||Lars Hammarstrand <email@example.com>|
||Chalmers tekniska högskola|
||2022-06-30 – 2023-01-01|
The most recent successes of AI have been based on Supervised Learning (SL) methods, fueled by large quantities of parallel compute power and humanly annotated training data. However, we are now at a point where it is becoming intractable to manually annotate new datasets that are of sufficient quantity and quality to further boost development. Additionally, there are sensor modalities, e.g. radar, where the availability of annotated data is scarce and where it, in contrast to image data, requires expert domain knowledge to label the data accurately making it even more expensive.
Instead, many believe that Semi-Supervised Learning (SSL) will drive the next AI revolution by using the vast amount of unlabeled data (and some labelled examples). Further, using SSL we can potentially have systems that continuously and automatically learn and adapt throughout their lifecycle. In the past year, we have seen the first indication of this, where SSL methods have outperformed SL on image classification tasks, even though vast quantities of labelled data are available. The basic premise of these SSL algorithms, and which makes them so effective, is that they exploit prediction consistency (pseudo-labels) between weakly and strongly augmented versions of the same image.
Although SSL has shown very promising results, current algorithms are known to be highly sensitive to the choice of augmentation strategy, optimization algorithm, network architecture and training schedule where poor choices lead to significantly worse performance. A possible reason is that current SSL algorithms lack proper statistical modelling of uncertainties. While this is also true for SL, as SSL relies on these uncertain model predictions (pseudo-labels) as a training signal, a proper uncertainty description is even more critical.
We, thus, believe that a proper treatment of statistical model uncertainty is key to both increasing SSL robustness as well as prediction performance. Further, we see great promise in generalizing the insights from current SSL schemes, using prediction consistencies between augmented versions of the same image, and using them to train on other (additional) consistencies which are much less explored (at least for high-level tasks) as, e.g., (i) temporal consistency, (ii) consistency between different sensor modalities (vision, radar, lidar) and (iii) viewpoint consistency between sensors of the same type. Also, for these cases, proper treatment of uncertainty is crucial to determine which data to trust.
In this project, we aim to extend the SSL revolution by developing more robust and precise SSL schemes that generalize to new situations. Our focus is on (i) improving how model uncertainty is considered during training, (ii) making more efficient use of data by combining SSL and fully unsupervised training and (iiI) adapting the insights from current SSL schemes for high-level image tasks to use on the abovementioned prediction consistencies. As a second step, we plan to extend our SSL scheme to new situations by, e.g., exploring how SSL augmentation schemes can be used for radar and lidar data.