the world&rsquo;s most advanced AI teams to legacy organizations &ndash; to develop and deploy impactful AI solutions. Scale combines ML technology with skilled human insight to ensure every AI application is built on a foundation of high-quality ground truth data. Scale is trusted by the largest technology companies in the world, including Microsoft, Toyota, Nuro, Samsung, Square, PayPal, SAP, Instacart, Pinterest, Brex, Flexport and U.S. government agencies like the U.S. Army and Air Force. Founded in 2016 and valued at $7.3 billion, Scale is backed by Founders Fund, Accel, Index, and Tiger Global.

Scale AI, the data infrastructure for AI, has been named a Cool Vendor in the 2022 Gartner Cool Vendors™ in Data-Centric AI.

Scale AI Named a 2022 Cool Vendor by Gartner® | Scale AI

Scale AI Named a 2022 Cool Vendor by Gartner®

Scale AI is dedicated to helping the Department of Defense (DoD) compete, fight, and win by providing high-quality, government-owned data that can power any AI/ML and autonomy effort.&nbsp;A first step towards this goal is organizing the DoD data layer. Technologies still relatively new for the DoD - like cloud computing and data repositories - are a good start. These are insufficient however, to provide the DoD with data processing capabilities required to fuel large-scale autonomy operations.&nbsp;Scale can provide the next critical layer in the software stack which can bring autonomy efforts out of the labs, into production, and onto the battlefield.
Beyond developing, training, and deploying individual autonomous systems, taking the time to &ldquo;get the data layer right&rdquo; can have broader impact, such as enabling the creation of &ldquo;Data Annexes&rdquo; to go along with Operational Plans (OPLANs).&nbsp;These would allow DoD leaders to provide the US Army&rsquo;s autonomous vehicle fleets with key, pre-validated datasets for critical operations in theater. These would include theater-specific perception models (LiDAR, camera, radar), segmentation-mapped avenues of approach, Order-of-Battle class and type IDs for detection and recognition, EO and SAR imagery mapped to friendly schemes of maneuver, and much more. With a single update, a vehicle fleet would be prepared to move out.&nbsp;
Scale provides the data and machine learning operations (MLOps) pipelines for the world&rsquo;s most ambitious robotics and autonomy teams.&nbsp;As the DoD explores novel approaches to human-machine teaming, it will require a sophisticated, government-owned data strategy with dedicated tooling.&nbsp;This will provide the numerous vehicle platforms and programs with the highest quality data to drive production-grade autonomy. Getting this right is a precursor to mission success.&nbsp;
Scale provides an open data architecture and MLOps pipeline in alignment with the DoD Modular Open Systems Approach (MOSA), decoupled from vehicle ownership.&nbsp;This allows the US Government to reuse training data and algorithms developed within Scale platforms for other programs. Thus the DoD can create a &ldquo;data flywheel&rdquo; that strengthens as more programs develop production-grade autonomy.
<img src="https://cdn.builder.io/api/v1/image/assets%2Fe0438815ba51486bbb6a202747122d4b%2Fceb82bb90c4741f2a98cdc32b2de19a4">
Scale currently executes billions of <a href="https://scale.com/guides/data-labeling-annotation-guide" target="_blank" rel="noopener noreferrer">annotations</a> per month, maps 3,800 kilometers of roads per week, and leads the industry as the AI perception data provider. Scale partners with self-driving vehicle companies training and building systems ranging from small delivery vehicles to long-haul trucks. Our partners include Kodiak Robotics, Voyage, Optimus Ride, Toyota Research Institute, Embark, General Motors, Honda innovations, Oshkosh, and Nuro - the first autonomous ground delivery vehicle to receive <a href="https://scale.com/customers/nuro" target="_blank" rel="noopener noreferrer">DoT approval</a> for public roads. Additionally, Scale is helping to develop autonomy in other unique domains, including maritime vessels (<a href="https://scale.com/customers/sea-machines-robotics" target="_blank" rel="noopener noreferrer">Sea Machines</a>), autonomous drones (<a href="https://scale.com/customers/skydio" target="_blank" rel="noopener noreferrer">Skydio</a>), and autonomous warehouse systems (<a href="https://scale.com/customers/otto-motors" target="_blank" rel="noopener noreferrer">OTTO Motors</a>).&nbsp;
Through our in-house capabilities, direct partnerships, and an extensive subcontractor network; Scale collects, curates, and <a href="https://scale.com/guides/data-labeling-annotation-guide" target="_blank" rel="noopener noreferrer">processes</a> more ground truth perception data than any other company in the world.
We are a $7B company with real-world applications at high levels of mission-readiness. Scale has over 700 employees focused on cutting edge ML development, human-machine teaming, content understanding, and novel applications for machine learning in the commercial and federal sectors. Our Modular Open System Architecture reduces cost and time to deployment, increases opportunities for iteration, and allows the government to pick and choose the tools and data it needs, all while maintaining full data rights.
Jonathan Hudgins is a member of the Federal Team at Scale where he focuses on bringing Scale&rsquo;s proven commercial solutions to solve Federal Government problems. Jonathan is a former active duty Air Force officer and F-15E Fighter Pilot.

