True ML Talks #14 - LLMs and Reinforcement Learning Co-founder @ CX Score

June 22, 2023
Share this post

We are back with another episode of  True ML Talks. In this, we again dive deep into LLMs, Reinforcement Learning, and CX Score and we are speaking with Ashwin Rao.

Ashwin Rao is a distinguished professional with a diverse background in academia, industry leadership, and entrepreneurship. He is currently a co-founder of CX Score, a seed-stage AI startup focused on empowering enterprises to enhance customer experiences on web and mobile applications.


Our conversations with Ashwin will cover below aspects:
- CX Score.
- Challenges and Applications of LLM in Retail.
- Reinforcement Learning.
- Applications of RL in the field of finance
- Using Reinforcement Learning to enhance LLMs
- Ensuring safe, unbiased, and high-quality responses in LLMs

Watch the full episode below:

TrueML talk with Ashwin Rao

CX Score

Overview of CX Ops and CX Score

CX Ops extends DevOps principles to improve the digital customer experience. It involves a collaborative approach to continuously enhance websites, web apps, and mobile apps.

The CX Score assesses customer experience using insights from a synthetic user—an AI bot that behaves like a human. It identifies issues like malfunctions, design inconsistencies, security concerns, and more, generating tickets for developers and designers.

Cross-functional teams address flagged issues and strive for continuous improvements. The synthetic user retests after issue resolution, contributing to the improvement of the CX Score over time.

Integrating CX Ops into DevOps ensures customer experience is a key focus throughout the development process. This creates seamless and engaging digital platforms for customers.

How CX Score Mimics Human Interactions

The CX Score employs a learning approach to mimic human interactions and understand what makes a digital experience intuitive and user-friendly. By observing and analyzing human behavior on websites and apps, the synthetic user, or AI bot, can learn from the signals and patterns exhibited by real users.

Supervisory data is collected to gain insights into how users navigate through digital platforms. This data includes metrics such as the time spent on different pages, the sequence of actions taken, and instances of abandonment. These signals provide valuable information about user confusion, frustrations, and areas where the experience falls short.

For example, if users frequently encounter difficulties in reaching a specific goal, such as deploying a machine learning model, the synthetic user can be trained to recognize this as a suboptimal user experience. By comparing the behavior of real users who struggle with the process against those who complete it effortlessly, the bot can understand the difference and learn what makes the experience more intuitive.

The AI bot's learning process relies on having a substantial amount of data and feedback from real users. By analyzing and mapping user journeys, it becomes possible to identify pain points, bottlenecks, and areas of improvement. This data-driven approach enables the bot to distinguish between user-friendly interactions and those that may cause frustration or confusion.

By continuously learning from human behavior, the CX Score aims to optimize the digital customer experience, making it more intuitive, streamlined, and aligned with user expectations. The goal is to ensure that the synthetic user can accurately mimic human interactions and provide valuable insights into areas where the experience can be enhanced.

Challenges and Applications of LLM in the Retail Sector

The retail industry has witnessed significant advancements in the application of AI, ML, and LLM (Large Language Models) to solve various challenges and enhance customer experiences. Here, we explore the challenges faced by the retail sector and the emerging applications of LLMs in addressing these issues.

Challenges in the Retail Industry

  1. Operations and Supply Chain: Retailers encounter complexities in managing inventory, logistics, and supply chain operations efficiently. Optimizing these processes to ensure seamless product movement and timely deliveries is crucial.
  2. Customer Experience: Providing personalized and engaging customer experiences is a top priority for retailers. This includes accurate search results, personalized recommendations, targeted marketing, and creating layouts tailored to individual preferences.

Applications of LLM in Retail

  1. Operations Optimization: LLMs can analyze vast amounts of data to optimize inventory management, demand forecasting, and supply chain operations. By leveraging LLMs, retailers can enhance their decision-making processes, improve operational efficiency, and reduce costs.
  2. Personalized Recommendations: LLMs excel in understanding customer preferences and product similarities. By utilizing customer and product embeddings, LLMs can generate highly personalized recommendations, enabling retailers to deliver targeted product suggestions and improve sales.
  3. Enhanced Search Capabilities: LLMs can transform the search experience in retail. Instead of relying solely on keyword-based searches, conversational chatbots powered by LLMs can engage in natural language dialogues, understanding context and intent to provide more accurate and relevant search results.
  4. Intelligent Customer Service: LLMs have the potential to revolutionize customer service in the retail sector. As LLM technology advances, intelligent chatbots will be able to engage in meaningful dialogues, assisting customers in finding the right products, providing guidance on prices, offering personalized shopping assistance, and handling return requests effectively.
  5. Future Possibilities: With further advancements, LLMs have the potential to become highly intelligent shopping assistants, understanding individual preferences, purchase history, and suggesting relevant products based on personalized needs. This can create a more seamless and intuitive shopping experience for customers

