Dynamic Adjustment of Hypothesis Space Size
Dynamically delete and resample hypotheses throughout the course of an episode.
The current Monty system initializes a fixed hypothesis space at the start of each episode and maintains it until the episode completes.
While this approach simplifies hypotheses updating, it introduces several constraints that limit adaptability and scalability.
Most importantly, it ties inference to episode boundaries and prevents Monty from adapting its hypothesis space as new observations are collected.
There are several important limitations in the current DefaultHypothesesUpdater design:
-
Fixed hypothesis space for the full episode.
A fixed hypothesis space implies a supervised notion of an episode's start and end.
Monty initializes hypotheses at the beginning and deletes them at the end, which prevents true continuous inference.
Ideally, Monty should operate without the concept of episodes, able to explore, move across objects, switch contexts, and update its hypotheses in an online unsupervised manner. -
Dependence on the first observation.
Monty relies heavily on the initial observation to sample its hypotheses.
As a result, the entire hypothesis space is anchored to the morphological features present in that first observation.
This makes it difficult for Monty to handle changes in object identity within a compositional model, or to accommodate pose changes that happen during interactions and behaviors.
Sensor noise in the first observation can also affect all the generated hypotheses from this single step.
More context for this future work item is provided here -
Inefficiency from maintaining unlikely hypotheses.
Since all hypotheses remain stored throughout the episode, even unlikely hypotheses require some computation and memory.
This overhead becomes significant as Monty learns about more objects.
Some optimizations exist, such as updating only the top hypotheses based on confidence thresholds (i.e.,x_percent_threshold), but the overall hypothesis space remains fixed and cannot shrink or grow.
The goal of this future work item is to replace the fixed hypothesis space with a dynamic hypothesis space.
The system should be able to identify and delete hypotheses that fail to accumulate evidence, while also sampling new hypotheses that are informed by more recent observations, when needed.
This would enable a shift toward continuous, episode-free inference where the system can navigate and understand its environment in a more natural way.
See the feat.dynamic_resizing feature branch to see the current prototype. This prototype validated that dynamically deleting and sampling hypotheses based on their evidence slope helps with accuracy (both object and pose detection) while speeding up inference as most steps have a very small hypothesis space. This feature is planned to be integrated into Monty. However, further ideas for heuristics to delete or resample new hypotheses are appreciated.
Help Us Make This Page Better
All our docs are open-source. If something is wrong or unclear, submit a PR to fix it!
Updated about 17 hours ago