Autonomy Table Stakes:  DoD’s Data Layer

On November 28th, OpenAI released a new addition to the GPT-3 model family: <code>davinci-003</code>. This latest model builds on <a href="https://openai.com/blog/instruction-following/" target="_blank" rel="noopener noreferrer">InstructGPT</a>, using reinforcement learning with human feedback to better align language models with human instructions. Unlike <code>davinci-002</code>, which uses supervised fine-tuning on human-written demonstrations and highly scored model samples to improve generation quality, <code>davinci-003</code> is a true reinforcement learning with human feedback (RLHF) model. It is trained with <a href="https://openai.com/blog/openai-baselines-ppo/" target="_blank" rel="noopener noreferrer">PPO</a> to optimize the generated text&rsquo;s score against a separate &ldquo;reward model&rdquo;, which is trained on rating comparisons by human graders of different model outputs. More details can be found in OpenAI&rsquo;s <a href="https://beta.openai.com/docs/model-index-for-researchers" target="_blank" rel="noopener noreferrer">model index for researchers</a>. The net result is that <code>davinci-003</code> is tuned to produce outputs that it thinks would be scored highly by humans. Scale is proud to partner with OpenAI to provide this human feedback.
OpenAI&rsquo;s announcement email mentions the following improvements for <code>davinci-003</code>:
<ul>
<li>&ldquo;It produces higher quality writing. This will help your applications deliver clearer, more engaging, and more compelling content.</li>
<li>It can handle more complex instructions, meaning you can get even more creative with how you make use of its capabilities now.</li>
<li>It&rsquo;s better at longer form content generation, allowing you to take on tasks that would have previously been too difficult to achieve.&rdquo;</li>
</ul>
But how much better is <code>davinci-003</code>, really? We decided to put it to the test quantitatively. Using <a href="https://scale.com/spellbook" target="_blank" rel="noopener noreferrer">Scale Spellbook</a>, the platform to build, compare and deploy large language model apps, we evaluated <code>davinci-003</code> versus <code>davinci-002</code> across tasks ranging from few- and zero-shot classification, summarization, and poetry writing. Here&rsquo;s what we found.
<h3>Classification</h3>
It&rsquo;s well known that <code>davinci-002</code> performs very well on classification with a few-shot prompt, and it appears that <code>davinci-003</code> offers comparable (but slightly worse) performance few-shot. Using Spellbook&rsquo;s built-in evaluation feature on 250 Yelp reviews, we compare the two models&rsquo; classification of these reviews as Positive, Negative, or Neutral in a four-shot prompt&mdash;meaning four labeled examples are included in the prompt itself.
<img src="https://cdn.builder.io/api/v1/image/assets%2Fe0438815ba51486bbb6a202747122d4b%2F65f8dfd9c9744c15967d03341c5cdc73">
The davinci-003 model has 87.2% accuracy on this 250 Yelp review dataset sample, slightly lower than davinci-002&rsquo;s 91.6%.
However, if we ask the two models to classify these same reviews in a zero-shot prompt, we find that <code>davinci-003</code> scores more than 30% higher in accuracy than <code>davinci-002</code>&mdash;and on par with a <code>davinci-002</code> few-shot prompt! This is a substantial improvement in zero-shot classification performance. In this classification setting, adding examples to the prompt actually lowers <code>davinci-003</code>&rsquo;s accuracy.
<img src="https://lh6.googleusercontent.com/Wr30hJNdDAekBQTxtBRBN40_GR1CIElZlygaJSrAbGd6OK20V-pGS_bI5BoCF2SS4az3z7EFPYde3k5TnoVKHZd-68dwlhtHctxB7XtW5zqOXOjQkY2P8SxCkAjzlmg1PpKQ1hzEgob2W18laWD8FfSjNYShqjurevSldtMENyh8DUALgcMeYi4HQKi-">
Beyond accuracy, we found that asking <code>davinci-003</code> to explain its rationale in a zero-shot prompt works reliably, and provides a better explanation than a one-shot <code>davinci-002</code> about 75% of the time, based on Scale&rsquo;s human evaluation workforce judging a sample of 100 outputs.
<h3>Text Generation</h3>
The most notable performance leap comes in how <code>davinci-003</code> understands word and sentence structure, especially rhyming and syllables. In prompting for a poem that rhymes, <code>davinci-003</code> constructs a desired output, whereas <code>davinci-002</code> largely fails to implement rhyming.
