Datasets:
File size: 3,681 Bytes
6c1c833 | 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31 32 33 34 35 36 37 38 39 40 41 42 43 44 45 46 47 48 49 50 51 52 53 54 55 56 57 | {
"field_id": "F048",
"field": "continual learning",
"directions": [
{
"direction_id": "G0",
"name": "Preserving knowledge through prior-learning evidence",
"scope": "When complete information from old tasks cannot be used directly during subsequent training, maintain earlier capabilities by retaining, compressing, or generating historical experience, or by using old-model predictions, feature targets, and pseudo-labels. These two types of prior-learning evidence need not be used together."
},
{
"direction_id": "G1",
"name": "Interference-limited model updating and capacity use",
"scope": "As new task capabilities are acquired sequentially, control changes to shared parameters or allocate, reuse, and expand specialized capacity to avoid damaging existing capabilities. This also includes constrained, compact adaptation of a pretrained backbone."
},
{
"direction_id": "G2",
"name": "Incremental class discrimination",
"scope": "As classes arrive, establish and maintain effective discrimination for both new and old classes. This includes adding class models over substantially stable features and correcting old–new class decision bias caused by sequential training."
},
{
"direction_id": "G3",
"name": "Accumulating reusable representations across stages",
"scope": "As tasks or data arrive sequentially, extend features and foundation-model capabilities that can be used later while preventing degradation of existing representations. This includes transferable features learned from labeled tasks, representation learning from unlabeled or sparsely labeled streams, and continued pretraining of language and multimodal foundation models."
},
{
"direction_id": "G4",
"name": "Adaptation under evolving distributions",
"scope": "As observed domains or populations change over time, keep a predictor effective on the current distribution while also accounting for earlier and unseen domains. This includes determining when adaptation is worthwhile."
},
{
"direction_id": "G5",
"name": "Identification of previously learned tasks",
"scope": "When task identity is unavailable at prediction time, determine which previously learned task or expert should handle an input."
},
{
"direction_id": "G6",
"name": "Continual discovery of semantic categories",
"scope": "Identify and incorporate previously unlearned semantic categories from sequentially arriving unlabeled data while maintaining recognition or organization of known categories."
},
{
"direction_id": "G7",
"name": "Distributed integration of continually changing knowledge",
"scope": "Integrate, transfer, and maintain useful knowledge among multiple clients or agents whose data and tasks change over time, subject to privacy and heterogeneity constraints."
},
{
"direction_id": "G8",
"name": "Evaluation and explanation of continual-learning behavior",
"scope": "Make knowledge retention, subsequent learning, transfer, and practical trade-offs measurable or explainable. This includes establishing sequential evaluation settings and protocols and analyzing how forgetting and plasticity vary with task relationships, representations, and optimization."
},
{
"direction_id": "G9",
"name": "Annotation-budgeted sequential acquisition",
"scope": "Given sequentially arriving unlabeled data and a limited annotation budget, decide which labels to request to support acquisition of new knowledge while accounting for existing knowledge."
}
]
}
|