We are proud to announce that six of NavInfo Europe’s research papers have been accepted by the Conference of Lifelong Learning Agents (COLLAs), taking place from the 18th to the 19th of August for the virtual conference, and the 22nd to 24th of August for the on-site conference. The conference is the first of its kind and will be hosting leading figures in AI research, with papers accepted from various prestigious organizations including MIT, Stanford, and Google to discuss the key topic of lifelong learning, how this can be achieved, and the contribution this will have to the machine learning and deep learning field.
For this conference, papers that are at the intersection of neuroscience and machine learning and have relevance to real-world applications were also accepted. Our research reflects those principles through six detailed papers and will be presented during session B of the virtual conference, on Friday the 19th.
The first paper, “Task Agnostic Representation Consolidation: A Self-Supervised Based Continual Learning Approach” addresses challenges of continual learning over non-stationary data streams in deep neural networks, as they become prone to catastrophic forgetting. We hypothesize that learning task-agnostic representations in addition to task-specific representations can potentially mitigate aforementioned problems by improving forward facilitation while reducing backward interference in CL. Although Self-supervised pre-training has been widely regarded to learn task-agnostic generalizable representations, it suffers from domain shift and runs a risk of learnt representations being overwritten in longer task sequences. We therefore propose an online Task-Agnostic Representation Consolidation (TARC), a two-stage generic CL training paradigm that intertwines task-specific and task-agnostic learning through self-supervised learning and multi-objective learning. Our method integrates task-agnostic learning into CL training thereby avoiding problems associated with self-supervised pre-training. Watch our presentation on Friday morning, during session B at B8.
The second paper we will present, “Consistency Is The Key To Mitigating Catastrophic Forgetting in Continual Learning” examines the role of consistency regularization in the Experience Replay (ER) framework under various continual learning scenarios. This research proposes to cast consistency regularization as a self-supervised pretext task, thereby enabling the use of a wide range of self-supervised learning methods as regularizes. Our team will be presenting this paper during session B at B9.
The third paper, “InBiased: Inductive Bias Distillation To Improve Generalization and Robustness Through Shape-Awareness” introduces the ‘InBiased’ framework which distills inductive bias and brings shape-awareness to the neural networks. The method, InBiaseD or Inductive Bias Distillation, distills prior knowledge from implicit information already existing in the data. We focus on “shape” as one of the meaningful inductive biases as it is also observed that humans focus more on global shape semantics to make decisions. Through extensive analysis, we show that InBiaseD reduces shortcut learning and texture bias behaviors. Further, it helps in improving generalization performance and robustness. This method includes bias alignment objectives that enforce the networks to learn more generic representations that are less vulnerable to unintended cues in the data, which results in improved generalization performance. Watch our team present their findings on this paper during session B at B13.
The fourth paper, “Synergy Between Synaptic Consolidation And Experience Replay For General Continual Learning” suggests a general continual learning method that mimics human brain functions that protect from memory erasure. The research proposes a general Continual Learning (CL) method that creates a synergy between SYNaptic consultation and dual memory Experience Replay (SYNERgy). This method maintains a semantic memory that accumulates and consolidates information across the tasks and interacts with episodic memory for effective replay. Watch our team’s presentation during Session B, at B15.
The fifth paper, “Differencing based self-supervised pretraining for scene change detection”, proposes a novel self-supervised method (DSP) to learn representations belonging to the changed region when the quantity of available labeled data is limited. DSP uses feature differencing to learn discriminatory features corresponding to the changed regions while simultaneously tackling the irrelevant changes (such as illumination, seasonal variations, and viewpoint differences between the image pairs) by enforcing temporal invariance across views. We demonstrate the effectiveness of DSP method, specifically to differences in camera viewpoints and lighting conditions, compared to transfer learning from ImageNet dataset which trains with millions of labeled images. DSP is shown to provide improved generalization, robustness to seasonal changes, distribution shift, and learning under limited labeled data. Watch our team present their findings during session B at B19.
Lastly, the sixth paper we will present, “Curbing Task Interference using Representation Similarity-Guided Multi-Task Feature Sharing” presents a method to mitigate task interference while sharing task decoders. The method focuses on determining a sharing scheme across different decoder layers based on the intent of reducing task interference. Dubbed the “Progressive Decoder Fusion (PDF), the method determines task grouping at a decoder layer using Centered Kernel Alignment between task representations at the concerned layer based on a trained network with grouping done up to the previous layer. This grouping is done progressively for all decoder layers. PDF is shown to provide improved generalization, robustness, and inter-task prediction consistency. Tune in for the findings of this presentation during session B at B20.
A word from NavInfo Europe’s Chief AI Scientist
“Learning agents must constantly adapt to their ever-changing surroundings without losing sight of the previously acquired knowledge. Deep neural networks, however, exhibit catastrophic forgetting and therefore struggle with continually learning from an environment and fail to generalize to scenarios they have not encountered during training. Humans on the other hand excel at continual learning and generalization to previously unseen scenarios. The concepts and mechanisms that enable our brains to perform effectively in complex and dynamic environments can serve as a great source of inspiration for developing next-generation AI models. Our AI department, therefore, carries out fascinating research at the intersection of neuroscience and artificial intelligence. We develop (neuro-inspired) AI models that can learn efficiently from multiple tasks.”
– Bahram Zonooz, Sr. Director and Chief Scientist at NavInfo Europe