<img src="https://cdn.builder.io/api/v1/image/assets%2Fe0438815ba51486bbb6a202747122d4b%2F31cb06da7e1a488a91f62ef14feee823">
<code>davinci-003</code> is also better equipped to handle requests for complex output structures, such as iambic pentameter (think Shakespeare). <code>davinci-002</code> certainly provides a poem, but ignores the request for lines of 5 pairs of stressed and unstressed syllables (buh-BUM). While not perfect, <code>davinci-003</code>&rsquo;s output is significantly more faithful to the requested prompt.
<img src="https://cdn.builder.io/api/v1/image/assets%2Fe0438815ba51486bbb6a202747122d4b%2Fc3fc8edf1aea432bb7bb5118a6dc8c82">
<h3>Summarization</h3>
One practical difference between the two variants is verbosity. In our experimentation, <code>davinci-003</code> consistently generates longer outputs than <code>davinci-002</code> on the same prompts. This is most visible in summarization tasks, where the value of the result may depend on its brevity.
In a sample of 60 articles, we tested both variants&rsquo; summarization capabilities by asking them to summarize the introductory paragraphs of lengthy Wikipedia pages. The additional text generated by <code>davinci-003</code> is typically beneficial, including information central to an article&rsquo;s importance. However, the resulting summaries are also wordy. In our analysis, <code>davinci-003</code> generates outputs about 65% longer than <code>davinci-002</code> under identical prompts.
<img src="https://cdn.builder.io/api/v1/image/assets%2Fe0438815ba51486bbb6a202747122d4b%2Fbbd0ed222dff4216ab9173d07d176890">
It will be important for GPT-3 users to bear in mind <code>davinci-003</code>&rsquo;s talkativeness when choosing between the two variants. In our testing, given the task of generating a 1-sentence summary, <code>davinci-003</code> produced 42 words on average, nearly doubling <code>davinci-002</code>&lsquo;s average of 23. The shortest summary <code>davinci-003</code> gave back was a beefy 32 words, compared to only 12 words for <code>davinci-002</code>. This tendency extends to multi-sentence summaries as well.
For use cases that value brevity, <code>davinci-002</code> is the safer choice. Nevertheless, for many use cases, the longer and more detailed completions provided by <code>davinci-003</code> can be beneficial.&nbsp;
<h3>Conclusion</h3>
Overall, <code>davinci-003</code> is a meaningful improvement from <code>davinci-002</code>, especially for zero-shot classification and complex text generation.
While many of the benefits of the latest iteration of GPT-3 are apparent at the surface, concrete evaluation metrics ensure models and prompt strategies are performing as well as possible.
Luckily, Spellbook makes that easy, allowing users to run instant programmatic evaluation in classification use cases, or sending model outputs to Scale&rsquo;s high-quality workforce to get quantitative evaluation metrics for generative use cases.
<a href="https://scale.com/spellbook" target="_blank" rel="noopener noreferrer">Sign up for Spellbook today!</a>
P.S. If you&rsquo;re looking for a little dramatic flair in your emails to coworkers, <code>davinci-003</code> has you covered. It performs significantly better at creatively translating text or even explaining code as different celebrities in a zero-shot setting.
&nbsp;
BranchCommitmaintestmaintest <img src="https://cdn.builder.io/api/v1/image/assets%2Fe0438815ba51486bbb6a202747122d4b%2Fa4f4e03e1cef4d28a44fcf465150aeca">
<pre class="ql-syntax" spellcheck="false"> Write a short rhyming poem about winter in New York davinci-002
davinci-003
</pre>
In New York City When winter descends
The air is so crisp and chilly
And the wind never ends
The snowflakes are so big and fluffy
You just want to grab a handful
And then make a snowball
People hurry to and fro
Bundled up in coats
Hats, and scarves, and gloves
No one wants to be caught
Lights are twinkling in the night, It's wintertime in New York City so bright.
The snow is glistening, the wind is cold,
The atmosphere is really bold.
The snowflakes fall with a gentle grace,
Adorning the city with a lovely lace.
The air is crisp, and the people merry,
It's a magical winter in New York City.
Text Generation
The most notable performance leap comes in how davinci-003 understands word and sentence structure, especially rhyming and syllables. In prompting for a poem that rhymes, davinci-003 constructs a desired output, whereas davinci-002 largely fails to implement rhyming.

