Exploring the frontiers of Data Science: top 3 highlights from ODSC East 2023
I recently had the privilege of attending ODSC East (the Open Data Science Conference) in Boston, U.S., from May 9th to 11th this year. As a leading AI and Data Science Conference, ODSC East gathered speakers and presenters to debate the most recent developments in the world of AI & ML. In this blog, I will report on the top three highlights of the conference.
Over the course of three days, over 5000 participants and 250 speakers joined together in a series of talks, workshops, and demos that covered the latest developments in Data Science & AI, with the conference being spread across multiple tracks including Machine Learning, MLOps, NLP, Deep Learning, Responsible AI and many more. Among many captivating discussions within all tracks, three themes stood out at the conference: AI Ethics and model responsibility, the implementation of Large Language Models (LLMs) technology such as Chat-GPT4, and accelerating AI & ML initiatives within organizations. In this blog post, I will delve into the insights by highlighting three sessions that I resonated most with.
AI Ethics & Model Responsibility
Dr. Cansu Canca, Ph.D., the program leader of the AI ethics lab at Northeastern University, provided an intriguing session on AI ethics and model-building responsibility. Her presentation emphasized the imperative for developers and organizations to adopt an ethical framework when building AI models, ensuring that technology is designed to serve humanity responsibly. Dr. Canca provided a guide and a toolset that empower companies to transpose their AI strategy into an ethical framework.
Her session focused on an open-source tool, which enables the comparison of AI ethics principles among different organizations, available at https://aiethicslab.com/big-picture/. This tool can aid in benchmarking ethical standards for companies across the globe and allows individuals to evaluate the ethical strengths and weaknesses of their own AI initiatives. By considering dimensions such as autonomy, harm-benefit, and justice, developers can gain valuable insights into the ethical implications of their AI models. Dr. Canca highlighted that ethics should guide actions throughout the AI lifecycle, prompting developers to constantly ask themselves what is the right thing to do and how to implement fair policies.
AI Principles by Organisation Type from https://aiethicslab.com/big-picture/
Another key aspect of Dr. Canca’s presentation was the emphasis on aligning AI practices with fundamental values. She aptly quoted:
“AI actors should respect the rule of law, human rights, and democratic values, throughout the AI system lifecycle. These include freedom, dignity and autonomy, privacy and data protection, non-discrimination and equality, diversity, fairness, social justice, and internationally recognized labor rights.”
To ensure adherence to these principles, AI actors should implement appropriate mechanisms and safeguards. This includes incorporating human determination into the decision-making process, ensuring the context is considered, and keeping up with the state of the art in AI technology. By integrating these measures, we can forge a path towards responsible AI development and deployment, fostering a harmonious coexistence between humans and machines.
Large Language Models (LLM) Implementation
A second presentation that stood out was the keynote by Hagay Lupesko, VP of Engineering at Mosaic ML, a US-based AI advisor, and Jay Jackson, VP of AI/ML at Oracle. Their keynote focused on debunking the misperception surrounding the implementation of Large Language Models (LLMs) and providing a practical guide to incorporating this technology into your own company. It is commonly believed that integrating LLMs into a company process or product requires exorbitant investments, with training costs reaching millions and model sizes growing to unmanageable scales.
AI-generated image from https://openai.com/dall-e-2
However, Mosaic ML explained that implementation of this technology is attainable and accessible by any company and can be initiated by first understanding what type of LLM implementation fits best for the company’s scope. Thus, they outlined three types of LLM implementation, each with its own set of considerations and trade-offs:
API Gated LLM:
The API Gated LLM approach offers a fast and convenient solution for implementing LLMs. The most common example would be using the API service of an existing model, such as Chat-GPT4, and implementing it in an own product/service offering. This can require minimal expertise as no model building has to be performed, allowing organizations to leverage pre-trained models quickly. However, this convenience comes at a price, as model compute usage prices can be high and costly at scale for companies. Privacy concerns also might arise since data is processed outside of the company boundaries, thus using confidential, R&D, or customer data within this approach might create security and privacy risks. Lastly, in this method, customization options are limited as no control over the model’s code or implementation is held.
Open Source LLM:
An Open Source LLM is a pre-built model that is readily available to a public audience. For example, META AI (https://ai.facebook.com/resources/) has built an ecosystem of pre-built AI frameworks and models which companies can use towards their own use case. The advantage to opting for an open-source path is that there is already a plethora of existing models to select from and tune to your own use case. While this approach offers flexibility, quality control can be challenging as models are built and trained outside a company’s boundaries. This requires a lot of fine-tuning to be made for adapting to specific use cases, thus requiring significant expertise in machine learning. Lastly, attention must also be given to licensing and rights associated with these models, since these can differ depending on the original creator of the model.
