● Method: The team developed a dynamic neural network com- prising two parts: dynamic net- work expansion for data across dissimilar domains and knowl- edge distillation for data in simi- lar domains [Figure 3(a)]. The domain similarity was deter- mined by the accuracy of the pre- vious model before training on the current task. Unibo Team (First Runners Up) ● Title: Efficient Continual Learning with Latent Rehearsal ● Members: Gabriele Graffieti, Lorenzo Pellegrini, Vincenzo Lomonaco, and Davide Maltoni ● Affiliation: University of Bologna, Italy ● Method: The team proposed a new lifelong learning approach based on latent rehearsal, namely, the replay of latent neural network activation Input instead of raw images at the input level [see the architecture and corre- sponding Android application in Figure 3(b)]. The algorithm can be deployed on the edge with low latency. Details can be found in [22]. Guinness Team Title: Learning Without Forgetting Approaches for Lifelong Robotic Vision ● Known Input Shared Convs Shared Convs Shared Convs Head 1 Head 1 Head 1 Head 1 Head 2 y1 y 1 _Known y1 y1 y2 Shared Convs Knowledge Distillation Expand KD Loss Task 1: Full Training Input Cls Loss Task 2: Knowledge Distillation With Head 1 Cls Loss Task 3: Expand Network Training Head 2 (a) External Storage (Replay Patterns) Latent Replay Layer Backward Pass (All Patterns) Class-Specific Discriminative Features (Training at Full Pace) Forward Pass (All Patterns) Output Layer (Classes) Low-Level Generic Features (Slow Training) Input Layer (Images) Backward Pass (Native Patterns) Forward Pass (Native Patterns) Concat (at Mini-Batch Level) (b) Figure 3. The challenge solutions. (a) HIK_LIG Team: dynamic network expansion for data across dissimilar domains and knowledge distillation for data in similar domains. (b) Unibo Team: replay of latent neural network activation instead of raw images at the input level (architecture and corresponding Android application). Cls: class; Convs: convolutions; KD: knowledge distillation. [(a) Source: HIK_LIG Team; used with permission; (b) source: Unibo Team; used with permission.] 14 * IEEE ROBOTICS & AUTOMATION MAGAZINE * JUNE 2020