We evaluate davinci-003 across a range of classification, summarization, and generation tasks.

Evaluating davinci-003: Performance, Improvements, and Gaps

How Much Better is OpenAI’s Newest GPT-3 Model?

<h3>Background</h3> The rise of social media platforms has driven the popularity of user-generated content (UGC) and it is changing from the bottom up at a pace faster than ever. Platforms are competing for user attention and engagement - some are winning, but some are falling behind. How can a social media platform stay ahead of the game? One of the most effective ways to incentivize users to create and engage with content on the platform is by curating and promoting trends. A trend is a type of content repeated and shared amongst creators -&nbsp;a specific subject matter, action, or format that rises in popularity and eventually becomes viral. Common types of trends include dance challenges, transitions, and comedy skits. While videos in a trend follow a similar format, creators usually express their creativity and individuality by adding a personal touch to the videos. For that reason, the variety of the videos within the same trend can be wide and unpredictable. The key elements of these video trends include the video’s image, motion, audio, and text overlay. For example, a transition trend is a trend where the participants show a sudden change in their appearance, often signified by a dramatic sound effect in the background music and sometimes accompanied by dance moves, lip syncing, and text overlay. The lifecycle of a trend can vary from days to months depending on various factors. To remain competitive, social media platforms need to detect the trends while they are still emerging and put them to use immediately by matching them to the right creators and promoting them to the suitable audience.&nbsp; <a href="https://scale.com/content-understanding" rel="noopener noreferrer" target="_blank">Scale Content Understanding</a> improves platform experiences by enriching content metadata, discovering trend insights, and flagging sensitive content. One feature of Content Understanding is content data enrichment for better trend detection. To facilitate this process, we developed proprietary algorithms to detect trends early, quickly, and thoroughly.&nbsp; <h3>ML Clustering Overview</h3> <img src="https://drive.tiny.cloud/1/0lbhu31lrngl0y5eo8e6p0jckqgeikc94mnkaig43yuahupv/449d8647-5509-470c-a6cf-ea38c14a62f3"> We took a human-in-the-loop (HitL) machine learning approach to detect trends. We represent trends as clusters of videos estimated from video-extracted embeddings. Once formed, these clusters are sent to a human pipeline (which we call “Cluster Labeling”) where taskers answer a series of questions about the videos in a given cluster - namely, if a sample of the videos corresponds to a trend and the potential name of the trend. The following are the phases of the ML Clustering process: <h3>Phase 1: Embedding Extraction</h3> To extract embeddings, we use a variety of multi-modal information extraction machine learning models, each of which is trained on large-scale, common datasets with high distribution overlap to the social media domain. <h3>Phase 2: Clustering</h3> Once the embeddings are extracted, we perform a clustering pipeline to take in these embeddings and output clusters. Overview Given the embeddings, we first perform <a href="https://arxiv.org/pdf/1802.03426.pdf" rel="noopener noreferrer" target="_blank">Uniform Manifold Approximation and Projection (UMAP)</a> dimensionality reduction (inspired