concept
active
concept:gradient-interferenceGradient interference
Frequent-task updates overwrite rare-task features.
Neighborhood — ranked by edge-count
Papers (1)
paper
Concepts (2)
concept
- Gradient starvationassociated_withPrior account of infrequent tasks having gradients starved out.
- Update-and-forget loopextendsSmall models learn then lose rare-task updates between observations.
Related by similarity (8)
cosine ≥ 0.65 · no typed edgeEntities in the same semantic neighborhood but without a typed relation to this one — candidates for new edges or unrecognized duplicates.
- When gradients of different tasks have negative cosine similarity, harming multi-task learning.
- Optimization technique that computes weight changes by following the gradient of an error function; contrasted with evolutionary stochastic search.
- The property that qualities vary slowly, subtly, gradually across the extent of each living thing; gradients arise as natural responses to changing circumstances and create field-like character that points toward and establishes centers
- Used for updating hidden state expectations; provides dynamical process theory testable against neuronal data
- Adversarial robustness phenomenon by which a transcoder could cheat Jacobian matching.
- When non-orthogonal features cause logistic regression to identify a suboptimal probe direction
- Baseline method against which probe-based ranking is compared; more computationally expensive.
- Used by APD; poor approximations of causal importance.