Own-Trained LLM:
The own-trained LLM approach involves building and training LLMs in-house. This approach offers full control, privacy, and compliance, making it a viable option for organizations. It is also easier to make this cost-effective at scale, allowing for proprietary technology that aligns with specific business scaling needs. There are already many cloud platforms on the market such as Azure or AWS that provide ML & AI usage plans to host and run the processing of these models at scale. However, implementing this approach also implies having an in-house data science team with technical LLM knowledge to build and maintain the product, which can become cost prohibitive. Moreover, building an LLM in-house implies having access to a large labeled dataset onto which to train your model, which, depending on the use case, could be difficult and costly to obtain.
By understanding these different approaches to LLM implementation, organizations can make informed decisions based on their specific requirements, resources, goals, and use cases. It is essential to carefully evaluate the advantages, limitations, and potential trade-offs associated with each approach before jumping into using LLM in your organization.
Accelerating AI/ML Initiatives
In a third talk, Shan Chidambaran, head of AI at Fujitsu North America, delved into the reasons behind the failure of organizational AI Proof of Concepts (PoCs). He highlighted three main reasons why many PoCs for AI & ML do not achieve their desired outcomes:
Skill Resources Shortage:
A scarcity of skilled data scientists and engineers poses a significant challenge for organizations. The lack of talent results in an inability to effectively combine AI technologies to solve business challenges. The difficulty in recruiting AI talent exacerbates the problem, hindering progress in AI/ML initiatives.
Long Model Preparation Times:
The time-consuming nature of model preparation often leads to decreasing stakeholder and customer interest. Lengthy iterations and delays in delivering tangible results can dampen enthusiasm and hinder the success of AI/ML PoCs. Swift action is required to maintain momentum and engagement throughout the process.
Lack of End-State Design:
Frequently, AI models are not designed with the end state in mind. For example, a model could be built without carefully considering how it would be integrated at a scale in an organization's IT architecture or into stakeholder decision-making processes. This oversight creates roadblocks and delays down the line when efforts are made to integrate the ML model. This hampers the scalability of AI solutions and prevents gaining the necessary sponsorship and adoption within organizations. Without a clear vision and strategic approach, the potential of AI/ML initiatives remains untapped.
Shan Chidambaran proposed a promising solution to address these challenges — AutoML. AutoML, which stands for Automated Machine Learning, is a research activity that streamlines AI application by automating a big chunk of the development processes. By leveraging AutoML, organizations can eliminate the extensive trial-and-error phases typically associated with building ML models. This, in turn, significantly reduces the time required for model development and delivery to the stakeholder. In fact, the application of AutoML has the potential to reduce the time to delivery by a staggering 80%, with optimism that a month-long endeavour can be transformed into a three-day process.
Moreover, AutoML enables data scientists to leverage pre-integrated AI frameworks designed for specific data structures. This eliminates the need to reinvent the wheel and can help accelerate the ML/AI adoption process. This could also free more resources from data scientists to spend more time on designing the end-state of a model so that final model integration into company processes & infrastructure is performed smoothly.
On the other hand, overly relying on AutoML for ML experimentation could bring its own pitfalls. Such might include the possibility of overfitting the model, creating ‘black box’ models in which results are difficult to explain, multicollinearity, or a total omission of integrating data outliers in the model. When using AutoML, a data scientist should be aware of these potential issues and use a balanced approach between automating model building and running manual model experiments.
Overall, AutoML can be a powerful tool for organizations to overcome skill shortages, expedite model building, and keep the stakeholder interest in AI/ML initiatives alive, as long as a balanced approach is used by the data scientist when using this technology. In this way, common pitfalls could be avoided while taking advantage of the full potential of this technology.
Image taken from https://unsplash.com/
Closing Notes
My visit to ODSC East 2023 resulted in valuable insights into the current and future state of AI. The discussions on AI ethics, model responsibility, and Large Language Models implementation, shed light on the critical considerations organizations must address when implementing AI. Driven by the need for fairness, transparency, and accountability, integrating AI ethics principles throughout the AI lifecycle is essential. Simultaneously, adopting the appropriate approach to LLM implementation empowers organizations to harness the power of language models effectively. Moreover, by considering AutoML in their model-building processes, organizations could overcome skill shortages, reduce model preparation times, and maintain stakeholder sponsorship within the organization. Overall, ODSC East proved to be an inspiring event, paving the path for the AI community for a more ethical, responsible, and innovative future.
