Events Add an event Speakers Talks Collections
 
SigOpt Summit 2021
November 16, 2021, Online
SigOpt Summit 2021
Request Q&A
SigOpt Summit 2021
From the conference
SigOpt Summit 2021
Request Q&A
Video
Optimizing Pre-Trained Transformers in Conversational AI for Faster Inference, Better Accuracy
Available
In cart
Free
Free
Free
Free
Free
Free
Free
Free
Free
Free
Free
Free
Add to favorites
71
I like 0
I dislike 0
Available
In cart
Free
Free
Free
Free
Free
Free
Free
Free
Free
Free
Free
Free
  • Description
  • Transcript
  • Discussion

About the talk

As competition increases, brands are evolving to make customer service a critical differentiator. The rule of the day is to provide faster and smarter contact center experiences. Delivering this level of service starts with having well-trained agents that know your product, but increasingly AI can be used to boost these agents – making them even more effective. 

At the same time, the emergence of transformers has seeded this AI opportunity, increasing the potential for conversational AI. And libraries of pre-trained transformers like HuggingFace promise to democratize access to these models for these use cases. But at the same time, these pre-trained transformers have their limitations, especially when applied to real-world use cases. So how can modeling teams fine-tune pre-trained transformers to make them more impactful on their real-world modeling problems? 

In this talk, Mindtree deep learning architects Bhanu Prakash and Sulata Patra will discuss techniques to refine pre-trained transformers from the HuggingFace library in an attempt to boost accuracy and accelerate inference time. They’ll focus their talk on automating conversation summarization through the development of a custom version of the BART summarization model from the HuggingFace library. In the course of this discussion, they’ll share technical details on how they implemented solutions like SigOpt to enable this refinement through a guided experimentation and hyperparameter optimization process. And they’ll provide real-world examples for how pre-trained BART performance compares to refined BART – and how this makes a big difference for their business.

About speakers

Bhanu Prakash
Program Architect at Mindtree
Sulata Patra
Data Scientist at Mindtree

Bhanu is an accomplished machine learning and software architect with over 20 years of experience in the consumer electronics, semiconductor, and automotive industries. He has extensive experience with short-range wireless networks, Linux, firmware, machine learning and deep learning. Currently, Bhanu draws on this expertise to address a variety of problems in conversational AI at Mindtree. Within Mindtree, he is the go-to person for the most difficult and pressing technical challenges that require data science and optimization solutions. He is passionate about fostering the next generation of technical and architects, serving as a mentor and advisor to emerging engineers across these domains.

View the profile

Sulata is a results-driven leader and IT expert with over ten years of experience in manufacturing, finance, data science, and engineering. Since 2016, she has worked as a data scientist for Mindtree to address a wide variety of AI problems, using Python, Machine Learning, Deep Learning, Natural Language Processing, Computer Vision, Power BI, and Databricks. Most recently,s he led efforts to build conversational AI to enhance customer support at Mindtree. Sulata is passionate about using various AI models to solve different problems., and has expertise in data modeling, data mining, analytics, and artificial intelligence.

View the profile
Share

Hello everybody. My name is Marlon on how do I pass the jackpot? Let's see. What is Ellen DeGeneres problem areas, and how many more to Haru collaborated later and part of time until 1? Bell Finance, Burbank. Engineering consulting company, complex problems. Sytv, parent company. I have a poster ideas of experience. I'm a program 11% I also might have liked several of our protection laws affect graduate. So what's happening in the contact center transformation area? Is it current the global

market in 2027? What is the customer service? Looking for a 1200 resolution is the key. Of course? The agents were the backbone of this should be a very highly efficient to 8.2 different tools and systems are used by agents agents being bad. Helping family issues. Areas of the way, I can help help writing a second element. These intelligent automation or intelligent. You have a mini, you can put, you can customize the best of the agent. And, of course, call Quality Transport to improve

relations management. Let's take a small example of case study. Turn up to 10 minutes for call Action before. Continue with the same quality, analysis recover, all calls. What we have. We have built a mitochondria to customer service, which is supporting sport, can be customized metal detector analysis. What is Microsoft Pro 4. It also has a water pipeline? You can plug unplugged key feature for Omni Channel, email chart, voice voice call on what's happening in the light wallpaper. Call Tyson motor

and there is a possibility. This is as I told us at work, when they have option to a new set of Recon training and also they have option to choose my create a new model with an improved. Once we already had some last week. Why don't we use until to improve the performance of the sky? Better performance reduced 180 analysis. Stupid. Call, Raha. No more details will be provided by my colleague. What is the performance be? So this is kind of a snapshot of bodyguard.

1.55 x 2 x in performance important significant s i s w e r t closer to call one. Cancel Sunday night match with dark later to 32545 Highway 1. What are the best looks like? As I said before, we have a condo pre-trained model and a little bit of fine-tuning in the programming and we finally using the bar from the band Pasta, Too different. Mystical connection with what I said. What are planning to make a plan to get paper in the coming week? When is the next time I start using cigar?

