WP5: Inference and Learning of New Knowledge

WP5 explores how well the anticipatory representations allow the recognition and preprocessing of events that have not been seen before over the training. Considering the missing information during the training, the aim is to investigate the effect of spatial and temporal dimensions of data over the inference and the assimilation of new knowledge.

Currently, neuroscientific insights represent a common source of inspiration for the advancements in machine learning. For example, in [5], Pervasive Internal Regression (PIR) is proposed as a solution that models the activations of the internal layer at a future timestep. The model is developed to mimic the prediction-error signals in biological brains. To develop qualitative analogical reasoning systems, it is important to understand how the human brain responds to different real-world stimuli. In [15], the effectiveness of fine-tuned and prompt-tuned supervision for learning neural representations is assessed under language guidance. Since most recent methods are developed for probing the brain’s language representation of Germanic languages, such as English, the study discussed in [14] focuses on the Chinese language. The inference of the temporal dimension of data is analysed from different perspectives as event timelines [12, 4, 1], explanation regeneration [6] or continual learning [3, 8, 9, 11, 13]. In the case of event timelines, the aim is to extract temporal information about different events mentioned in the text, while in the case of explanation regeneration, chains of facts are created for autoregressive reasoning in the question-answering problems.

By continual learning, we allow different systems to permanently refine and update their knowledge about the real world. In [3], a novel replay-based online continual learning method is proposed to prevent catastrophic forgetting. In [8], the effect of the sampling strategy usually used to implement the replay-based methods is analysed over the model performance. The study presents case studies for text classification and question-answering. In [9], a novel method called Entropy-based Stability-Plasticity (ESP) is proposed to address the stability-plasticity dilemma that prevents models from accumulating knowledge over lifelong learning. Given that current methods for continual learning tend to overestimate recently observed data, a new methodology that recognises and corrects the prediction bias is discussed in [11]. Further on, in [13], the problem of overestimating the effect of the recent observations is broadly addressed by rearranging the dataset in order to obtain continuously nonstationary data. Again, this work is particularly important, knowing that the current work dedicated to task-free continual learning assumes that the data is mostly stationary and only changed at a few distinct moments of time.

Unlike the previous works related to the temporal aspect of data, the inference of new knowledge can also refer to the spatial 3-D configurations. For example, the autonomous agent called Layout-aware Dreamer (LAD) introduced in [10] can navigate in a previously unseen environment and localize a remote object. The spatial reasoning in unseen environments is also explored in [7, 2] for controlling self-driven cars using natural language commands.