Work

Percepire, Ragionare, Agire: la Nuova Frontiera dell’Embodied AI

Year: 2022
Type: dissertation
Abstract: How to bridge the performance gap for simple Embodied AI tasks?Our efforts to minimize the discrepancy between human and artificial performance begin with an analysis of the role of time.Traditional methods for vision-and-language have experienced the difficulties of dealing with time series.For instance, to generate a caption for an image or video, one must be aware of the concept of time.In other words, the reasoning models need to know what is before and what is after.The same applies to Embodied AI, but on a completely different scale.The agent must not only reason about time but also perceive and act inside time.During the chain of events represented by a navigation episode, the observations made by the agent greatly depend on its previous decisions and actions.The first question arising when dealing with simple Embodied AI tasks is: How to deal with time series with long-term dependencies?Immediately after, the newborn agent needs a simple yet effective way to perceive the environment and act inside it.The action of exploring an unknown area is a simple form of interaction between the agent and its environment.For this reason, in this thesis, we consider the task of embodied exploration as a crucial testbed for our agents.The second question in our journey in Embodied AI is: How to merge visual perception and actions to enable exploration?While exploration is conceptually simple, it can enable a series of more complex downstream tasks.For instance, given an appropriate architecture for the agent, it is possible to employ the knowledge gathered in exploration to perform goal-driven navigation.In the same way, it is possible to influence the learning process with task-specific rewards to teach the agent different simple tasks.The third question is: How to combine learning methods and architectural modules to enable new tasks?Finally, we want our agent to understand our instructions.Therefore, we need to include the language modality in the loop.In our specific case, we want the agent to interpret textual instructions.Examples of such instructions may be "turn Chapter 2Working Memory Connections for LSTM Recurrent Neural Networks with Long Short-Term Memory (LSTM) make use of gating mechanisms to mitigate exploding and vanishing gradients when learning long-term dependencies.For this reason, LSTMs and other gated RNNs are widely adopted, being the standard de facto for many sequence modeling tasks.Although the memory cell inside the LSTM contains essential information, it is not allowed to influence the gating mechanism directly.In this Chapter, we improve the gate potential by including information coming from the internal cell state.The proposed modification, named Working Memory Connection, consists in adding a learnable nonlinear projection of the cell content into the network gates.This modification can fit into the classical LSTM gates without any assumption on the underlying task, being particularly effective when dealing with longer sequences.Previous research effort in this direction, which goes back to the early 2000s, could not bring a consistent improvement over vanilla LSTM.As part of our work, we identify a key issue tied to previous connections that heavily limits their effectiveness, hence preventing a successful integration of the knowledge coming from the internal cell state.We show through extensive experimental evaluation that Working Memory Connections constantly improve the performance of LSTMs on a variety of tasks.Numerical results suggest that the cell state contains useful information that is worth including in the gate structure.
Authors: Cucchiara, Rita, Bergamaschi, Sonia, Landi, Federico
Language: Italian
Open Access status: green
OpenAlex ID: W7032998610
Cites: 149
Cited by:
Related to: 10