Gradient-Based Learning Applied to Document Recognition.¶
Lecun, Y., Bottou, L., Bengio, Y., & Haffner, P. (1998). Gradient-Based Learning Applied to Document Recognition. Proceedings of the IEEE, 86(11), 2278-2324.
Cited by¶
4 citations across 4 artifacts.
Each citation links to the sentence it supports in the citing article.
Primes¶
- Convolution
- Neural computation. A neuron computes a weighted sum of its inputs over time; convolutional networks replace per-position parameters with a shared kernel to encode translation invariance, and early visual receptive fields are well modelled as oriented convolutions.
This sourceConvolutional networks share a fixed kernel across positions to encode translation invariance, modeled on local receptive fields.
Supported in partVerified against the work's full text
LeCun et al. show that replicating shared weights across space gives shift invariance and that local receptive fields extract oriented edges, but they do not model early visual receptive fields as oriented convolutions.
“In convolutional networks, as described below, shift invariance is automatically obtained by forcing the replication of weight”
- Neural computation. A neuron computes a weighted sum of its inputs over time; convolutional networks replace per-position parameters with a shared kernel to encode translation invariance, and early visual receptive fields are well modelled as oriented convolutions.
- Neighborhood
- The neuroscience notion of a receptive field as a bounded input neighborhood transferred into convolutional networks and then into graph neural networks, the same primitive carrying through three substrates with the same role.
This sourceEstablishes convolutional receptive fields — bounded input neighborhoods — carrying the receptive-field primitive from neuroscience into neural networks.
Supported in partVerified against the work's full text
LeCun et al. tie convolutional local receptive fields to Hubel and Wiesel's visual-cortex neurons, but say nothing about graph neural networks.
“Each unit in a layer receives inputs from a set of units located in a small neighborhood in the previous layer.”
- The neuroscience notion of a receptive field as a bounded input neighborhood transferred into convolutional networks and then into graph neural networks, the same primitive carrying through three substrates with the same role.
- Receptive Field
- Receptive field licenses several characteristic moves. Coverage analysis: given the units and their fields, compute the union (covered region) and the complement (gaps), where gaps predict silence and overlaps predict contention. Tiling design: choose unit count and field size under a budget — many small fields give high resolution at high cost, few large fields are cheap but blur — a trade-off that is universal across substrates. Hierarchical composition: stacking units whose fields pool over many lower-level fields produces larger effective fields with selectivity for higher-level features, the move shared by convolutional depth, organisational hierarchy, regulatory federation, and visual cortex.
This sourceConvolutional architecture with local receptive fields drawn directly from visual neuroscience; each unit responds to a local input patch and deeper layers pool over many lower fields to build larger effective fields.
Supported in partVerified against the work's full text
LeCun et al. support the convolutional-network and visual-neuroscience part: local receptive fields are combined by later layers into higher-order features; the organisational and regulatory analogies are not in the paper.
“These features are then combined by the subsequent layers in order to detect higher order features.”
- Receptive field licenses several characteristic moves. Coverage analysis: given the units and their fields, compute the union (covered region) and the complement (gaps), where gaps predict silence and overlaps predict contention. Tiling design: choose unit count and field size under a budget — many small fields give high resolution at high cost, few large fields are cheap but blur — a trade-off that is universal across substrates. Hierarchical composition: stacking units whose fields pool over many lower-level fields produces larger effective fields with selectivity for higher-level features, the move shared by convolutional depth, organisational hierarchy, regulatory federation, and visual cortex.
Mechanisms¶
- Sliding-Window or Kernel Bank
- Its strength is efficient local sensitivity at scale: one small bank of detectors, applied everywhere, finds a pattern regardless of where it sits in the input — the translation-invariance that makes convolution so effective is exactly this reuse of the same local responder across positions.
This sourceExplains convolutional networks through local receptive fields and shared weights that replicate the same feature detector across positions and provide shift-invariant recognition.
- Its strength is efficient local sensitivity at scale: one small bank of detectors, applied everywhere, finds a pattern regardless of where it sits in the input — the translation-invariance that makes convolution so effective is exactly this reuse of the same local responder across positions.
Verification¶
Does it exist? Confirmed. This work's DOI resolves to a registered record, which fixes its identity. That is all it fixes.
Does it back the claim? Read against the text for 3 of 4 citations: 3 supported in part. Each verdict is shown under its citation below, with what in the work backs the sentence.
Support is checked per citation rather than per work — the same source can be cited soundly in one article and wrongly in another. Per-citation recording began recently, so a citation with no recorded check is a gap in the record rather than evidence it went unchecked.
See how references were verified.
Registry ID ref:b8739fc00515 · see in the full table