Hello g19, philosopher's club and I'm here to talk about the on the experiments that Nancy love and dear to us before that. Let me introduce myself such as manufacturing, finance and data science and engineering. No. Since we're talkin 6816. I have been working as data scientist in technology, like Fightin in the machine learning, deep learning. And I'm really enthusiastic about using videos and models different problems. Yeah, this light experiments done in Asia. Before going to see, go

without cigar, I did the training and just giving me you can see the store without you. And with only two parameters and then I went with a sensitive. Am I do to help? Mary, did, you know, she was different differences in the squad, but there is definitely a difference in Roscoe. You can see the Rouge on Andrews to and finally. You know, I look forward to hugging face hugging face so that and that I can pass through window. Grand Theft Auto V do scientists keep passing from the end.

Rooster is 30.0. And finally, for that, even when I trade with mini mini parameters, you think she got that thing. Read that I did. Because I say so multimatic, definitely will help us. Getting a note to Linda. Only the right is used to school. Also the inference timers. That normally it has been production. Environment. Sentences by the total immigrants in the sentences, just messing up the diagram in sentences sentences. Okay. The more, you know, how many applications it may be necessary to maximize two competing with us giving the best results.

This is your son is referred. Yes. Temple architecture diagram. With the, you know, the Box model, Bart bass modern. We have taken the preachin modern pieces of base model, Brandon. You can see me though. It has been trained on large English Wikipedia. And so, on top of this Peyton model, you can see here. The Samsung data said we have had it and you customize data supposed to be to be considered a process. In badminton can be used to secure a p. I r a m, a

r s. You know, all the range of the parameters. I was using. Those details when I get in and out of that in you. Sohbat model. I'm just explaining Transformer model and a notary, okay. Yeah, so what is not particularly effective for some reason and translation problems even answering problem solved? So, Samsung data settings. They move on to next flight. So this is the from this late. I will go to the city, and we will see all the experiments, all the inconsistent. And you know, what time we

have news, that is straining. Training training class and inference time. And apart from them before before I did this summer, has an experiment with bot model in cedar trees. In the first experiment. I did what I wanted to do, like my friend from the newest one, simple function that was in order to function to give me the words for the machine learning simple problem. That was so complete to see how it works. And how do I parameter setting, you know the distance happening from and

which will help me, you know, the model Ultra pants. Yeah, so now I think we can go for the demo to show all the experiments from Auntie metric and and other experiments as well with the BART model. The on experiments the projects if you could still be here. This is the project. You know, I experiment with the two Matrix with the imprint Stein and roosters. And this is the Matrix where the roots school and the loss medicine laws. Are you considered as an to Mother other methods are used for the store?

Okay, before. This is this experiment we did. Are you just wanted to see. You know how it works with the Transformer. Transformer. Invaders from Google. Let me go inside Pegasus. but if you experiment with only You can see the bad sites and they're learning. And here, if you see the graph. See, you can see the progress happen and then it does. And you can see my front door. And we could have got better with many more parameters. But since the bot model was

giving better results. My phone. So, he already been seedo review. This is the highest score. and you can see through experiments with cigarettes, even the best with Ian, this is go inside this. Show property property where we can see the range of matrix. I'm so sorry mortal Matrix, we have used. And I see the mystery. Route. 110 time taken. And bruised one as the other two. Sunday. Yeah, man, go to an ellipsis. Americans. And don't. And The Matrix. You can see.

Yeah, so this is basically one live. Wake me up in go to experiment. So this is about 11 coordinates. Land on the parameters, we have used. Alarm bit older. Yeah, and if you kill grab side of port and starboard. To understand the morning. Yeah, that's all. Thank you.

Cackle comments for the website

Buy this talk

Access to the talk “Optimizing Pre-Trained Transformers in Conversational AI for Faster Inference, Better Accuracy”
Available
In cart
Free
Free
Free
Free
Free
Free
Free
Free
Free
Free
Free
Free

Full access

Get access to all videos “SigOpt Summit 2021”
Available
In cart
Free
Free
Free
Free
Free
Free
Free
Free
Free
Free
Free
Free
Ticket

Interested in topic “Education, Training and EdTech”?

You might be interested in videos from this event

November 1 - 4, 2021
Lisbon
17
3.2 M
ai, brand, chatbot, fintech, it, maas, marketing, sport, startup, technoligies

Similar talks

Michael McCourt
Head of Engineering at SigOpt
Available
In cart
Free
Free
Free
Free
Free
Free
Free
Free
Free
Free
Free
Free
Available
In cart
Free
Free
Free
Free
Free
Free
Free
Free
Free
Free
Free
Free
Rafael Gomez-Bombarelli
Professor at MIT
Available
In cart
Free
Free
Free
Free
Free
Free
Free
Free
Free
Free
Free
Free
Available
In cart
Free
Free
Free
Free
Free
Free
Free
Free
Free
Free
Free
Free
Scott Clark
Co-founder and CEO at SigOpt
Available
In cart
Free
Free
Free
Free
Free
Free
Free
Free
Free
Free
Free
Free
Available
In cart
Free
Free
Free
Free
Free
Free
Free
Free
Free
Free
Free
Free

Buy this video

Video
Access to the talk “Optimizing Pre-Trained Transformers in Conversational AI for Faster Inference, Better Accuracy”
Available
In cart
Free
Free
Free
Free
Free
Free
Free
Free
Free
Free
Free
Free

Conference Cast

With ConferenceCast.tv, you get access to our library of the world's best conference talks.

Conference Cast
946 conferences
37606 speakers
14373 hours of content
Bhanu Prakash
Sulata Patra