Reinforcement Learning

Reinforcement learning (RL) is an advanced field of machine learning where agents learn through trial and error.

In RL, an agent interacts with an environment, such as a self-driving car navigating roads filled with obstacles and traffic. The agent observes the environment's current state and selects actions to maximize cumulative rewards over time.

Rewards are numerical values that reflect the quality of an agent's decisions, considering factors like efficiency and safety. By accumulating rewards, RL agents learn to navigate effectively.

RL incorporates stochasticity to handle uncertainties in the environment, enabling agents to make optimal decisions despite unpredictable circumstances.

RL finds applications in finance, retail, robotics, and self-driving vehicles. It has also contributed to improving language models like ChatGPT, enhancing their performance and generating more accurate responses. Understanding RL's fundamentals enables us to appreciate its potential for solving complex decision-making problems and advancing AI capabilities

You get rewards and punishments for your actions depending on the rewards you get. This is how humans learn, which is why I found the field very interesting.
- Ashwin


importance of negative reward in RL:
Negative rewards in reinforcement learning (RL) are crucial for shaping agent behavior and promoting desirable outcomes. Instead of relying on human judgments, the best approach is to design systems where rewards are organic and based on actual outcomes. For example, in the context of driving, negative rewards can be associated with accidents or significant deceleration. By focusing on objective measurements like time efficiency and comfort, RL agents can learn to make optimal decisions without requiring subjective human labeling. This approach ensures robust and effective learning without the complexities of varying opinions and judgments.

Applications of Reinforcement Learning in the field of finance

  1. Portfolio Management: Reinforcement learning can be used to dynamically allocate investments based on changing market conditions, optimizing the allocation of funds across different assets and adjusting risk levels.
  2. Derivatives Pricing: Reinforcement learning techniques can be employed to accurately price and hedge complex derivatives, such as options, contributing to improved risk management in financial markets.
  3. Algorithmic Trading: Reinforcement learning can facilitate real-time trading decisions, including optimal execution strategies for large block trades and bid-ask spread control for market makers, enhancing trading efficiency and profitability.

These applications represent just a subset of the potential use cases for reinforcement learning in finance. As the field continues to evolve, more opportunities for leveraging RL are expected to emerge, leading to increased adoption and advancements in financial decision-making processes.

How RL can handle different timeframes for investments.

When considering different timeframes for investments in finance, the concept of the time value of money becomes crucial. The time value of money recognizes that the value of money received in the future is less than the same amount of money received in the present. Reinforcement learning (RL) frameworks account for this by incorporating a discount factor, which allows for the valuation of future rewards in the present.

In finance, the discount factor is determined based on the risk-free rate of return. For example, if the risk-free rate is 4%, a reward of $1 received in one year would be worth approximately $0.96 in present value terms. This discounting mechanism within RL helps capture the time value of money and the importance of different time horizons for investments.

Another consideration when maximizing financial returns is the trade-off between risk and reward. While maximizing expected returns is a common goal, it exposes investors to varying levels of uncertainty and risk. Each individual has their own risk appetite and preference for balancing potential rewards and risks. This trade-off between return and risk is a key aspect of utility theory, which addresses how individuals value different outcomes based on their risk preferences.

In finance, the reward function goes beyond mere dollar amounts and includes risk-adjusted returns. Defining an objective that incorporates risk-adjusted returns allows investors to align their investment strategies with their risk tolerance and desired trade-off between risk and reward. Utility theory provides a framework for understanding and quantifying this trade-off, helping investors make informed decisions.

Exploring the intricate relationship between timeframes, risk-adjusted returns, and investor preferences requires a deeper understanding of finance and utility theory, which can be further explored in comprehensive resources such as Ashwin Rao's book on reinforcement learning for finance.

