input × gradient, explainable ai
Simple attribution method.
3,145 technical terms and definitions
Simple attribution method.
Input filters screen prompts for policy violations or attacks.
Clean and validate user inputs before sending to LLM to prevent attacks.
Input-dependent depth networks use different layer counts per sample based on complexity.
Normalize each sample independently.
Instant Neural Graphics Primitives use hash-based encoding for fast NeRF training and rendering.
Instruct-pix2pix edits images following natural language instructions.
BLIP fine-tuned for instructions.
GPT-3 fine-tuned with RLHF for following instructions.
Instruction datasets contain task descriptions and input-output pairs for supervised learning.
Instruction models are fine-tuned to follow natural language commands.
Instruction tuning fine-tunes models on diverse instruction-following tasks.
Image editing following text instructions.
Path integral of gradients for attribution.
Use second-order information for attribution.
Inter-pair skew is delay variation across multiple differential pairs affecting parallel bus timing.
Interaction blocks in molecular GNNs update node and edge features through learned message passing.
InterCode evaluates interactive coding agents on execution-based tasks.
Generate sequences mixing images and text.
Combine at intermediate layers.
Costs from defects found before shipping.
InternLM is Shanghai AI Lab model. Multilingual research model.
Interpretability makes model decisions understandable to humans.
Understanding why a model makes specific predictions or decisions.
Interpretability helps understand model decisions. Attention visualization, probing, mechanistic. Build trust.
XAI techniques explain model decisions: attention viz, feature importance, counterfactuals. Important for trust.
Compute output bounds for input regions.
Intra-pair skew is delay mismatch within differential pair reducing common-mode noise immunity.
Test what transformations preserve predictions.
Inventory accuracy measures agreement between physical counts and system records.
Inverted residuals expand then compress feature maps using depthwise convolutions in between for efficiency.
I can help structure a pitch deck story: problem, solution, traction, market, moat, and financial/roadmap slides.
Ion exchange removes specific ions from water replacing them with hydrogen or hydroxide ions.
Image prompt adapter.
IP-Adapter enables image prompting by injecting image features into cross-attention layers.
Detect Fe-B complexes affecting lifetime.
Isolation forest adapted for temporal data detects anomalies by measuring path lengths in random decision trees built on windowed features.
Isolation Forest for time series identifies anomalies through path length in random trees over sliding windows.
Non-parametric calibration method.
Categorize and route bug reports.
Recursively amplify human supervision.
Iterated amplification decomposes oversight into manageable subtasks recursively.
One update of model weights (one batch processed).
Gradually prune and retrain.
Iterative pruning alternates between pruning and retraining gradually reducing model size.
Identify attempts to bypass safety guardrails or content policies.
Crafted inputs designed to bypass safety restrictions.
Jailbreaking bypasses safety measures through carefully crafted prompts.
Jailbreaks bypass model safety filters. Cat-and-mouse with defenses. Responsible disclosure.
Prompt injection and jailbreaks try to bypass system rules. Defenses: layered prompts, content filters, tool isolation, and conservative permissions.