by <a href="https://arxiv.org/abs/2203.05794" rel="noopener noreferrer" target="_blank">BERTopic</a>). UMAP takes in high-dimensional data and reduces it down to a lower dimensional space scalably (e.g. UMAP performs computation on tens of thousands of points on the order of tens of seconds). It reduces dimensionality in a way that preserves global Riemannian geometric structure and pairwise distances. On these low-dimensional embeddings, we perform an iterative hierarchical clustering process. Each individual point in embedding space is initially represented as a cluster.&nbsp;We then calculate the proximity between all pairs of clusters. We calculate the proximity of two clusters in a given pair as the variance in embedding space (i.e. the average sum of squares from the joined cluster mean embedding vector) of the points corresponding to if these clusters were to be joined. The two clusters that are closest to each other are joined given the above definition of proximity. This cluster-joining process is performed iteratively until all clusters are above a specified variance that is provided as a hyperparameter. <img src="https://drive.tiny.cloud/1/0lbhu31lrngl0y5eo8e6p0jckqgeikc94mnkaig43yuahupv/24053e18-8fbc-46a8-91ba-4ea274599636"> Stop Condition Variance Hyperparameter We carefully tune the variance hyperparameter in the stop condition to balance between making the clusters too coarse or too granular. For large variance hyperparameters, the final clusters are allowed to span a wide region in embedding space, meaning they are not semantically similar enough to represent trends (which tend to be extremely specific). Diametrically opposed, for small variance hyperparameters, the clusters end up with too few videos for users in the cluster labeling pipeline to recognize the trend. We therefore carefully tune the stop condition variance hyperparameter. This provides a balance between the clusters being fine enough that they represent groups of videos that are semantically similar enough to correspond to trends, and coarse enough that users from the cluster labeling pipeline can recognize the trend. This hyperparameter is normalized in order to reduce the amount of per-dataset manual tuning. <h3>Phase 3: Cluster Post-Processing</h3> To improve the quality and throughput of the taskers, we prune out clusters that are composed of a number of videos that is smaller than our definition of a trend. We also add a post-processing density score to help taskers focus on the clusters which are tightest in embedding space - i.e. the ones that are most likely to be semantically similar. This density score is calculated as the intra-cluster variance of each cluster in the dimensionality-reduced embedding space. <h3>Conclusion</h3> Our trend detection product consistently performs at a higher than 90% recall and 90% precision in production. Our HitL ML approach enables us to stay flexible with customization towards specific content niches while maintaining a high throughput. We are able to process tens of millions of videos with a few hours of turnaround. Our priority is to not only detect trends promptly, but also equip our customers with the data and insights on these trends to take actions immediately.&nbsp; At Scale, we believe investing in data is the key to building better platform experiences and discovering trend insights. If you’re interested in learning more about <a href="https://scale.com/content-understanding" rel="noopener noreferrer" target="_blank">Scale Content Understanding</a>, speak to our team today.&nbsp;

