Some of the biggest advances today in large language models (LLMs) are driven by challenges, such as those related to privacy. Now, engineers have a way to deal with privacy concerns around sensitive data in AI. It’s called split learning.
In a split learning scenario, a model does work on a data set, before handing the project off to another LLM for completion. The initial model works up to what is called a “cut layer,” and then hands off the baton. The second model takes the inputs and completes the work without being given the identifiers and personal data that is being protected.
I found this explanation of how this works, in terms of forward propagation and backpropagation, from the MIT Media Lab, where some of my colleagues work:
“The outputs at the cut layer are sent to another entity which completes the rest of the training without looking at raw data from any client that holds the raw data. This completes a round of forward propagation without sharing raw data. The gradients are now back propagated again from its last layer until the cut layer in a similar fashion. The gradients at the cut layer (and only these gradients) are sent back to radiology client centers. The rest of back propagation is now completed at the radiology client centers. This process is continued until the distributed split learning network is trained without looking at each other’s raw data.”
Clear as mud?
Benefits of Split Learning
Using split learning can reduce the costs of communication, because less data has to travel the full trajectory involved in the workflow. But also, importantly, the approach can help AI systems to comply with privacy rules like HIPAA, and generally safeguard sensitive data assets.
A guide from Dong Nguyen Minh at Medium shows how this can work in a healthcare setting:
“Two entities holding pathology test results and radiology data can collaborate on training a deep learning model for patient diagnosis without sharing their raw data,” Minh writes. “Smaller hospitals can contribute data to an aggregate model (ML model trained by aggregating data and knowledge from various sources and combines the strengths of different models to improve its overall effectiveness) without sharing any raw data, enabling them to effectively serve those in need.”
Minh’s own description of the method uses the terms “front-end” and “back-end” as follows:
“The front-end is kept on a local device, and it processes the input data up to a certain point, typically right before the last few layers of the model that produce the final output. The output of the front-end is sent to a central server, which has the back-end of the model. The server then continues to process the output of the front-end and generates the final output.”
Function Calls
One thing that this reminds me of is what you encounter as a young computer science student getting acquainted with coding for the first time. Complex codebase scenarios use function calls as a way to port along data that needs to be acted on in some way. The function takes in inputs, and returns an output, and passes that output along. Newbies learn this process, and it forms part of the backbone of how they write code.
In this case, though, the goal is to segment workflows to protect data.
Thoughts from TEDxBoston
Recently, I did an interview with Otkrist Gupta of IIT Delhi fame, who also has significant enterprise experience with this kind of thing. Gupta broke down the idea of split learning, and its uses.
“You can prompt the models,” he said, “as people have seen, and it will exfiltrate the data for you, it’ll exfiltrate credit card numbers … the idea was: what if you never shared the data in the first place? What if you could train the model without sharing the data? So we came up with a way that you would not transmit your raw data. You would not transmit the data that you have locally. You would only transmit an intermediary representation of the neural network, and that intermediate representation would then go further and train the rest of the neural network.”
Gupta talked about how his work at Lendbuzz is informed by this advancement, and how it allows the company to lend to those with poor credit information, who might not otherwise be approved.
So split learning is being integrated into business. It is effective. And it does represent a partial response to some of those big data ownership questions. We hear about this all the time: how can people keep data safe when it’s flying around the world at the speed of light? Deliberation, it seems, is the key. Split learning will help to accomplish that.
