Welcome to the April 2023 edition of Baseline, Accenture Federal Services’ machine learning newsletter. In Baseline, we share insights on important advances in machine learning technologies likely to impact our federal customers. This month we cover the following topics:

  • Major updates in Generative AI
  • Controlling Stable Diffusion with ControlNet
  • NIST’s guidance on AI Trustworthiness

Click here to subscribe to email updates: Receive Baseline every month in your inbox and stay up-to-date on the latest advancements in machine learning.


Generative AI Updates – LLaMA, ChatGPT API, GPT-4, and Google’s PaLM API



The generative AI space is moving rapidly – spurred by the interest in OpenAI’s ChatGPT model that was released in November 2022. This past month Meta AI announced the release of their large language model (LLM), LLaMA. In order to curb misuse of the model, they are only making it available under a noncommercial license to which researchers can apply to gain access.

Additionally, LLaMA was trained using only publicly available datasets, excluding proprietary and inaccessible data. Even with these data restrictions, developers are able to achieve state-of-the-art results on LLM benchmark tasks.  LLaMA is also released at several different model sizes to enhance accessibility for users. Ultimately, LLaMA addresses typical shortcomings of LLMs such as data provenance and compute requirements, which will improve future models.


OpenAI made their ChatGPT and Whisper models available as paid APIs. ChatGPT is the conversational LLM that OpenAI released in November, and Whisper is their automatic speech recognition (ASR) system built for audio ML tasks. This allows developers to iterate rapidly and harness the most up-to-date models from OpenAI. The APIs are cheaper than previous models, further lowering the barrier to integrating ChatGPT and Whisper models in developer applications.


Recently, OpenAI released their next generation GPT model, GPT-4. Unlike its predecessors, GPT-4 is a multimodal model which accepts text AND image as inputs, and outputs text responses. Additionally, GPT-4 allows users interacting with it to “steer” the model – adjusting the tone and style of its responses. Although GPT-4 still has some of the same limitations as previous models, such as the propensity to hallucinate factually incorrect answers, those errors are reduced. The expanding set of use cases available with GPT-4 will increase the set of business cases these models can be used for.

<<< Start >>>

A man is ironing clothes on the back on the van, which GPT-4 reads as unusual.

An example of GPT-4 accepting an image as an input and explaining what is unusual about it in the form of a text response. Source: Openai.com.

<<< End >>>

PaLM API and MakerSuite

Google announced their large language model, PaLM, in April 2022 and released the architecture to the open source community in December 2022. At that point, it was still difficult to utilize, as it required large compute capacity to train a model which was unattainable for most users. Now, Google has announced that they are making PaLM available via API to select developers with a waitlist coming soon.

Additionally, Google announced the release of MakerSuite, a toolset designed to help developers prototype and build applications using generative models. MakerSuite also includes access to PaLM API, prompt iteration tools,  synthetic data to help training, and custom model tuning. As they continue to roll out access to these tools, AI developers will have more intuitive ways to build out and improve upon their applications.


Controlling Stable Diffusion with ControlNet

When Stable Diffusion was released, it made waves for its ability to generate realistic images of concepts based on text inputs. ControlNet is a neural network structure that allows users to control diffusion models by adding extra conditions to the model. With previous models, there was no effective way to instruct the model to do tasks such as maintaining the background of an image, while modifying or replacing an object in the foreground.

Other uses for ControlNet include: image matting, inpainting, and style transfer. We’ve seen instances where it can be used to erase the backdrop from an image, fill in empty spaces, or copy the look of one image onto another. The ability to tune the models' outputs more directly makes ControlNet a useful tool for practitioners in the image generation space.

<<< Start >>>

<<< End >>>


NIST’s Guidance on AI Trustworthiness

As AI continues to be more powerful, it is important not to lose sight of risks, ethics, and trustworthiness concerns. To better manage these challenges faced by individuals, organizations, and society, the National Institute of Standards and Technology (NIST) has released a voluntary Artificial Intelligence Risk Management Framework (AI RMF) along with a companion AI RMF Playbook.

It is a sector and use-case agnostic framework split into two parts. The first part of the AI RMF focuses on helping organizations frame AI risks, and it outlines the characteristics of trustworthy AI systems: valid and reliable, safe, secure and resilient, accountable and transparent, explainable and interpretable, privacy-enhancing, and fair - with harmful bias managed. The second part dives into the AI RMF Core functions to provide an actionable approach during any part of the AI life cycle: Govern, Map, Measure, and Manage.

Through this socio-technical framework, organizations can responsibly accelerate the development of AI, while also protecting society. This important step forward sets the foundation to power the development of best practices and standards. An updated playbook is expected to be released this spring.

<<< Start >>>

A diagram showing the AI RMF Playbook which outlines how organizations can govern, manage, map and measure their AI strategy.

AI RMF Core functions, Image source

<<< End >>>

Accenture Federal Services is a leader in artificial intelligence for the U.S. federal government. Our Machine Learning Center of Excellence, Discovery Lab, and Advanced Research Group continually assess, develop, and adapt the world’s most innovative techniques and emerging technologies for mission-critical applications.

Shauna Revay, Ph.D.

Senior Manager – Accenture Federal Services, Machine Learning

Nalin Senjalia

Analyst – Accenture Federal Services, Analytics

Amber Chin

Analyst – Accenture Federal Services, Data Engineering

Danielle Heymann

Associate Manager – Accenture Federal Services, Data Science

Subscription Center
Subscribe to Accenture's Federal Viewpoints Blog Subscribe to Accenture's Federal Viewpoints Blog