Discovering short-form video trends is challenging. In this blog post, we explain how we use an ML + HiTL approach for faster and higher-quality trend detection.

Discovering Emerging Short Form Video Trends with Clustering and HitL

Becoming a veteran requires a transition from the regimented lifestyle as a service member into the civilian world. The majority of veterans <a href="https://www.cfr.org/backgrounder/demographics-us-military" target="_blank" rel="noopener noreferrer">joined out of high school or college</a>, and their adult lives have been shaped by their military service. We entrust our service members at a young age with responsibilities that they otherwise would never encounter in the civilian world. These jobs not only require a highly trained technical force, but are often life or death situations.
In my case, I was only 20 when I received orders to my ship. I was responsible for operating &amp; maintaining the computers and networks that powered the multi-million dollar AEGIS Combat system. By the time I was 22, I was on a 7 month deployment in the Persian Gulf, leading a team of fellow technicians supporting combat operations in Iraq &amp; Afghanistan during the 2011 withdrawal from Iraq.
<img src="https://cdn.builder.io/api/v1/image/assets%2Fe0438815ba51486bbb6a202747122d4b%2Fa540352ee602468a940cb12fc334d5e6">
My experience isn&rsquo;t unique. The average age of the US Military service member is<a href="https://download.militaryonesource.mil/12038/MOS/Reports/2018-demographics-report.pdf" target="_blank" rel="noopener noreferrer"> 28 years old</a>, with the vast majority of service members serving one tour and becoming a veteran before the age of 25. My fellow U.S. Navy Veteran at Scale, Nick Falconi, was only 25 when the Navy trusted him to oversee the operations of the nuclear reactor that powered the submarine he served onboard. He did this alone while in port as the Duty Officer, and acted as Senior Officer while the other Senior Officers slept at sea.
Serving in the military requires a commitment that is hard to describe. Having the privilege to do amazing things at such a young age makes it hard to define an identity outside of the military. Becoming a veteran means giving up the military which has defined your adult life up until that point. One day you are a fighter pilot, special forces, or intelligence analyst in the military, and the next you are just a person trying to figure out how to succeed in the civilian world.
It&rsquo;s already hard to give that up, and enter the civilian world, but it&rsquo;s even harder to translate that experience into a civilian career. After I left the Navy and transitioned, I made the decision to go to college back home in St. Louis, but I struggled being a 23 year old freshman in classes with 18 year old kids. After a year I decided to enter the workforce while still pursuing my degree. I had difficulty finding entry level IT positions, even though the AEGIS weapon system I supported was similar to the computers &amp; network equipment found at most companies. I use the same technical skills I did in the Navy daily in my current role as the Lead of the Infrastructure &amp; Security team for Scale&rsquo;s Federal Team.
It wasn&rsquo;t until I started getting involved in the St. Louis Veterans Community and volunteering that I found the mentors and resources that helped me succeed. I learned from other veterans in the St. Louis community like Steve Wahle &amp; Micheal Bamber how to use my military experience to my advantage. They introduced me to community leaders like Maxine Clark, the founder and former CEO of Build-A-Bear Workshops, who helped me land my first civilian job supporting the computers and networks that powered the Build-A-Bear retail empire.&nbsp;
I strongly believe a successful transition to civilian life is not only key to reintegrating veterans into the community, but it is also a good business decision. Veterans have deep technical knowledge &amp; leadership skills, along with a tenacity that is hard to train that makes successful employees who can jump right into tough problems. Scale recognizes this, and has been successful in its commitment to hire veterans. I&rsquo;m proud to say that Scale&rsquo;s veteran population is nearly 5% of the US based workforce exceeding the <a href="https://www.prweb.com/releases/new_research_from_findem_sheds_light_on_status_of_veterans_in_the_workforce/prweb17533903.htm" target="_blank" rel="noopener noreferrer">percentage of veterans in the tech workforce,</a> and one-third of veteran employees are recently separated.&nbsp;
I encourage everyone this Veteran&rsquo;s Day to commit to helping new veterans. If you are a transitioning service member looking for a new challenge &amp; mission, consider <a href="https://scale.com/blog" target="_blank" rel="noopener noreferrer">Scale.</a> Our mission is to accelerate the development of AI. If that sounds exciting to you, let's talk.

