Related Context and Background for This Question
The OP had previously asked a very similar question that described a team where two machine learning specialists were “blocked” because they were discovering or reporting previously undisclosed dependencies on a single backend developer part-way through the Sprint. With that additional context regarding the team’s dynamic, my previous answer to the OP’s self-deleted question still seems relevant so I’m reposting it here.
Quick Analysis
In a properly functioning Scrum Team there is one product and one Product Backlog. While various Developers within the team may have specialties, they are still one single team and collectively responsible for product delivery and product quality. This requires close collaboration for all activities required to deliver the current Sprint Goal and Product Increment.
It is likely that the underlying problem is due to insufficient task decomposition during Sprint Planning, and a lack of ongoing collaboration between the backend and machine learning members of the team throughout the Sprint. This needs to be addressed structurally as a whole-team process problem.
Based on your description, you have an implied hierarchy where the machine learning people are treating the backend work as something to be delegated and "tossed over the wall" as someone else's problem instead of working with the rest of the team to define and deliver any prerequisites they may need. Scrum defines a team-based shared responsibility model for deliverables, and this certainly includes building essential technical runway such as any backend work required for the project to succeed.
Problems and Solutions
As an educated guess, not enough time is being spent on decomposing Sprint Backlog items during Sprint Planning. By the end of Sprint Planning, dependencies and timelines for the current Sprint should be fairly well understood by all involved.
In addition, the Daily Scrum is an opportunity for Developers to communicate about and collaborate on near-term dependencies. If the ML team members are working a day or more ahead of their backend teammate, this is a queuing problem that needs to be discussed. There should be enough collaboration that, at a bare minimum, the whole team knows what is needed today to get ready for the current day's work, and ideally what work will be in scope tomorrow so that near-term dependencies can be prioritized when possible.
The Daily Scrum is when anything that may prevent smooth flow is identified before it becomes an issue. If the ML team members' work will always be ahead of the backend work, then the team needs to collaboratively prioritize their tasks and coordinate their respective queues in order to maintain effective team flow.
Even if the ML developers must work from queues that empty faster than the backend queue can fill it, the fact that the ML queue developers are uncovering dependencies post facto means that the dependencies between the queues were not properly captured ahead of time, or that the Developers are not collaborating closely enough to maintain proper flow.
These planning and queue management issues should be addressed at the earliest opportunity, or at the team's next Sprint Retrospective at the latest. The whole Scrum Team needs to identify the root cause of the process problem (the "five whys" technique may be useful here) and work together rather than pointing fingers to find a way to prevent the dysfunction between the team member's respective pull-queues.