Using Reinforcement Learning to enhance LLMs

Reinforcement learning (RL) has played a significant role in enhancing Language Models (LLMs) like Chat GPT. While RL might not be widely recognized in the mainstream, it has been a crucial technique behind the advancements in LLMs.

The journey towards developing Chat GPT began a few years ago with earlier versions like GPT-2 and GPT-3. However, these models often produced nonsensical or irrelevant responses, limiting their usability. But within a relatively short period, remarkable improvements were observed in the quality of responses generated by models like Chat GPT.

The key breakthrough came from incorporating RL as a means to control the model's responses. Imagine using Chat GPT-4 on a daily basis, where after each response it generates, you have the ability to provide feedback. You can indicate whether the response was great, valuable, or if it seemed nonsensical or irrelevant. This feedback acts as a reward or punishment for the model, shaping its future responses.

In the context of a conversation, this feedback loop creates an RL framework. The model receives the reward or punishment based on how users respond to its answers. This continuous interaction enables the model to learn and improve over time. The RL framework captures the sequential nature of conversations, with state transitions occurring as the dialogue progresses.

Through this RL framework, Chat GPT learns to understand what constitutes a sensible response versus a nonsensical one. It also helps address the issue of hallucinations, where the model generates output that might be incorrect or fabricated. By receiving feedback on these instances of hallucination, the model can learn to control and minimize them.

RL for LLMs can thus be seen as a method of hallucination control, ensuring a balance between generating creative and coherent responses without going too far into the realm of nonsensical output. By leveraging RL techniques, LLMs like Chat GPT can continually improve their performance and enhance the overall user experience.

The integration of RL into LLMs represents an important direction for future developments in language processing and understanding. It enables models to adapt and refine their responses based on real-time user feedback, leading to more accurate, relevant, and context-aware interactions.

Ensuring safe, unbiased, and high-quality responses in LLMs

Approaches for Ensuring Safe, Unbiased, and High-Quality Responses in LLMs:

  1. Incorporating Human Feedback: Human evaluators can identify and provide feedback on situations where LLM responses may be unsafe or harmful. This feedback helps train the model to recognize and avoid such cases.
  2. Defining Ethical Boundaries: Tech companies can establish predefined boundaries or limitations for certain areas like morals, ethics, and predefined behaviors. These boundaries are hardcoded and not subject to modification through RL training, ensuring consistent behavior aligned with ethical standards.
  3. Formal and Systematic Modeling: Ensuring safe, unbiased, and high-quality responses requires a more formal approach to modeling and shaping LLM behavior. This involves systematic processes for addressing bias, safety concerns, correctness, and response quality beyond just providing rewards.
  4. Continuous Monitoring: Ongoing monitoring of LLM behavior is crucial to detect and address any potential issues. Regular evaluation and analysis help identify areas where improvements can be made to enhance the safety and quality of responses.
  5. Striking a Balance: The training process must strike a balance between providing flexibility and adhering to safety and quality standards. This involves careful consideration of trade-offs and constant refinement to optimize the model's behavior.
  6. Research and Improvement: Continuous research and improvement of training techniques are essential to enhance the robustness and reliability of LLMs. This includes staying vigilant against potential exploitation by bad actors and proactively addressing emerging challenges.

Read our previous blogs in the True ML Talks series:

Keep watching the TrueML youtube series and reading the TrueML blog series.

TrueFoundry is a ML Deployment PaaS over Kubernetes to speed up developer workflows while allowing them full flexibility in testing and deploying models while ensuring full security and control for the Infra team. Through our platform, we enable Machine learning Teams to deploy and monitor models in 15 minutes with 100% reliability, scalability, and the ability to roll back in seconds - allowing them to save cost and release Models to production faster, enabling real business value realisation.

Discover More

March 28, 2024

Applications of GenAI at Google

True ML Talks
March 22, 2024

Programmatic Data Labelling and Training LLMs at

True ML Talks
January 4, 2024

TrueML Talks #26 - Enterprise GenAI and LLMOps with Labhesh Patel

True ML Talks
February 15, 2024

TrueML Talks #29 - GenAI and LLMs for Location Intelligence @ Beans.AI

True ML Talks

Related Blogs

No items found.

Blazingly fast way to build, track and deploy your models!