If you are a transitioning service member looking for a new challenge & mission, consider Scale. 

Scale is committed to hiring veterans

Veterans Day @ Scale

Today, we&rsquo;re excited to announce the Early Access of Scale Forge (Beta) for AI-generated product imagery in seconds. With just a few clicks, marketers and brands can create stunning visual content for ad creatives, campaigns, and social media. Learn more and <a href="https://scale.com/catalog" target="_blank" rel="noopener noreferrer">sign-up for Early Access</a>.
<h2>Unlocking Stunning Product Imagery</h2>
Currently, retailers and advertisers are limited by the amount and quality of product photography they have to create an engaging shopping experience for customers. The investment needed for photoshoots, scale of product catalogs, and the diverse tastes of audiences adds to this burden. This limits the growth of businesses, and ultimately their ability to communicate their story and unique product offering to their audience.
Building on top of our existing strengths and technology in eCommerce, we built Forge to solve this problem.
<h2>Catalog Forge</h2>
With Forge, we&rsquo;re enabling customers to create AI-generated imagery via an easy-to-use design studio. Forge uses novel ML methods to allow for high-fidelity product and brand preservation for retail products.&nbsp;
With Forge, teams can:
<ol>
<li>Experiment and prototype quickly to decide the aesthetic of images</li>
<li>Operationalize and scale product image generation to enable servicing thousands of brands and products</li>
<li>Improve conversion with enhanced product images to maximize ROI&nbsp;&nbsp;</li>
</ol>
<a href="https://scaleai.wistia.com/medias/sk212njbxf" target="_blank" rel="noopener noreferrer">Video</a>
To get started, simply provide your product image, select a scene or describe your own, and create imagery through two ways:&nbsp;
<ol>
<li>Scene Enrichment: Create unique customizable scenes to generate product images that align with your brand values and product identity. Retain the original pose of each image. Best for: rigid objects</li>
<li>Full Transformation: Experiment with different poses, zoom levels, whole scenes, complementary products, and more. Only available for select demo products right now. Best for: clothing, fashion accessories</li>
</ol>
Forge enables you to create unique and memorable product images for your brand powered by AI.
<h2>New Era of Creativity in eCommerce and Retail&nbsp;</h2>
We&rsquo;re thrilled to introduce Scale Forge for eCommerce. We&rsquo;re excited to see what companies can create once they have access to the best quality AI tools. At Scale, we believe investing in data is the key to unlocking success for eCommerce companies, and we&rsquo;re just getting started.&nbsp;
To get Early Access for Forge, <a href="https://scale.com/catalog" target="_blank" rel="noopener noreferrer">sign-up for the waitlist</a>.

Catalog Forge is the fastest way to create AI-generated product imagery in seconds. Forge is now in Early Access.

Announcing Scale Forge: Available for Early Access 

The 2022 TransformX conference may now be behind us, but you can still catch all of our sessions on-demand! Nearly 32,000 virtual and in-person attendees joined the 3-day event, which included almost 80 sessions and 108 speakers. 

Blog